Theory Is All You Need: AI, Human Cognition, and Causal Reasoning

Abstract

Scholars argue that AI can generate genuine novelty and new knowledge, and in turn, that AI and computational models of cognition will replace human decision making under uncertainty. We disagree. We argue that AI’s data-based prediction is different from human theory-based causal logic and reasoning. We highlight problems with the decades-old analogy between computers and minds as input-output devices, using large language models (LLMs) as an example. Human cognition is better conceptualized as a form of theory-based causal reasoning rather than AI’s emphasis on information processing and data-based prediction. AI uses a probability-based approach to knowledge and is largely backward-looking and imitative, while human cognition is forward-looking and capable of generating genuine novelty. We introduce the idea of “data-belief asymmetries” to highlight the difference between AI and human cognition, using the example of “heavier-than-air flight” to illustrate our arguments. Theory-based causal reasoning provides a cognitive mechanism for humans to “intervene” in the world and to engage in directed experimentation to generate new data. Throughout the article we discuss the implications of our argument for understanding the origins of novelty, new knowledge, and decision making under uncertainty.

Full text

Theory Is All You Need:
AI, Human Cognition, and Causal Reasoning†
Teppo Felin
Utah State University
& University of Oxford
Matthias Holweg
University of Oxford
† Arguments related to this paper were presented at the Strategy Science “Theory-Based View” conference at Bocconi
University as well as Harvard Business School, Aalto University, and the University of Illinois Urbana-Champaign. We are
grateful for feedback from many participants and audience members that have helped us improve our arguments. This
paper is also much improved due to feedback from editors and peer review.

2
Theory Is All You Need:
AI, Human Cognition, and Causal Reasoning
ABSTRACT
Scholars argue that AI can generate genuine novelty and new knowledge, and in
turn, that AI and computational models of cognition will replace human decision
making under uncertainty. We disagree. We argue that AI’s data-based prediction
is different from human theory-based causal logic and reasoning. We highlight
problems with the decades-old analogy between computers and minds as input-
output devices, using large language models (LLMs) as an example. Human
cognition is better conceptualized as a form of theory-based causal reasoning
rather than AI’s emphasis on information processing and data-based prediction.
AI uses a probability-based approach to knowledge and is largely backward-
looking and imitative, while human cognition is forward-looking and capable of
generating genuine novelty. We introduce the idea of “data-belief asymmetries”
to highlight the difference between AI and human cognition, using the example
of “heavier-than-air flight” to illustrate our arguments. Theory-based causal
reasoning provides a cognitive mechanism for humans to “intervene” in the
world and to engage in directed experimentation to generate new data.
Throughout the article we discuss the implications of our argument for
understanding the origins of novelty, new knowledge, and decision making under
uncertainty.
Key words: cognition, artificial intelligence, prediction, causal reasoning,
decision making, strategy, theory-based view

3
INTRODUCTION
Artificial intelligence (AI) now matches or outperforms humans in any number of games, standardized tests,
and cognitive tasks that involve high-level thinking and strategic reasoning. For example, AI engines can
readily beat humans in chess, which for decades served as a key benchmark of AI capability (Bory, 2019;
Simon, 1985). AI systems also now perform extremely well in complex board games that involve sophisticated
negotiation, complex interaction with others, alliances, deception, and understanding other players’ intentions
(e.g., Ananthaswamy, 2022). Current AI models also outperform over 90% of humans in various professional
qualification exams, like the Bar exam in law and the CPA exam in accounting (Achiam et al., 2023). AI has
also made radical strides in medical diagnosis, beating highly-trained medical professionals in diagnosing some
illnesses (e.g., Zhou et al., 2023). These rapid advances have led some AI scholars to argue that even the most
human of traits, like consciousness, will in principle soon be replicable by machines (e.g., Butlin et al., 2023;
Goyal and Bengio, 2022). In all, AI is rapidly devising algorithms that “think humanly,” “think rationally,” “act
humanly,” and “act rationally” (Csaszar and Steinberger, 2022).
Given the astonishing progress of AI, Daniel Kahneman asks (and answers) the logical next question:
“Will there be anything that is reserved for human beings? Frankly, I don’t see any reason to set limits on what
AI can do…And so it’s very difficult to imagine that with sufficient data there will remain things that only
humans can do…You should replace humans by algorithms whenever possible” (2018: 609-610, emphasis
added).
Kahneman is not alone in this assessment. Davenport and Kirby argue that “we already know that
analytics and algorithms are better at creating insights from data than most humans,” and that “this
human/machine performance gap will only increase” (2016: 29). Many scholars claim that AI is likely to
outperform humans in most—if not all—forms of reasoning and decision making (e.g., Grace et al., 2024,
Legg and Hutter, 2007; Morris et al., 2023). Some argue that strategic decision making might also be taken
over by AI (Csaszar, Ketkar and Kim, 2024), or even that science itself will be automated by “AI scientists”
(e.g., Lu et al., 2024; Manning, Zhu and Horton, 2024). One of the pioneers of AI, Geoffrey Hinton, argues
that large language models already are sentient and intelligent, and that “digital intelligence” will inevitably

4
surpass human “biological intelligence”—if it has not already done so (see Hinton, 2023; also see Bengio et al.,
2023).
Compared to machines, the cognitive and computational limitations of humans are obvious. Humans
are biased (Chater et al., 2018; Kahneman, 2011). Humans are selective about what data they attend to and
sample, and they are susceptible to confirmation bias, motivated reasoning, and hundreds of other cognitive
biases (nearly two hundred as of last count). In short, humans are “boundedly rational”—significantly
hampered by their ability to compute and process information (Simon, 1955), particularly compared to
computers (cf. Simon, 1990). And the very things that make humans boundedly rational and poor at decision
making, are seemingly the very things that enable computers to perform well on cognitive tasks. The
advantage of computers and AI is that they can handle vast amounts of data and process it quickly and in
powerful ways.
In this paper we offer a contrarian view of AI relative to human cognition—including its implications
for strategy, the emergence of novelty, and decision making under uncertainty. AI builds on the idea that
cognition—both by machines and humans—is a generalized form of information processing, a type of “input-
output” device. To illustrate cognitive differences between humans and computers, we use the example of
large language models versus human language learning. We introduce the notion of “data-belief (a)symmetry”
and the role this respectively plays in explaining AI and human cognition, using “heavier-than-air” flight as an
extended example. Human cognition is forward-looking, necessitating data-belief asymmetries which are
manifest in theories, causal reasoning, and experimentation. We argue that human cognition is driven by
forward-looking theory-based causal logic, which is distinct from the emphasis AI and computational models
of cognition place on prediction and backward-looking data. Theory-based causal reasoning enables the
generation of new and contrarian data, observations, and experimentation. We highlight the implications of these
arguments for understanding the origins of novelty, new knowledge, and decision making under uncertainty.
1
1
We need to briefly comment on the title of this paper, “theory is all you need.” Our title echoes the title of the “attention
is all you need” article that introduced the transformer architecture which (among other technologies) gave rise to recent
progress in AI (Vaswani et al., 2017). But just as “attention” is not all an AI system or large language model needs, so
theory of course is not all that humans need. In this article we simply emphasize that theory is a foundational—often
unrecognized—aspect of human cognition, one that is not easily replicable by machines and AI. We emphasize the role of
theory in human cognition, particularly the ways in which humans counterfactually think about, causally reason,
experiment, and practically “intervene” in the world.

5
AI=MIND: IS COGNITION COMPUTATION?
Modeling the human mind—thinking, rationality, and cognition—has been the central aspiration and ambition
behind AI from the 1940s to the present (McCulloch and Pitts, 1943; Turing 1948; also see Simon, 1955;
Hinton, 1992; McCorduck, 2004; Perconti and Plebe, 2020). As put by the organizers of the first conference
on AI (held at Dartmouth in 1956), their goal was to “proceed on the basis of the conjecture that every aspect
of learning or any other feature of intelligence can in principle be so precisely described that a machine can be
made to simulate it” (McCarthy et al., 2007: 12). The commonalities between models of AI and human
cognition are not just historical, but these linkages have only deepened in the intervening decades (for a
review, see Sun, 2023; also see Laird et al., 2017). Computation also underlies many other models of cognition,
including the concept of mental models (Johnson-Laird, 1983), the Bayesian brain, and predictive coding or
processing (e.g., Friston and Kiebel, 2009; Hohwy, 2013, 2020). In fact, cognitive scientist Johnson-Laird goes
so far as to argue that “any scientific theory of the mind has to treat it as an automaton” (1983: 477).
AI sees cognition as a general form of computation, specifically where “human thinking is wholly
information-processing activity” (Feigenbaum, 1963: 249; also see Simon, 1980). This logic is also captured by
computational neuroscientist David Marr who states that “most of the phenomena that are central to us as
human beings—the mysteries of life and evolution, of perception and feeling and thought—are primarily
phenomena of information processing” (1982: 4; cf. Hinton, 2023). Both mind and machine are a type of
generalized input-output device, where inputs such as stimuli and cues (“data”) are processed to yield varied
types of outputs, including decisions, capabilities, behaviors, and actions (Simon, 1980; 1990; Hasson et al.,
2020; McCelland and Rumelhart, 1981). This general model of information processing has been applied to any
number of issues and problems at the nexus of AI and cognition, including perception, learning, memory,
expertise, search, and decision making (cf. Russell and Norvig, 2022). Furthermore, the idea of human mental
activity as computation is pervasive in evolutionary arguments. For example, Cosmides and Tooby focus on
the “information-processing architecture of the human brain” and further argue that “the brain is a computer,
that is, a physical system that was designed to process information” (2013: 202-203).
Now, our overall purpose is not to exhaustively review models of AI and cognition, particularly as
excellent reviews can be found elsewhere (e.g., Aggarwal, 2018; Russell and Norvig, 2022; Goodfellow et al.,
2016). Rather, we simply want to point out the strong emphasis that past, current, and ongoing research has

6
placed on the similarities between models of AI and human cognition. To assure the reader that we are
creating a caricature of existing work, we have provided relevant, additional detail about AI-cognition
similarities in Appendix 1. There we more exhaustively point out examples of how scholars—from the 1950s
to the present—have sought to create an equivalence between AI, machines and human cognition. In all of
this work, cognition and computation (and AI) are seen as deeply connected: the underlying premise of this
work is that machines and humans are a form of input-output device, where the same underlying mechanisms
of information processing and learning are at play. The focus on computation and information processing also
is the axiomatic basis for the concept of bounded rationality (for a review, see Felin, Koenderink, and
Krueger, 2017). Bounded rationality has focused on human “computational capacities” and their limits
(Simon, 1955: 99)—and this idea has deeply shaped fields such as economics, decision theory, strategy, and the
cognitive sciences (e.g., Chater et al., 2018; Gigerenzer and Goldstein, 2024; Kahneman, 2003; Puranam et al.,
2015).
We disagree with the idea that AI and human cognition share significant similarities as forms of
computation, for reasons to be discussed next. That said, our aim in making this claim is not to take away from
the exciting breakthroughs in AI. Rather, we highlight how the analogy between AI and humans quickly
breaks down when it comes to understanding the mind and cognition, with important derivative consequences
for how we think about the emergence of novelty, new knowledge, and decision making under uncertainty. In
the next section we delve into a specific example, namely language learning by machines versus humans, to
enable us to make this point more carefully.
Machine Versus Human Learning: Different Inputs, Different Outputs
While the input-output model of minds and machines—whether we are talking about symbolic or
subsymbolic approaches (see Appendix 1 for further detail)—has been a central emphasis of AI and cognitive
science, next we highlight some important differences between machine learning and human learning. An apt
context for highlighting these differences is to focus on language. Language arguably is “the most defining trait
of human cognition (language and its relation with thought)” and therefore it “can be a true ‘window into the
mind’ ”(Chomsky, 2020: 321; also see Pinker, 1994).
2
Language provides an important “test” and context for
2
Recent comparisons between large language models (LLMs) and humans have revealed intriguing insights into formal
versus functional linguistic competence. In humans these two forms of competence rely on different neural mechanisms
(Mahowald et al., 2024).

7
understanding human and artificial intelligence. Furthermore, some have already argued that large language
models are sentient, with a few even arguing that they already closely mirror or exceed human cognition (e.g.,
Binz and Schulz, 2023; Hinton, 2023)—an assumption which we challenge.
At the most basic level, to study any system and its behavior we need to understand its inputs and
outputs. Alan Turing (1948) argued that any form of intelligence, whether human or machine, can be studied
as an input-output system. In discussing the possibilities of artificial intelligence—or “intelligent machinery” as
he called it—Turing made the analogy to an “untrained infant brain.” An infant brain is largely a blank slate,
“something like a notebook” with “little mechanism, and lots of blank sheets” (1950: 456, emphasis added; cf.
Turing, 1948). According to Turing, these blank sheets are (or need to be) filled with inputs via the process of
training and education. Through the early course of its life, an infant or child is taught and receives inputs in
the form of language and spoken words that it hears and encounters. Education and training represent the
inputs that eventually account for human linguistic capacities and outputs. And in the same way, Turing
argues, one can think of an “analogous teaching process applied to machines” (1948: 107), where machines
learn from their inputs. Turing lists various settings in which a thinking machine might show that it has
learned—including games like chess or poker, cryptography, or mathematics—and he argues that “learning of
languages would be the most impressive, since it is the most human of these activities” (Turing, 1948: 117). As
human and machine learning are often seen as a similar process, we next focus on key differences using
language learning as our example. We then highlight the implications of these differences in learning for
decision making and knowledge generation both in scientific and economic contexts.
How Machines Learn Language. To illustrate the process of machine learning, next we carefully
consider modern large language models (LLMs) and how they learn. LLMs offer a useful instantiation of
machine learning. Learning is essentially generated from scratch—bottom up, directly from the data—through
the introduction of vast amounts of training data and the algorithmic processing of the statistical associations
and interactions amongst that data. In the context of an LLM, the training data is composed of enormous
amounts of words and text, pulled together from various public sources and the Internet. To appreciate just
how much data and training these models incorporate, the latest LLMs (as of early 2024) are estimated to
include some 13 trillion tokens (a token being the rough equivalent of a word). To put this into context, if a

8
human tried to read this text—say at a speed of 9,000 words/hour (150 words/minute)—it would take over
164,000 years to read the 13 trillion words of a training dataset.
The vast corpus of text used to train an LLM is tokenized to enable natural language processing. This
typically involves converting words (or sub-word units or characters) into numerical sequences or vectors. To
illustrate, a sentence like “The cat sat on the mat” might be tokenized into a sequence like [10, 123, 56, 21, 90,
78]. Each token is passed through an embedding layer, which converts the token into a dense vector
representation that captures semantic information, such as its frequency and positional embedding. The
embedding layer has its own set of parameters (weights) that are learned during training. The attention
mechanism introduced with the “transformer” architecture (Vaswani et al, 2017), touched on by us previously,
allows the model to consider each token in context of all other surrounding tokens, thus to gain an
understanding of the wider context. Deep artificial neural networks have turned out to be extremely general
and applicable not just to text but varied domains like image recognition and computer vision, including multi-
modal applications that combine various types of data (and thus can create images from text prompts for
example).
3
From the vast data that serves as its training input, the LLM learns associations and correlations
between various statistical and distributional elements of language: specific words relative to each other, their
relationships, ordering, frequencies, and so forth. These statistical associations are based on the patterns of
word usage, context, syntax, and semantics found within the training dataset. The model develops an
“understanding” of how words and phrases tend to co-occur in varied contexts. The model does not just learn
associations but also understands correlations between different linguistic elements. In other words, it discerns
that certain words are more likely to appear in specific contexts.
Now, while the above is not a technical introduction to LLMs, it offers the broad outlines of the
process to the degree that it is relevant for our argument (for a detailed review, see Chang et al., 2024; Minaee
3
In terms of the training of an LLM, the tokenized words are submitted for algorithmic processing based on a
predetermined sequence or input length. Sequence length is important because it allows the LLM to understand context.
The (tokenized) text is not fed into the system as one long string but rather in chunks of predetermined length. This
predetermined length is variously called the context window, input or sequence length, or token limit. Recent LLM
models (as of early 2024) typically use input lengths of 2,048 tokens. (Newer models are exploring longer sequence
lengths.) Therefore, a 13 trillion token training dataset is parsed into 2,048-length sequences, enabling the algorithm to
learn language. Learning language is a statistical exercise where the LLM learns from the word patterns, context, and
dependencies found in the training data. It then uses this learning to stochastically generate outputs through next-word
prediction.

9
et al., 2024; Naveed et al., 2023; also see Resnik, 2024). The end-result of this training is an AI model that is
capable of language: more specifically, the model is capable of generating fluent and coherent text by using a
stochastic approach of “next-word prediction” in response to a prompt. In short, LLM outputs are based on
conditional probabilities given the structure of the inputs they have encountered in their training data.
Based on this broad outline of how an LLM is trained, we compare this to how humans learn
language. We should reiterate, as discussed at the outset of this article, that the basic premise behind models of
AI is that there is a symmetry between how machines and humans learn. We think it is important to carefully
point out differences, as these provide the foundation for our subsequent arguments about cognition and the
emergence of novelty.
How Humans Learn Language, Compared to Machines. The differences between human and
machine learning—when it comes to language (as well as other domains)—are stark. While LLMs are
introduced to and trained with trillions of words of text, human language “training” happens at a much slower
rate. To illustrate, a human infant or child hears—from parents, teachers, siblings, friends, and their
surroundings—an average of roughly 20,000 words a day (e.g., Gilkerson et al., 2017; Hart and Risley, 2003).
So, in its first five years a child might be exposed to—or “trained” with—some 36.5 million words. By
comparison, LLMs are trained with trillions of tokens within a short time interval of weeks or months.
4
The inputs differ radically in terms of quantity (sheer amount), but also in terms of their quality.
5
Namely, the spoken language that an infant or young child is (largely) exposed to is different from the written
language on which an LLM is trained on. Spoken language differs significantly from written language in terms
of its nature, structure, and purpose. Here the research on the differences between spoken and written
language is highly instructive (e.g., Biber, 1991). Spoken language is spontaneous (not meaningfully edited),
informal, repetitive, and often ephemeral. Written language—on the other hand—is visual and permanent,
4
For an infant to be exposed to the same 13 trillion tokens represented by the training of current LLMs, it would take or
roughly 1.8 million years.
5
Of course, an infant is not just “trained” through the language it might be exposed to by auditory means, but also
through other modalities and systems (including visual, olfactory, gustatory, and tactile ones). LLMs are largely
monomodal, though various multimodal models of AI are of course in development. But setting aside questions of
multimodality or even the “amount” of text or information that a system might be trained with, there are also deeper
questions. That is, how humans are able to learn from the things they encounter in the first place, and what they learn (or
what humans notice in the first place), is a key puzzle. Undoubtedly the biological nature and evolutionary history of
humans is central to understanding these types of questions, as is the associated ability of humans—as we emphasize in
this paper—to engage with their surroundings in novel and forward-looking ways.

10
more carefully crafted, planned, and edited. It is also denser, featuring more complex vocabulary (e.g.,
Halliday, 1989; Tannen, 2007). Importantly, the functional purposes and uses of spoken versus written
language also differ significantly. Spoken language is immediate, interactive, focused on coordinating,
expressing, and practically doing things. While written language also serves these purposes, the emphasis is
more on the communication of complex information. The vast bulk of the training data of the LLM is not
conversational (for models trained on spoken language or “raw audio,” see Lakhotia et al., 2022). Rather,
written language is more carefully thought-out. An LLM is likely to be trained with the works of Shakespeare
and Plato, academic publications, public domain books (e.g., from Project Gutenberg), lyrics, blog posts, news
articles, and various material from the Internet. This data is far “cleaner,” far more correct grammatically, and
organized. Arguably the inputs received by an LLM—in the form of written, edited and published text—are
linguistically far superior. In a statistical sense, LLM training data contains less “noise” and thus offers greater
predictive power. Even the vast stores of Wikipedia articles that are included in most LLM training datasets
are the end result of thousands of edits to ensure readability, accuracy, and flow.
Clearly humans learn language under different conditions and via different types of inputs. In short, it
can readily be argued that the human capacity for language develops differently from how machines learn
language in both quantity and quality. Humans (somehow) learn language from extremely sparse,
impoverished, and highly unsystematic inputs and data (Chomsky, 1975). Compared to LLMs, human
linguistic capabilities are radically “underdetermined” by the inputs. That is, the relatively sparse linguistic
inputs can scarcely account for the radically novel outputs generated by humans.
6
Beyond the quantitative and qualitative differences in inputs (when it comes language learning by
LLMs versus humans), it is important to compare the linguistic outputs and capabilities of machines versus
humans. In terms of output, LLMs are said to be “generative” (the acronym GPT stands for “generative
6
This logic is aptly captured by Chomsky: “One can describe the child’s acquisition of knowledge of language as a kind of
theory construction. Presented with highly restricted data, he constructs a theory of language of which this data is a sample
(and, in fact, a highly degenerate sample, in the sense that much of it must be excluded as irrelevant and incorrect—thus the
child learns rules of grammar that identify much of what he has heard as ill-formed, inaccurate, and inappropriate). The child’s
ultimate knowledge of language obviously extends far beyond the data presented to him. In other words, the theory he has in
some way developed has a predictive scope of which the data on which it is based constitute a negligible part. The normal use
of language characteristically involves new sentences, sentences that bear no point-by-point resemblance or analogy to
those in the child’s experience (1975: 179, emphasis added).” Our goal is not to endorse Chomsky’s theory of universal
grammar. Rather, we concur with this specific quote in terms of its characterization of the input-output relationship,
where human linguistic outputs are underdetermined by the inputs children receive. Broadly this also links to the
alternative approach that we focus on (the theory-based view of cognition), discussed in the second half of the paper.

11
pretrained transformer”). But in what sense are LLMs generative? They are generative in the specific sense
that they are able to create novel outputs by probabilistically sampling from the vast combinatorial possibilities
in the associational and correlational network of word frequencies, positional encodings, and co-occurrences
encountered in the training data (Vaswani, 2017).
7
The LLM is generative in the sense that the text that is
produced is not simply plagiarized or copied verbatim from existing sources contained in the pretraining data
(McCoy et al., 2023). In the process of generating text, the parameters (weights and biases) determine how
much influence different parts of the training data probabilistically have on the output. For example, in a
sentence completion task, the weights—developed from the corpus of the training data—help the model
decide which words are most likely to come next, based on the context provided by the input. The output is
statistically derived (or put differently, probabilistically drawn) from the training data’s underlying linguistic
structure. The outputs therefore have compositional novelty (in terms of novel ways of saying the same
thing—more on this below), and they also manifest some analogical generalization (McCoy et al., 2023). That
said, any assessment of how “good” an LLM is needs to recognize “the problem that [LLMs] were trained to
solve: next-word prediction” (McCoy et al., 2023). And as next-word prediction engines, LLMs certainly
demonstrate exceptional capabilities.
BEYOND MIRRORING: CAN AI GENERATE GENUINE NOVELTY?
So far we have summarized the central elements of a particular AI system—an LLM—and compared it with
humans. Next we further address whether an AI can be said to be “intelligent” and whether it can generate
genuine novelty and new knowledge. While our focus remains on LLMs, we extend our arguments to other
forms of AI and cognitive approaches that focus on data and prediction. We concurrently raise questions
about whether an AI system meaningfully can originate new knowledge and engage in decision making under
uncertainty.
7
Relative to the idea of next-word prediction (and the probabilistic “draw” of the next word), there are different ways for
this to happen. For example, a model might always pick the most likely next word (greedy). Or a model might explore
multiple sequences simultaneously (beam search), along with many other approaches (top-k sampling, top-p sampling etc).
In practice, different types of prompts (depending on prompt context, length, tone, style) lead to different types of
sampling and next-word prediction (Holtzman et al., 2019), as will changing the “temperature” setting of the model.

12
AI: Intelligence and New Knowledge?
As we have foreshadowed above, an AI like an LLM seems to “mirror” the inputs it has been trained with
rather than meaningfully manifest some form of intelligence. But beyond next-word prediction and linguistic
fluency, could an LLM do a better job than humans in decision making under uncertainty (e.g., Csaszar et al.,
2024; cf. Kahneman, 2018)—or could an LLM or “AI scientist” perhaps even “automate” science itself (e.g.,
Lu et al., 2024; Manning et al., 2024; also see Agrawal et al., 2024; Kiciman et al., 2023)?
8
Without question, LLMs seem to manifest sparks of intelligence. But intelligence is not simply
memorization, the ability to restate or paraphrase information in various ways. We argue that LLMs appear
intelligent because they capitalize on the fact that the same thing can be stated, said, and represented in
indefinite ways This is readily illustrated by the fact that the revolutionary breakthrough that gave rise to
LLMs—the transformer architecture—was developed in the context of language translation (Vaswani et al.,
2017). In an important sense, LLMs can be seen as “translation generalized.” They represent a generalized
technology for translating one way of saying things into another way of saying the same thing. Translation after all is an
effort to represent and accurately mirror something in a different way—to represent the same thing in a
different language or with a different set of words, or more abstractly: to represent the same thing in a
different format. LLMs serve this representational and mirroring function remarkably well. This
representational and mirroring function from language to language is generalized to a process that takes one
way of saying something and generates another way of saying the same thing. Stochastic next-word prediction
using conditional probabilities—based on the weights and parameters derived from vast training datasets—
allows for surprisingly rich combinatorial outputs. The learning of the LLM is embodied in the relationships
found between words which are sampled to enable stochastic generativity, where the outputs mirror past
inputs. With vast data, an LLM is good at probabilistically and fluently predicting the next word. But, as we
8
AI can, of course, be (and has been) a powerful aid in scientific discovery. For example, modern AI techniques have
analyzed astronomical datasets far more quickly and accurately than humans, helping identify new planets and celestial
phenomena, as seen with Kepler's laws of planetary motion. Similarly, DeepMind's AlphaFold has revolutionized protein
structure prediction, a critical task for understanding biological processes and developing new medications (e.g., Jumper et
al., 2021). Yet it is important to state that in both of these cases AI is not somehow independently doing the science by
forming hypotheses or conducting experiments, but that these hypotheses were provided by human scientists in the form
of patterns and reward functions, respectively. AI has significantly accelerated research by enabling scientist to process
large datasets and uncover novel patterns, allowing scientists to focus on hypothesis generation and experimental design
rather than “number crunching.”

13
will discuss, the fluency with which LLMs seem to predict and generate outputs dupes us into seeing them as
intelligent—as if they are engaging in far more than mere mirroring or translation.
Before revisiting our question of whether an AI like an LLM could actually originate novelty or
engage in some form of forward-looking decision making, it is worth highlighting metaphorical similarities
between AI and cognitive architectures based on prediction. For example, consider a cognitive approach like
predictive processing (Pezzulo, Parr and Friston, 2024: which shares broad similarities with active inference,
the free energy principle, the Bayesian brain, and predictive coding). At a high level, both LLMs and predictive
processing seek to engage in a similar process, namely, error minimization and iterative optimization, where
the systems are essentially navigating a high-dimensional space to find a state that minimizes both error and
surprise. LLMs learn from the training data and predictive processing learns from its environment (cf. Hohwy,
2020). LLMs aim to reduce the difference between their probabilistic predictions (the next word in a sentence)
and the actual outcomes (the real next word), thereby improving their accuracy. Predictive processing, as a
cognitive theory, posits that the brain continuously predicts sensory input and minimizes the error between its
predictions and actual sensory input. The capability of each to predict—whether a word or a perception—is a
function of past inputs. Large language models seek to predict the most likely next word based on training
data, and active inference seeks to predict the most likely next percept or action. Both approaches are wildly
conservative (tied to past data) as they seek to reduce surprise—or to engage in prediction as “error
minimization” (Hohwy, 2013).
9
Back-propagation, a fundamental mechanism in training neural networks, and
the concept of error minimization in predictive processing (Friston et al., 2009), share a broad conceptual
similarity in that both involve iterative adjustments to minimize some form of error or discrepancy. Both
generate a prediction based on past inputs. Both back-propagation and error minimization in predictive
processing involve adjusting an internal model (neural network weights in AI, and hierarchical brain models in
neuroscience) to reduce error (or, in machine learning terms, minimize the loss function).
With this architecture—focused on error minimization and surprise reduction—can an LLM or any
prediction-oriented cognitive AI truly generate some form of new knowledge? Beyond memorizing,
9
This leads to the problem of surprise and the famous “dark room” problem of predictive processing. For an attempt to
deal with this, see Clark, 2018.

14
translating, restating, or mirroring the text with which it has been trained, can an LLM generate new
knowledge?
We do not believe LLMs or input-output based cognitive systems can do this, at least not beyond
random flukes that might emerge due to their stochastic nature.
10
There is no forward-looking mechanism or
unique causal logic built into these systems. It is important to clearly delineate why this is the case, as some
argue and anticipate that LLMs will replace human decision makers in uncertain contexts like strategy and
even science itself. For example, Csaszar et al (2024) argue that “the corpora used to train LLMs encompass
information necessary for [strategic decision making], including consumer preferences, competitor
information, and strategy knowledge” and point to how an AI can use various decision making tools to
generate business plans and strategy (Csaszar et al., 2024). And Manning et al (2024) even argue that LLMs will
“automate” social science given their seeming ability to generate hypotheses and causal models, including testing
them (also see Lu et al., 2024).
These claims are vastly overstated. One way to think about this is that a prediction-oriented AI like an
LLM can essentially be seen as possessing “Wikipedia-level” knowledge. On any number of topics (if
contained in the training data), an LLM can summarize, represent, and mirror the words they have
encountered in various different and new ways. On any given topic—again, if sufficiently represented in the
training data—an LLM can generate indefinite numbers of coherent, fluent, and well-written Wikipedia articles
by drawing on the conditional probabilities it has learned. But just as a subject-matter expert is unlikely to
learn anything new about their specialty from a Wikipedia article within their domain of expertise, so an LLM
is unlikely to somehow bootstrap knowledge beyond the combinatorial possibilities of the word associations it
has encountered in the past. It has no forward-looking mechanism for doing so.
There is also good evidence to suggest that when an LLM encounters (is prompted with) a reasoning
task, it merely reproduces the linguistic answers (about reasoning) it has encountered in the training data rather
than engaging in any form of actual, on-the-fly reasoning. If the wording of a reasoning task—like the Wason
selection task or the Monty Hall problem—is changed only slightly, LLM performance declines significantly
below human performance, where the mistakes of the LLM are glaringly obvious to humans (e.g., Hong et al.,
10
Though we of course recognize that there is significant disagreement on this point (for example, related to AI versus
human creativity, see Franceschelli and Musolesi, 2023).

15
2024). LLMs are not meaningfully engaged in any form of real-time reasoning (as assumed by Lu et al., 2024
and Manning et al., 2024). Rather, they are merely repeating the word structures associated with reasoning,
which they have encountered in the training data. This effect can also be shown empirically, as training LLMs
on their past output leads to a rapid decline in performance and even their “collapse” (Shumailov et al., 2024).
Importantly, LLMs memorize and regurgitate the words associated with reasoning but do not engage in on-
the-fly reasoning of any sort.
11
This is why Francois Chollet (2019) has created the “abstraction and reasoning
corpus,” as a challenge or test to see if an AI system can actually solve new problems (that is, problems it has
not encountered in its training data), without merely resorting to memorized answers and solutions
encountered in the past (which captures the present state of AI systems including LLMs).
12
That said, our goal is not to dismiss the remarkable feats of LLMs, nor other forms of AI or
applications of machine learning. The fact that an LLM can outperform most humans in varied types of tests
and exams is remarkable (Achiam et al., 2023). But this is because it has encountered this information,
memorized it, and is able to repeat it in fluent ways. An LLM essentially has a superhuman capacity for
memorization, and an ability to summarize memorized word structures in diverse ways. In all, certainly the
idea of LLMs as “stochastic parrots” or “glorified auto-complete” (Bender et al., 2021) underestimates their
ability. But equally, ascribing LLMs the ability to actually reason and generate new knowledge vastly
overestimates their ability. LLMs are essentially powerful and creative “imitation engines” in stochastically and
probabilistically assembling words, though not linguistically innovative compared to children (see Yiu, Kosoy,
and Gopnik, 2023). The idea that LLMs somehow generate new-to-the-world knowledge—or feature
something like human consciousness—seems to be a significant stretch (though, see Butlin et al., 2023;
Hinton, 2023). In sum, the generativity of these models is a type of “lower-case-g” generativity that shows up
in the form of the unique sentences that creatively summarize and repackage existing knowledge.
11
The capabilities of AI are of course rapidly evolving and future developments are hard to anticipate. In this paper we
have discuss AI in its past and current state—comparing it with human cognition—rather than speculate about what AI
might be capable of in the future. It might be that the forms of human reasoning and cognition that we emphasize (and
claim, in this paper, to be unique to humans) could be mimicked or replicated by future AI systems.
12
Beyond the ability of a human or AI to solve previously-unseen, new problems (which is the focus of Chollet’s ARC
challenge), there is an even higher form of intelligence in being able to specify and formulate problems in the first place
(Felin and Zenger, 2017). This is a skill that is manifest in humans—in theorizing and causal reasoning—but not evident
in AI. As we discuss later, it was the ability of the Wright brothers to formulate the right problems (lift, propulsion, and
steering) that enabled them to then identify the right data, specific forms of experimentation and relevant solutions.

16
To illustrate the problem of generating something novel—like new knowledge—with an LLM,
imagine the following thought experiment. Imagine an LLM in the year 1633, where the LLM’s training data
incorporates all the scientific and other texts published by humans to that point in history. If the LLM were
asked about Galileo’s heliocentric view, how would it respond? Since the LLM would probabilistically sample
from the association and correlation-based word structure of its vast training data—again, everything that has
so far been written (including all the scientific writings about the structure of the cosmos)—it would only
restate, represent, and mirror the accumulated scientific consensus. The training dataset for the LLM would
overwhelmingly feature texts with word structures supporting a geocentric view, in the form of the work of
Aristotle, Ptolemy, and many others. Ptolemy’s careful trigonometric and geometric calculations, along with
his astronomic observations, would be included in support of a geocentric view, as represented in the many
texts that would have summarized the geocentric view (like de Sacrobosco’s popular textbook De saphera
mundi). These texts would feature word associations that highlight how the motions and movements of the
planets could be predicted with remarkable accuracy with the predominant geocentric view. The evidence—as
inferred from the repeated word associations found in the training data—would overwhelmingly be against
Galileo. LLMs do not have any way of accessing truth (for example through experimentation or
counterfactuals) beyond mirroring and restating what is found in the text.
Even if alternative or heretical views were included in the training data (like the work of Copernicus,
even though his work was largely banned), the logic of this work would be dwarfed by all the texts and
materials that supported the predominant geocentric paradigm.
13
The overwhelming corpus of thousands of
years of geocentric texts would vastly outweigh Galileo’s view, or anything supporting it. . An LLM’s model of
truth or knowledge is solely statistical, relying on frequency and probability. Outputs are influenced by the
frequency with which an idea is mentioned in the training data, as reflected by associated word structures. For
example, the frequency with which the geocentric view has been mentioned, summarized, and discussed in the
training data necessarily imprints itself onto the output of the LLM as truth. As the LLM has no actual
grounding in truth—beyond the statistical relationships between words—it would say that Galileo’s view and
belief is delusional and in no way grounded in science.
13
While Copernicus’s On the Revolution of the Heavenly Spheres was published in 1543, the theory contained within the book
represented a fringe view within science. Given the fringe nature of the Copernican view, his book was withdrawn from
circulation and eventually censored (Gingerich and Maclachlan, 2005).

17
A neural network like an LLM might in fact include any number of delusional beliefs, including
beliefs that turned out to eventually be correct (like Galileo’s), but also beliefs that objectively were (and still are)
delusional. Ex ante there is no way for an LLM to arbitrate between the two. For example, the eminent
astronomer Tyco Brahe made and famously published extensive claims about astrology, the idea that celestial
bodies and their movement directly impact individual human fates as well as political and other affairs. His
astrological writings were popular not just among some scientists, but also among the educated elite. A
hypothetical LLM (in 1633) would have no way of arbitrating between Galileo’s (seeming) delusions about
heliocentrism nor Brahe’s (actual) delusions about astrology. Our hypothetical LLM would be far more likely
to have claimed that Brahe’s astrological claims are true than that Galileo’s argument about heliocentrism is
true. The LLM can only represent and mirror the predominant and existing conceptions—in this case, the
support for the geocentric view of the universe—it finds in the frequencies and statistical association of words
in its training data.
In sum, it is important to recognize that the way an LLM gets at truth and knowledge is via a
statistical exercise of finding more frequent mentions of (hopefully) a true claim (in the form of statistical
associations between words) and less frequent mentions of a false claim. LLM outputs are probabilistically
drawn from the statistical associations of words it has encountered while being trained. When an LLM makes
truthful claims, these are an epiphenomenon of the fact that true claims happened to have been made more
frequently. There is no other way for the LLM to assess truth, or to reason. Truth—if it happens to emerge—
is a byproduct of statistical patterns and frequencies rather than from the LLM developing an intrinsic
understanding of—or ability to bootstrap or reason—what is true or false in reality.
Some LLMs have sought to engineer around the problem of their frequency-based and probabilistic
approach by creating so-called “mixture of experts” models where the outputs are not simply the “average”
result of “outrageously” large neural networks but can be fine-tuned toward some forms of expertise (Du et
al., 2023; Shazeer et al., 2017). Another approach is retrieval-augmented generation, which uses the general
linguistic abilities of the LLM but limits the data used for prediction to a confined and pre-selected set of
sources (Lewis et al., 2020). Furthermore, ensemble approaches—which combine or aggregate diverse
architectures or outputs—have also been developed (Friedman and Popescu, 2008; Russell and Norvig, 2022).
However, even here the outputs would necessarily also be reflective of what any particular experts have said

18
within the training data, rather than any form of forward-looking projection or on-the-fly causal reasoning on
the part of the LLM. This problem is further compounded in situations that are characterized by high levels of
uncertainty and novelty (like many forms of decision making), where the idea of expertise or even bounded
rationality is hard to specify given an evolving and changing world (Felin, Kauffman, Koppl and Longo,
2014).
14
Finally, it is critically important to keep in mind that the inputs of any LLM are past human inputs,
and therefore outputs also roughly represent what we know so far. Inherently an LLM cannot go beyond the
realms covered by the inputs. There is no mechanism to somehow bootstrap forward-looking beliefs about
the future—nor causal logic or knowledge—beyond what can be inferred from the existing statistical
associations and correlations found in the words in the training data.
The Primacy of Data versus Data-Belief Asymmetry
The central problem we have highlighted, so far, is that learning by machines and AI is necessarily backward-
looking and imitative. Again, this should not be read as a critique of these models, rather, merely as a
description of their structural limits. While they are useful for many things, an AI model—like an LLM—is
not able to generate new knowledge or solve new problems. An LLM does not reason. And an LLM has no
way of postulating beyond what it has encountered in its training data. Next we extend this problem to the
more general emphasis on the primacy of data within both AI and cognitive science. Data itself of course is
not the problem. Rather, the problem is that data is used in theory-independent fashion (Anderson, 2007). To
assure the reader that we are not caricaturing existing AI-linked models of cognition by simply focusing on
LLMs, we also extend our arguments into other forms of cognitive AI.
The general emphasis on minds and machines as input-output devices places a primary emphasis on
data. This suggests a model where data—such as cues, stimuli, text, images—essentially are “read,” learned
and represented by a system, whether it is human or computational one. The world (any large corpus of
14
A deeper issue here is the “frame problem” (McCarthy and Hayes, 1969) and the implications it has for not just
artificial intelligence but also for understanding decision making under uncertainty (Felin, Kauffman, Koppl, and
Longo, 2014). In the latter context, the frame problem refers to the challenge of determining which aspects of a
situation are relevant or irrelevant when making decisions (cf. Felin and Koenderink, 2022). The frame problem
highlights the difficulty of specifying all the possible consequences of an action in a dynamic environment,
particularly when only some aspects of the world are affected by the action while others remain unchanged. In
decision-making under uncertainty, the frame problem underscores the complexity of reasoning about the
implications of actions when the system must account for numerous potential variables and outcomes, often leading
to difficulties in efficiently processing information and making reliable predictions.

19
images, text, or environment) has a particular statistical and physical structure, and the goal of a system is to
accurately learn from it and reflect it. This is said to be the very basis of intelligence. As put by Poldrack, “any
system that is going to behave intelligently in the world must contain representations that reflect the structure of the
world” (2021: 1307, emphasis added; cf. Yin, 2020). Neural network-based approaches and machine learning—
with their emphasis on bottom-up representation—offer the perfect mechanism for doing this, because they
can “learn directly from data” (Lansdell and Kording, 2019; also see Baker et al., 2022). Learning is data-
driven.
15
Of course, cognitive systems may not be able to learn perfectly, but an agent or machine can
“repeatedly interact with the environment” to make inferences about its nature and structure (Binz et al.,
2023). This is the basis of “probabilistic models of behavior” which view “human behavior in complex
environments as solving a statistical inference problem” (Tervo, Tenenbaum, and Gershman, 2021).
16
Bayesian cognition also posits that learning by humans and machines can be understood in terms of
probabilistic reasoning about an environment, as captured in Bayesian statistical methods (e.g., Griffiths et al.,
2010). This framework conceptualizes sensory inputs, perceptions, and experiential evidence as data, which are
continuously acquired from the environment and then used to update one’s model of the world (or of a
particular hypothesis). The cognitive process involves sampling from a probability distribution of possible
states or outcomes, informed by incoming data. Crucially, Bayesian and related approaches to cognition
emphasize the dynamic updating of beliefs—where prior knowledge (a prior) is integrated with new evidence
to revise beliefs (posterior), in a process mathematically described by the Bayesian formula (Pinker, 2021). This
iterative updating, reflecting a continual learning process, acknowledges and quantifies uncertainty, framing
understanding and decision making as inherently probabilistic. This probabilistic architecture is (very broadly)
also the basis of large swaths of AI and the cognitive sciences.
It is worth reflecting on the epistemic stance—or underlying theory of knowledge—that is presumed
here. Knowledge is traditionally defined as justified belief, and belief is justified by data and evidence. As
15
The problems with this approach have not just been discussed by us. For example, Yin, 2020 for related points in the
field of neuroscience.
16
In the context of machine learning, it is interesting that while the approach is said to be “theory-free” (to learn directly
from data), nonetheless the architects of these machine learning systems are making any number of top-down decisions
about the design and architecture of the algorithms, and how the learning occurs and the types of outputs that are valued.
These decisions all imply mini-theories of what is important—a point that is not often recognized (cf. Rudin, 2019). This
involves obvious things like choice of data, model architecture, and hyperparameter settings, but also loss functions and
metrics, regularization and generalization techniques, valued outputs, type of human reinforcement, and so forth.

20
suggested by Bayesian models, we believe or know things to the extent to which we have data and evidence
for them (Pinker, 2021). Beliefs should be proportionate to the evidence at hand, because agents are better off
if they have an accurate representation or conception of their environment and the world (e.g., Schwöbel et al.,
2018).
17
Knowledge can be seen as the accumulated inputs, data and evidence that make up our beliefs. And
the strength or degree of any belief should be symmetrical with the amount of supporting data, or put differently,
the weight of the evidence (Pinker, 2021; also see Dasgupta et al., 2020; Griffin and Tversky, 1992; Kvam and
Pleskac, 2016). This is the foundation of probabilistic models of cognitive systems. These approaches focus on
“reverse-engineering the mind”—from inputs to outputs—and they “[forge] strong connections with the
latest ideas from computer science, machine learning, and statistics” (Griffiths et al., 2010: 363). Overall, this
represents a relatively widely-agreed upon epistemic stance, which also matches an input-output-oriented
“computational theory of mind” (e.g., Rescorla, 2015) where humans or machines learn “through repeated
interactions with an environment”—without “requiring any a priori specifications” (Binz et al., 2023). One
way to summarize the above literature is that there needs to be a symmetry between one’s belief and the
corroborating data. A rational decision maker will form (and weight) their beliefs about any given thing by
taking into account the available data and evidence.
But what about “edge cases?” That is, what about situations where an agent correctly takes in all the
data and evidence yet somehow turns out to be wrong? Models based on rational information processing do
not offer a mechanism for explaining change or new knowledge, nor an explanation of situations where data
and evidence-based reasoning might lead to poor outcomes (cf. Felin and Koenderink, 2022). Furthermore,
while learning-based models of knowledge enable “belief updating”—based on new evidence—there is no
mechanism for explaining where new data comes from, or what data should be considered as relevant and
what data should be ignored. And what if the data and evidence are contested? This is a particularly significant
problem in contexts that feature rampant uncertainty, including any type of forward-looking decision making
and scientific reasoning.
Explaining the emergence of novelty and new knowledge is highly problematic for computational,
input-output models of cognition that assume what we call data-belief symmetry. The basis of knowledge is the
quest for truth (Pinker, 2021), which is focused on existing evidence and data. But we argue that data-belief
17
The predictive processing and active inference approach has many of these features (e.g., Parr and Friston, 2017)

21
asymmetry in fact is essential for the generation of new knowledge and associated decision making. The
existing literature in the cognitive sciences has focused on one side of the data-belief asymmetry, namely its
downside—the negative aspects of data-belief asymmetry (e.g., Kunda, 1990; Scheffer et al., 2022). This
downside includes all the ways in which humans persist in believing something despite seemingly clear evidence
to the contrary (Pinker, 2021). This includes a large literature which focuses on human biases in information
processing—the suboptimal and biased ways that humans process, perceive, and use data and fail to
appropriately update their beliefs. This is evident in the vast literatures that focus on various data-related
pathologies and biases, including motivated reasoning, confirmation bias, selective perception and sampling,
and the availability bias. The emphasis on erroneous beliefs and human bias has powerfully influenced how we
think about human nature and decision making within various social and economic domains (e.g., Benabou
and Tirole, 2016; Bordalo et al., 2024; Chater, 2018; Gennaioli and Shleifer, 2018; Kahneman, 2011;
Kahneman et al., 2021).
But what about the positive side of data-belief asymmetry? What about situations where beliefs
appear delusional and distorted—seemingly contrary to established evidence and facts—but where these
beliefs nonetheless turn out to be correct? Here we are specifically talking about beliefs that may outstrip,
ignore, and go beyond existing evidence. Forward-looking, contrarian views are essential for the generation of
novelty and new knowledge. Due to the statistical and past-oriented nature of AI-based computational and
cognitive systems (focused on correlations, associations, and averages from past data), they are not able to
project or reason forward in contrarian ways, given the implicit insistence on symmetry between data and
beliefs. That said, notice that—as we will discuss—our focus on data-belief asymmetries is not somehow data-
independent or untethered from reality. Rather, this form of data-belief asymmetry is forward-looking, where
beliefs and causal reasoning enable the identification of new data and experimental interventions, and the
eventual verification of beliefs that previously were seen as the basis of distortion or delusion.
To offer a practical and vivid illustration of how data-belief symmetry can be problematic, consider
the beliefs that were held about the plausibility of “heavier-than-air” human, powered, and controlled flight in
the late 1800s and the early 1900s. (We introduce this example here and revisit it throughout the remainder of
the manuscript.) To form a belief about the possibility of human powered flight—or to even assign it a
probability—we would first want to look at the existing data and evidence. So, what was the evidence for the

22
plausibility of human powered flight at the time? The most obvious datapoint at the time was that human
powered flight was not a reality. This alone, of course, would not negate the possibility. So, one might want to
look at all the data related to human flight attempts to assess its plausibility. Here we would find that humans
have tried to build flying machines for centuries, and flight-related trials had in fact radically accelerated during
the 19th century. All of these trials of flight could be seen as the data and evidence we should use to update our
beliefs about the implausibility of flight. All of the evidence clearly suggested that a belief in human powered
flight was delusional. A delusion can readily be defined as having a belief contrary to evidence and reality
(Pinker, 2021; Scheffer, 2022): a belief that does not align with accepted facts. In fact, the DSM-4/5—the
authoritative manual for mental disorders—defines delusions as “false beliefs due to incorrect inference about
external reality” or “fixed beliefs that are not amenable to change in light of conflicting evidence.”
Notice that many people at the time—naïvely, it was thought—pointed to birds as evidence for the
belief that humans might also fly. This was a common argument.
18
But the idea that bird flight somehow
provided hope and evidence for the plausibility of human flight was seen as delusional by scientists and put to
bed by the prominent scientist Joseph LeConte. He argued that flight was “impossible, in spite of the testimony
of birds” (1888: 69). Like a good scientist and Bayesian, LeConte appealed to the data to support his claim. He
looked at bird species—those that fly and those that do not—and concluded “there is a limit of size and
weight of a flying animal.” According to LeConte, weight was the critical determinant of flight. With his data,
he clearly pointed out that no bird above the weight of 50 pounds is able to fly, and thus concluded that
therefore humans cannot fly. After all, large birds like ostriches and emus are flightless. And even the largest
flying birds, he argued—like turkeys and bustards—“rise with difficulty” and “are evidently near the limit”
(LeConte, 1888; 69-76). Flight and weight are correlated. To this, Simon Newcomb—one of the foremost
astronomers and mathematicians of the time—added that “the most numerous fliers are little insects, and the
rising series stops with the condor, which, though having much less weight than a man, is said to fly with
difficulty when gorged with food” (1901: 435).
18
As captured by a prominent engineer at the time: “There probably can be found no better example of the speculative
tendency carrying man to the verge of the chimerical than in his attempts to imitate the birds, or no field where so much
inventive seed has been sown with so little return as in the attempts of man to fly successfully through the air” (Melville,
1901: 820).

23
The emphasis that LeConte placed on the weight of birds to disprove the possibility of human
powered flight highlights one of the problems with data and belief updating based on evidence. It is hard to
know what data and evidence might be relevant for a given belief or hypothesis. The problem is—as succinctly
put by Polanyi—that “things are not labeled evidence in nature” (1957: 31). Is the fact that small birds can fly
and large birds cannot fly relevant to the question of whether humans can fly? What is the relevant data and
evidence in this context? Did flight have something to do with weight, size, or with other features like wings?
Did it have something to do with the “flapping” of wings (as Jacob Degen hypothesized)? Or did it have
something to do with wing shape, wing size, or wing weight?
19
Perhaps feathers are critical to flight. In short,
it is hard to know what data might be relevant and useful.
Of course, not all our beliefs are fully justified in terms of direct empirical data that we ourselves have
verified. We cannot—nor would we want to—directly verify all the data and observations that underlie our
beliefs and knowledge. More often than not, for our evidence we rightly rely on the expertise, beliefs, or
scientific arguments of others, which serve as “testimony” for the beliefs that we hold (Coady, 1992;
Goldman, 1999). The cognitive sciences have also begun to emphasize this point. Bayesian and other
probabilistic models of cognition have introduced the idea of “the reliability of the source” when considering
what data or evidence one should use to update beliefs and knowledge (e.g., Hahn, Merdes, and Sydow, 2018;
Merdes et al., 2021). This approach recognizes that not all data and evidence is equal. Who says what does
matter. The source of evidence needs to be considered. For example, scientific expertise and consensus are
critically important sources of beliefs and knowledge.
This is readily illustrated by heavier-than-air flight. So, what might happen if we weight our beliefs
about the plausibility of human flight by focusing on reliable, scientific sources and consensus? In most
instances, this is a rational strategy. However, updating our belief on this basis when it comes to heavier-than-
air flight during this time period would further reinforce the conclusion that human powered flight was
delusional and impossible. Again, scientists like LeConte and Newcomb argued that flight was impossible by
pointing to seemingly conclusive data and evidence. And not only should we update our belief based on this
19
Even if LeConte happened to be right that the delimiting factor for flight was weight (which of course is not the case),
he also did not take into account—or more likely, was not aware of—findings related to prehistoric fossils. For example,
in the mid and late 1800s, scientific journals reported about the discovery of prehistoric fossils—pelagornis sandersi—
with wingspans of up to 20-24 feet and conjectures of a weight up to 130 pounds.

24
evidence, but we should also further weight that evidence by the fact that it came from a highly prominent
scientist with seemingly relevant knowledge in this domain. LeConte for example became the eventual
President of the leading scientific association in the United States (the American Association for the
Advancement of Science). And LeConte was scarcely alone. He was part of a much broader scientific
consensus that insisted on the impossibility of human powered flight. For example, Lord Kelvin emphatically
argued—while serving as President of the British Royal Society—that “heavier-than-air flying machines are
impossible.” This is ironic, as Kelvin’s scientific expertise in thermo- and hydrodynamics, the behavior of
gases under different conditions (and other areas of physics) in fact features practical implications that turned
out to be extremely relevant for human powered flight. And the aforementioned, prominent mathematician-
astronomer Simon Newcomb (1901) also argued in the early 1900s—in his article, “Is the airship coming?”—
that the impossibility of flight was a scientific fact, as there was no combination of physical materials that
could be combined to enable human flight (for historical details, see Anderson, 2004; Crouch, 2002).
The question then is: how does someone still—despite seemingly clear evidence and scientific
consensus—hold onto a belief that appears delusional? In the case of human flight, the data, evidence, and
scientific consensus were firmly against the possibility. No rational Bayesian should have believed in heavier-
than-air flight. Again, the evidence against it was not just empirical (in the form of LeConte’s bird and other
data) and based on science and scientific consensus (in the form of Kelvin and Newcomb’s physics-related
arguments), but it also was observationally salient. Many aviation pioneers not only failed and were injured,
but some also died. For example, in 1896, the German aviation pioneer Otto Lilienthal died while attempting
to fly, a fact that the Wright brothers were well acquainted with (as they subsequently studied Lilienthal’s
notebooks and data). And in 1903—just nine weeks before the Wright brothers succeeded—the scientist
Samuel Langley failed spectacularly in his attempts at flight, with large scientific and lay audiences witnessing
the failures. Reflecting on recent flight attempts (including Langley’s prominent failure), the editorial board of
the New York Times (1903) estimated that it would take the “combined and continuous efforts of
mathematicians and mechanicians from one million to ten million years” to achieve human powered flight.
Now, we have of course opportunistically selected a historical example where a seemingly delusional
belief—one that went against existing data, evidence, and scientific consensus—turned out to be correct.
Cognitive and social psychologists often engage in the “opposite” exercise where they retrospectively point to

25
situations where humans doggedly persist in holding delusional beliefs—despite clear evidence against those
beliefs—due to biased information processing, selective perception or biased sampling of data (Festinger et al.,
1956; Kahneman, 2011; Kunda, 1990; Pinker, 2021; though see Anglin, 2019). Conspiracy theories provide a
frequently discussed example of beliefs that seem impervious to evidence (Gagliardi, 2023; Rao and Greve,
2024). Economists more generally have highlighted how humans can be “resistant to many forms of evidence,
with individuals displaying non-Bayesian behaviors such as not wanting to know, wishful thinking, and reality
denial” (Benabou and Tirole, 2016: 142).
20
Of course, some beliefs truly are delusional. But others—like
flight—may merely appear delusional.
We think that the other side of beliefs—beliefs that presently might appear delusional (beliefs that “go
against the evidence”) and are seemingly driven by motivated reasoning, but turn out to be correct—also need
to be addressed. Our example of flight offers an instance of a far more generalizable process where data-belief
asymmetries are essential for the emergence of novelty and new knowledge. Heterogenous beliefs and data-
belief asymmetries are the lifeblood of new ideas, new forms of experimentation, and new knowledge—as we
discuss next. Furthermore, this turns out to have important implications for computation-oriented forms of
AI and cognition.
THEORY-BASED CAUSAL LOGIC AND COGNITION
Building on the aforementioned data-belief asymmetry, next we discuss the cognitive and practical process by
which humans engage in forward-looking theorizing and causal reasoning that enables them to, in essence, go
“beyond the data”—or more specifically, to go beyond existing data, to experiment and produce new data and
novelty. We specifically emphasize how this form of cognitive and practical activity differs from
computational, data-driven, and information processing-oriented forms of cognition—the hallmarks of AI and
computational forms of cognition—and allows humans to “intervene” in the world in forward-looking
20
In economics there has similarly been an emphasis on how beliefs lead to negative outcomes. For example, Gennaioli
and Shleifer’s (2018; also see Benabou and Tirole, 2016) “theory of beliefs” focuses on beliefs that turn out to be
delusional and are the result of poor judgment, biased information processing, and selective perception. In a related vein,
Bordalo et al (2023) largely argue that humans are poor statisticians—selectively attending to and inappropriately
weighting evidence and feedback—leading to suboptimal outcomes. Here we highlight discrepant beliefs that appear
delusional and highly biased—to some, or even a majority, of actors—in the present, but turn out to be correct. Importantly,
our theory is one of belief asymmetry rather than bounded rationality, bias, or information asymmetry (cf. Felin,
Gambardella, Novelli and Zenger, 2024).

26
fashion. Approaches that focus on data-driven prediction take and analyze the world as it is, without
recognizing the human capacity to intervene (Pearl and Mackenzie, 2018)—and to realize beliefs that presently
seem implausible due to the apparent lack of data and evidence. We extend the example of heavier-than-air
flight to offer a practical illustration of this point, in an effort to provide a unique window into what we think
is a far more generalized and ubiquitous process.
Our foundational starting point—building on Felin and Zenger (2017)—is that cognitive activity is a
form of theoretical or scientific activity.
21
That is, humans generate forward-looking theories that guide their
perception, search, and action. As noted by Peirce, the human “mind has a natural adaptation to imagining
correct theories of some kinds (…). If man had not the gift of a mind adapted to his requirements, he could
not have acquired any knowledge” (1957: 71). As highlighted by our example of language, the meager
linguistic inputs of a child can scarcely account for the vast outputs, thus pointing to a human generative
capacity to theorize. The human capacity to theorize—to engage in novel problem solving and
experimentation—has evolutionary origins and provides a highly plausible explanation for evolutionary leaps
and the emergence of technology (Felin and Kauffman, 2023).
Importantly, theory-based cognition enables humans to do things, to experiment. This is also the basis
of the so-called “core knowledge” argument in child development (e.g., Carey and Spelke, 1996; Spelke et al.,
1992). Humans develop knowledge like scientists, through a process of hypotheisizing, causal reasoning, and
experimentation. While computational approaches to cognition focus on the primacy of data and
environmental inputs, a theory-based view of cognition focuses on the active role of humans in not just
learning about their surroundings but also their role in actively generating new knowledge (Felin and Zenger,
2017). Without this active, generative, and forward-looking component of theorizing, it is hard to imagine how
knowledge would grow—whether we are talking about practical or scientific knowledge. This is nicely
captured in the title of an article in developmental psychology: “If you want to get ahead, get a theory”
21
A central aspect of this argument—which we unfortunately do not have room to explicate in this paper—is that
humans are biological organisms. The theory-based view builds on the idea that all organisms engage in a form of forward-
looking problem solving. A central aspect of this approach is captured by the biologist Rupert Riedl who argued that
“Every conscious cognitive process will show itself to be steeped in theories; full of hypotheses” (1984: 8). To see the
implications of this biological argument on human cognition—particularly in comparison to statistical and computational
approaches—see Felin and Koenderink (2022; also see Roli et al., 2022; Jaeger et al., 2024). For the embodied aspects of
human cognition, see Mastrogiorgio et al., 2022.

27
(Karmiloff-Smith and Inhelder, 1974). This also echoes Kurt Lewin’s maxim, “there is nothing as practical as a
good theory” (1943: 118). The central point here is that theories are not just for scientists. Theories are
pragmatically useful for anyone seeking to understand and influence their surroundings—theories help us do
things. Theorizing is a central aspect of human cognitive and practical activity. Thus, as argued by Dewey, “the
entities of science are not only from the scientist” and “individuals in every branch of human endeavor should
be experimentalists” (1916: 438-442). We build on this intuition and extend it into new and novel domains,
along with contrasting it with AI-informed models of cognition.
The theory-based view—in the context of decision making and strategy—extends the above logic and
emphasizes the importance of theorizing and theories in economic contexts, with widespread implications for
cognition (Felin and Zenger, 2017). The central idea behind the theory-based view is that economic actors can
(and need to) develop unique, firm-specific theories. Theories do not attempt to map existing realities, but
rather to generate unseen future possibilities—and importantly, theories suggest causal interventions
(experiments and actions that need to be taken) that enable the realization of these possibilities.
Theories can also be seen as a mechanism for “hacking” competitive factor markets (cf. Barney, 1986),
enabling economic actors to see and search the world differently. Awareness for new possibilities is cognitively
developed top-down (Felin and Koenderink, 2022). Theories also have central implications for how to
efficiently organize or govern the process of realizing something that is new (Wuebker et al., 2023). This
approach has been empirically tested and validated (e.g., Agarwal et al., 2023; Camuffo et al., 2021; Novelli and
Spina, 2022), including important theoretical extensions (e.g., Ehrig and Schmidt, 2022; Zellweger and Zenger,
2023).
22
The practical implications of the theory-based view have also led to the development of managerial
tools to assist startups, economic actors, and organizations in creating economic value (Felin, Gambardella,
and Zenger, 2021).
Our goal in this section of the paper is not to exhaustively review the theory-based view. Rather, our
goal now is to further build out the cognitive and practical aspects of the theory-based view, with a specific
emphasis on causal reasoning and how this contrasts with backward-oriented, data-focused approaches to AI and
cognition. We highlight how the human capacity for theorizing and causal reasoning differs from AI’s
22
There are parallel literatures in strategy that focus on mental representations (e.g., Csaszar and Levinthal, 2016) and
forward-looking search and representation (e.g., Gavetti and Levinthal, 2000; also see Gans, Stern and Wu, 2019).

28
emphasis on data-driven prediction. A theory-based view of cognition allows humans to intervene in the
world, beyond the given data—not just to process, represent, or extrapolate from existing data. Theories
enable the identification or generation of nonobvious data and new knowledge through experimentation. This
differs significantly from the arguments and prescriptions suggested by computational, Bayesian, and AI-
inspired approaches to cognition. It is important to carefully establish these differences, as AI-based and
computational approaches—as extensively discussed at the outset of this paper—are said to be superior to
human judgment and cognition (e.g., Kahneman, 2018).
Cognition: Data-Belief Asymmetry Revisited
Heterogeneous beliefs provide the initiating impetus for theory-based causal reasoning and cognition. From
our perspective, for beliefs to be a relevant concept for understanding cognition and decision making, beliefs
do not necessarily—in the first instance—need to be based on data. We are specifically interested in forward-
looking beliefs, beliefs that presently lack evidence or even go against existing data, but which might turn out to be
true. Forward-looking beliefs, then, are more in search of data rather than based on existing data. At the
forefront of knowledge, data is an outcome of beliefs—coupled with causal reasoning and experimentation
(which we discuss in the next section)—rather than new knowledge being a direct outcome of existing data.
The problem is that it is hard to ex ante distinguish between beliefs that indeed are delusional versus those that simply
are ahead of their time. Data-belief asymmetry is critical in situations where data “lags” belief (or where data
might presently be non-existent), that is, situations where the corroborating data simply has not yet been
identified, found, or experimentally generated. In many cases, beliefs do not automatically verify themselves.
Rather, more often than not, they require some form of targeted intervention, action, and experimentation.
The search for data in support of an uncommon, contrarian or discrepant belief necessarily looks like
irrational motivated reasoning or confirmation bias (Kunda, 1990; cf. Hahn and Harris, 2014). To briefly
illustrate, Galileo’s belief in heliocentrism went against the established scientific data and consensus, and even
plain common sense. Geocentric conceptions of the earth’s place in the universe were observationally well-
established. And they were successful: they enabled precise predictions about the movement of planets and
stars. Even everyday observation verified that the earth does not move and that the sun seemingly circles the
earth. Galileo’s detractors essentially argued that Galileo was engaged in a form of biased, motivated reasoning

29
against the Catholic Church, by trying to take humankind and the immovable Earth away from the center of
God’s creation.
Before discussing how causal reasoning is essential for the realization of contrarian or delusional
beliefs, it is worth emphasizing the role of beliefs as motivators of action. Namely, the strength or degree of
one’s belief can be measured by one’s likelihood to take action as a result of that belief (Ramsey, 1931; also see
Felin, Gambardella, and Zenger, 2021). By way of contrast, the degree or strength of belief, based on
probabilistic or Bayesian models of cognition (cf. Pinker, 2021), is directly tied to existing data and the weight
of the available evidence (cf. Keynes, 1921), rather than the likelihood of taking action—a significant
difference.
Notice the implications of this in a context of our earlier example, human powered flight. Belief
played a central role in motivating action on the part of aviation pioneers despite overwhelming data and
evidence against the belief. In a sense, those pursuing flight did not appropriately update their beliefs. Much if
not most of the evidence was against the Wright brothers, but somehow they still believed in the plausibility of
flight. One of the Wright brothers, Wilbur, wrote to the scientist and aviation pioneer Samuel Langley in 1899
and admitted that “for some years I have been afflicted with the belief that flight is possible. My disease has
increased in severity and I feel that it will soon cost me an increased amount of money if not my life” (Wright
and Wright, 1881-1940, emphasis added). Wilbur clearly recognized that his belief about flight appeared
delusional to others, as is evident from his letters. But this belief motivated him to engage in causal reasoning
and experimentation that enabled him and his brother to make the seemingly delusional belief a reality (only
four short years later). Contrast the Wright brothers’ belief with the belief of Lord Kelvin, one of the greatest
scientific minds of the time. When invited to join the newly-formed Aeronautical Society a decade earlier,
Kelvin declined and said “I have not the slightest molecule of faith in aerial navigation.” Here Kelvin might
have been channeling a scientific contemporary of his—the mathematician William Clifford—who argued that
“it is wrong always, everywhere, and for anyone to believe anything on insufficient evidence” (2010: 79).
Kelvin did not want to lend support to what he considered an anti-scientific endeavor. Without the slightest
belief in the possibility of human flight, Kelvin naturally did not want to support anything that suggested
human powered flight might be possible. But for the Wright brothers, the possibility of powered flight was

30
very much a “live hypothesis” (James, 1967). Despite the data, they believed human flight might be possible,
and took specific steps to realize their belief.
Asymmetries between data and beliefs present problems for the very idea of rationality (cf. Chater et
al., 2018; Felin and Koenderink, 2022). After all, to be a rational human being, our knowledge should be based
on evidence. Our beliefs and knowledge should be proportionate to the evidence at hand. In a strict sense, the
very concept of beliefs is not even needed, as one can instead simply talk about knowledge—that is, beliefs
justified by evidence. This is succinctly captured by Pinker who argues “I don’t believe in anything you have to
believe in” (2021: 244). This seems like a reasonable stance. It is also the basis of Bayesian approaches where
new data (somehow) emerges, and where we can update our beliefs and knowledge accordingly—providing us
an “optimal way to update beliefs given new evidence” (Pilgrim et al., 2024). This is indeed the implicit stance
of cognitive approaches that focus on computational and probabilistic belief updating (e.g., Dasgupta et al.,
2020).
But data-belief asymmetries—where existing data presently does not corroborate beliefs, or even goes
against them—can be highly useful, even essential. They are the raw materials of technological and scientific
progress. They are a central ingredient of decision making under uncertainty. Data-belief asymmetries direct
our awareness toward new data and possible experiments to generate the evidence to support a belief. Of
course, the idea of “seeking-data-to-verify-a-particular-belief” is the very definition of delusion and a host of
associated biases, including confirmation bias, motivated reasoning, cherry picking, denialism, self-deception,
and belief perseverance. To an outsider, this looks like the perfect example of “the bad habit of seeking
evidence that ratifies a belief and being incurious about evidence that might falsify it” (Pinker, 2021: 13; also
see Hahn and Harris, 2014). Belief in human powered flight readily illustrates this, as there was plenty of
evidence to falsify the Wright brothers’ belief in the plausibility of heavier-than-air flight. Holding an
asymmetric belief seems to amount to “wishful thinking”, or “protecting one’s beliefs when confronted with
new evidence” (Kruglanski, Jasko and Friston, 2020: 413; though see Anglin, 2019). The Wright brothers were
continuously confronted with evidence that disconfirmed their belief, including Samuel Langley’s public
failures with flight or the knowledge of Lilienthal’s failed attempts (and his death due to a failed flight
attempt). But in these instances, ignoring the salient data and evidence—not updating beliefs based on
seemingly strong evidence and even scientific consensus—turned out to be the correct course of action.

31
There are times when being (seemingly) irrational—ignoring evidence, disagreeing about its
interpretation, or selectively looking for the right data—turns out to be the correct course of action. Human
powered flight of course is a particularly vivid illustration of this, though even more mundane forms of human
behavior are fundamentally characterized by a similar process (Felin and Koenderink, 2022). Most important
for present purposes, our argument is that beliefs have a causal role of their own and can be measured by our
propensity to act on them (Felin et al., 2020; Ramsey, 1931). Of course, having beliefs or having a willingness-
to-act on them does not assure us that they are true. But they are an important motivation for action (Ajzen,
1991; Bratman, 1987).
23
And again, notice that our emphasis on beliefs should not be seen as an attempt to
dismiss the importance of data. Rather, as we highlight next, beliefs can motivate theory-based causal
reasoning that directs human awareness toward actions and experiments that enable the generation of new
data, evidence, and realization of new knowledge.
From Beliefs to Causal Reasoning and Experimentation
The realization of beliefs is not automatic. A central aspect of beliefs is their propensity to lead to causal
reasoning and some form of directed experimentation. Beliefs enable actors to articulate a path for how to
“intervene” in their surroundings and generate the evidence needed (Felin, Gambardella and Zenger, 2021).
Our view of cognition and action here is more generally informed by the idea that theorizing can guide
humans to develop an underlying causal logic that enables us to intervene in the world (Pearl and Mackenzie,
2018; also see Ehrig et al., 2024). This intervention-orientation means that we do not simply take the world as
it is, rather we counterfactually think about possibilities and future states, with an eye toward taking specific
action, experimenting and generating the right evidence. This shifts the locus from backward-oriented
information processing and prediction (where the data is given), to doing and experimentation (where the right
data and evidence is identified or generated). This involves actively questioning and manipulating causal
structures, allowing for a deeper exploration of “what if” scenarios. Counterfactual thinking empowers
humans to probe hypothetical alternatives and dissect causal mechanisms offering insights into necessary and
sufficient conditions for an outcome (Felin, Gambardella, Novelli, and Zenger, 2024). This approach is
significantly different from input-output and information processing-oriented models of AI and
23
Beyond the work of Ramsey, Ajzen, and Bratman mentioned above, there is of course a large literature on how beliefs
motivate action. Our emphasis here is on the interaction between data and beliefs (and eventually, the role of theory-based
causal logic), as this has manifest in computational, Bayesian and probabilistic forms of AI and cognition.

32
computational cognition, and various “data-driven” or Bayesian approaches to decision making. AI-based
models of cognition largely focus on patterns based on past associations and correlations—prediction is based
on past data. But these approaches lack an ability to understand underlying causal structures, hypothetical
possibilities, and possible interventions (cf. Ehrig et al., 2024; Felin et al., 2020). This is the role of theory-
based causal logic.
A focus on plausible interventions and experimentation can be illustrated by extending our example
of human powered flight. This example also aptly illustrates the difference between how data-oriented and
evidence-based scientists thought about the possibility of human powered flight versus how more
intervention-oriented and causal logic-based practitioners like the Wright brothers thought about it. To
understand flight, the Wright brothers delved into the minutiae of why previous attempts at flight had not
succeeded, and more importantly, they developed a causal theory of flight. While failed flight attempts and the
death of Lilienthal (and others) were used by many as data to claim that flight was impossible, the Wright
brothers looked at the specific reasons why these attempts had failed.
24
And while scientists had used bird data
to argue that human flight was impossible (due to weight) (e.g., LeConte, 1888; Newcomb, 1901), the Wright
brothers paid attention to a different aspect of birds flight. Ironically, bird-related data—though different
aspects of it—provided seeming evidence for both those advocating for and against flight. LeConte focused
on the weight of birds, while the Wright brothers engaged in observational studies of the mechanics of bird
flight and anatomy (why birds were able to fly), for example, carefully studying the positioning of bird wings
when banking and turning.
The key difference was that the Wright brothers—with their belief in the plausibility of flight—were
building a “causal theory of flying” rather than looking for data that confirmed or disconfirmed whether flight
was possible. The Wright brothers ignored the data and the scientific arguments of the naysayers. From the
Smithsonian, the Wright brothers requested and received details about numerous historical flight attempts,
including Otto Lilienthal’s records. The Wright brothers notes and letters reveal that they carefully studied the
flight attempts and aircraft of earlier pioneers like George Cayley, Alphonse Penaud, and Octave Chanute
24
The Wright brothers respected Otto Lilienthal and carefully analyzed his data. Based on their own experimentation,
they found that some of his data on “lift” overestimated lift coefficients. Lilienthal tested one wing shape while the Wright
brothers experimented with various options. The Wright brothers constructed their own wind tunnel to gather
aerodynamic data. Their tests led them to develop new lift, drag, and pressure distribution data, which differed from
Lilienthal's findings. This data was critical in designing their successful aircraft.

33
(Anderson, 2002; McCoullough, 2015; Wright and Wright, 1841-1940). They studied various aspects of past
flight attempts: the types of airplanes used, details about wing shape and size, weather conditions, and
underlying aerodynamic assumptions.
Again, the Wright brothers sought to develop their own, causal theory of flying. Their theory was not
just motivated by their contrarian belief that flight was possible (a belief for which there did not seem to be
any evidence). Their confidence in the plausibility of flight grew as they carefully studied the underlying
mechanics of flight, as they investigated the causal logic of flight. Most importantly, their causal reasoning led
them to articulate the specific problems they needed to solve for human powered flight to be possible. The
Wright brothers reasoned that it was essential to solve three problems related to flight—namely (a) lift, (b)
propulsion, and (c) steering. To illustrate the power of developing a theory-based causal logic, and identifying
specific problems to solve, coupled with directed experimentation, we briefly discuss how they addressed one
of the problems: the problem of lift.
In terms of lift, the Wright brothers understood that to achieve flight they needed a wing design that
could provide sufficient lift to overcome the weight of their aircraft. Indeed, prominent scientists had argued
that the prohibiting factor of human flight was weight (again, pointing to insect flight and the weight of those
birds that fly and those that do not). The Wright brothers felt that the concern with weight was not
insurmountable. Informed by their investigations into bird flight (and the flight attempts of others), they
approached this problem through a series of experiments that included the construction and testing of various
airfoils. Their experimentation was highly targeted and data-oriented, testing various wing shapes, sizes, and
angles. They also quickly realized that not everything needed to be tested at scale and that their experiments
with lift could more safely and cost effectively be done in laboratory conditions. Thus they constructed their
own wind tunnels. Targeted tests within these tunnels allowed the Wright brothers to learn the central
principles of lift. They measured everything and kept meticulous track of their data—data that they generated
through ongoing experimental manipulation and variation. This hands-on experimentation allowed them to
collect data on how different shapes and angles of attack affected lift. By systematically varying these
parameters and observing the outcomes, they were effectively employing causal reasoning to identify the
conditions under which lift could be maximized. Their discovery and refinement of wing warping for roll

34
control was a direct outcome of understanding the causal relationship between wing shape, air pressure, and
lift.
The same processes of causal reasoning and directed experimentation were also central for addressing
the other two problems, propulsion and steering or control. And more generally, the Wright brothers were
careful scientists in every aspect of their attempt to realize their belief in human powered flight. For example,
to determine a suitable place for their flight attempts, they contacted the US Weather Bureau. They had
established what the optimal conditions might be for testing flight. They needed four things: consistent wind
(direction and strength), wide open spaces, soft or sandy landing surfaces, and privacy. They received several
suggestions from the US Weather Bureau and chose Kitty Hawk, North Carolina for the site of their “real
world” trial (Wright and Wright, 1881-1940).
The Wright brothers’ approach to flight offers a useful case study and microcosm of how theory-
based causal logic enables belief realization, even when beliefs seemingly are not supported by existing data,
evidence, or science. Based on their theorizing, study, and causal reasoning, the Wright brothers engaged in
directed experimentation—to solve the central problems of lift, propulsion, and steering. Their approach
exemplifies the application of causal logic to understand and intervene in the world, in the seeming absence of
data (or even when data is contrary to one’s belief). Their success with flight demonstrates how a systematic,
intervention-oriented approach can unravel the causal mechanisms underlying complex phenomena and
overcome the shortcomings of existing data.
As is implied by our arguments, we think scientific, economic, and technological domains are replete
with opportunities for those with asymmetric beliefs to utilize theory-based causal reasoning and engage in
directed experimentation and problem solving (Felin and Zenger, 2017). As we have argued, existing theories
of cognition are overly focused on data-belief symmetry rather than data-belief asymmetry and how the latter
enables causal reasoning that can enable the emergence of heterogeneity and the creation of novelty and value.
While there is much excitement about using AI to automate the generation of new knowledge and novelty
generation (e.g., Csaszar et al., 2024; Lu et al., 2024; Manning et al., 2024)—and even calls to replace biased
human decision making by AI (e.g., Kahneman, 2018)—we argue that human causal reasoning cannot, at least
presently, be mimicked by AI systems or computational approaches to cognition. Next we further explore the
implications of this argument for decision making under uncertainty and strategy.

35
DISCUSSION: THE LIMITS OF PREDICTION FOR DECISION MAKING UNDER
UNCERTAINTY
As we have extensively discussed in this article, AI and the cognitive sciences use many of the same
metaphors, tools, methods, and ways of reasoning about intelligence, rationality, and the mind. The prevailing
assumption in much of the cognitive sciences is that the human mind is a computational input-output system
(Christian and Griffiths, 2016). Computational and algorithmic systems emphasize the power of prediction
based on past data. The centrality of prediction is echoed by one the pioneers of AI, Yann LeCun (2017), who
argues that “prediction is the essence of intelligence.”
Clearly the predictive capabilities of AI are powerful. But is prediction the central for decision making
under uncertainty as well (that is, in unpredictable situations)? Many argue that this is the case (e.g., Davenport
and Kirby, 2017, Kahneman, 2018). For example, in their widely-acclaimed book Prediction Machines: The Simple
Economics of Artificial Intelligence—Agrawal, Gans, and Goldfarb emphasize that, stripped down to its essence,
“AI is a prediction technology” (2022: 22-32). And a central claim of their book is that “prediction is at the heart
of making decisions under uncertainty.” (2022: 7, emphasis added). One way to summarize our argument in this
paper is that we disagree with the importance placed on prediction—particularly in the form it is manifest in
AI (that is, prediction based on past data)—especially in situations of uncertainty. Since the emphasis on
prediction is commonplace, it is worth carefully pinpointing why we disagree with the importance placed on
prediction.
Agrawal et al’s (2022) summary offers a useful way for us to crystallize our more general concerns
with the emphasis that is placed on prediction. Their argument might be summarized by pointing to a
relatively common causal chain (of sorts), one that proceeds from data to information to prediction and to a
decision, or in short: dataà informationàpredictionàdecision.
25
They specifically argue that “data provides
the information that enables a prediction” and prediction in turn is “a key input into our decision making.”
This causal chain—from data to information to prediction and decision—certainly has intuitive appeal and
mirrors what AI systems are good at: taking in vast amounts of inputs and data, processing this information,
and then making predictions that can be used to make decisions. In short, as emphasized by Agrawal et al
25
This has parallels with the data-information-knowledge-wisdom or “DIKW” framework. For discussions of this see
Felin, Koenderink, Krueger, Noble and Ellis, 2021, and Yanai and Lercher, 2020.

36
(2022) and many others, data-driven prediction is at the heart of not just language models but AI more
generally, and also placed center stage in cognition.
But as we have highlighted throughout this paper, the problem is that data—data that is presently
available or given—is not likely to be the best source of information and prediction when making forward-
looking decisions. Data is snapshot of or mirror to the past. Even vast amounts of data are unlikely to
somehow enable one to anticipate the future (Felin, Kauffman, Koppl and Longo, 2014). What is needed is
some mechanism for projecting into the future and identifying the relevant data and evidence, or more likely,
experimentally generating new data. This is the role of a theory and some form of causal reasoning, which are
critical elements missing from data-first and prediction-oriented approaches to AI and cognition. We grant
that for various routine and repetitive decisions, prediction undoubtedly is a useful tool. Data-based prediction
can be highly powerful in predictable situations, situations that match or extrapolate from the past. This matches
what AI and prediction-based cognition is really good at, namely, the minimization of surprise and reduction of
error. More broadly this also matches the strong emphasis that many scholars of judgment and decision
making put on “consistency”—and the eagerness to avoid noise (see Kahneman, Sibony and Sunstein, 2021).
But many, important decisions are not meaningfully about uncertainty reduction through error
minimization using existing data. The purpose of large swaths of decision making is more about (in a sense)
maximizing surprise and error, or what to others might look like error. In a strategy context, the most impactful
opportunities and sources of value are not founded on immediately-available data. Rather, important decisions
like this require the development of a theory, founded on some kind of heterogeneous belief, that maps a
causal path or logic for how to test the theory, experiment, and gather new evidence to realize the belief. In an
important sense, strategic decision making has more to do with unpredictability and the maximization of
surprise rather than prediction and the minimization of surprise. Some decisions are highly-impactful, low-
frequency and rare, and fraught with uncertainty (Camuffo et al., 2022), and therefore simply not amenable to
algorithmic processing using existing data. This is why theory-based causal reasoning is not about
appropriately representing the structure of the environment, or about bounded rationality or listening to
customers—rather it is about developing a forward-looking theory and causal logic about how to experiment
and create value (Felin, Gambardella, Novelli and Zenger, 2024).

37
Notice that our focus on unpredictability and surprise does not mean that we are somehow outside
the realms of science or data. Quite the contrary. The process of making forward-looking decisions is about
developing an underlying theory-based causal logic of how one might intervene in the world—essentially,
outlining a causal path of how one might get from point A (the current state of the world) to point B (a
hypothesized future state of the world). Theories create salience for the right interventions, experiments, and
new data that will enable the realization of beliefs that initially appear implausible. Theories play a central role
in generating salience for what can be observed—the very idea of “data” (or observation) is theory-dependent.
As put by Einstein, “whether you can observe a thing or not depends on the theory which you use. It is the
theory which decides what can be observed” (Polanyi, 1974: 604). Salience to the right (or new) data or forms
of experimentation is given by a theory, not by past data. In this sense, theories could be said to have a
“predictive” function, though here prediction is not a data-driven or error-minimizing process as it has been
defined and operationalized within the context of AI (Agrawal et al., 2022) and cognitive science (cf. Clark,
2018). Now, if the task at hand is routine and mundane—for example, “predict the next word in this
sentence” or “tell me what you expect to see next”—then prediction with existing data can be useful. But the
theory-based view is more focused on the forward-looking aspects of cognition, and how human agents
realize beliefs by developing a multi-step causal path that enables the realization of beliefs through
experimentation and problem solving. This is precisely what our example of the Wright brothers’ theory of
flying—and causal reasoning and experimentation—illustrates. It serves as microcosm of a far more general
process of how humans intervene in their surroundings and realize novel beliefs. The economic domain is full
of examples of how economic actors engage in this process (Felin and Zenger, 2017).
Our emphasis on surprise and unpredictability—rather than predictability and the minimization of
error—is particularly important in competitive contexts. If everyone has access to the same prediction
machines and AI-related information processing tools, then the outcomes are likely to be homogeneous.
Strategy, if it is to actually create new value, needs to be unique and firm-specific. And this firm-specificity is
tied to unique beliefs and the development of a theory-based logic for creating value that is unforeseen (not
predictable) by others. Theories enable economic actors to “hack” competitive factor markets (Barney, 1986),
to develop unique expectations about the value of assets and activities. Theories also enable firms to “search”
in a more targeted fashion (Felin, Kauffman and Zenger, 2023), rather than engaging in costly and exhaustive

38
forms of global search. Prediction based engines, while there are attempts to fine-tune them, are inherently
based on past frequencies, correlations, and averages, rather than extremes. And in many instances, it is the
extremes that turn out to be far more interesting, as these provide the seeds of the (eventual) beliefs and data
that we later take for granted.
In all, we disagree with the emphasis that has been placed on prediction, algorithmic processing, and
computation in decision making and cognition (e.g., Agrawal et al., 2022; Christian and Griffiths, 2016).
Human decision making should not be relegated to AI (cf. Kahneman, 2018). AI and AI-inspired models of
cognition are based on backward-looking data and prediction rather than any form of forward-looking theory-
based causal logic. Emphasizing or relying on data and prediction is a debilitating limitation for not just
decision making and cognition, but also for understanding knowledge generation and even scientific progress.
Therefore we have emphasized the importance of heterogenous beliefs in human cognition, and the
development of theory-based causal logic that enables experimentation and the generation of new data and
novelty.
FUTURE RESEARCH OPPORTUNITIES
The above arguments suggest a number of research opportunities, particularly when it comes to understanding
AI, the emergence of novelty, and decision making under uncertainty. First, there is an opportunity to study
when and how AI-related tools might be utilized by humans (like economic actors) to create new value or to aid
in decision making. If AI—as a cognitive tool—is to be a source of competitive advantage, it has to be utilized
in unique or firm-specific ways. AI that uses universally-available training data will necessarily yield generic and
non-specific outputs. There is the risk that off-the-shelf AI solutions will be susceptible to the “productivity
paradox” of information technology (Brynjolffson and Hitt, 1998), where investments in AI actually do not
yield any gains to those buying these tools (rather, only to those selling these technologies). Thus there is an
opportunity to study how a specific decision maker’s—like a firm’s—own theory of value can drive the
process of AI development and adoption. For AI to actually be a useful tool for strategy and decision making,
AI needs to be customized, purpose-trained, and fine-tuned—it needs to be made specific—to the theories,
unique causal reasoning, datasets, and proprietary documents of decision makers like firms. For example,
advances on “retrieval-augmented generation” seem to offer a promising avenue to enhance specificity when

39
using AI in strategic decision-making. Any adoption of AI should be deliberate about which corpora and
training data are utilized (and which not) when seeking unique AI-driven outputs. After all, the outputs of an
AI—tailored to use specific data—are also the product of human agents who make decisions about which
data are relevant and (which are not) for the decision at hand. It is here that we see an opportunity to
understand how humans might uniquely interact with AI to generate these tools and associated human-AI
interfaces. Early work has begun to look at how firms utilize AI to increase innovation or how various human-
AI hybrid solutions enable better decision making (e.g., Babina et al., 2024; Bell et al., 2024; Choudhary,
Marchetti, Shrestha, and Puranam, 2023; Girotra et al., 2023; Gregory et al., 2021; Jia et al., 2024; Kemp, 2023;
Kim et al., 2023; Raisch and Fomina, 2023). But there are promising opportunities to study how a particular
economic actor’s or firm’s own theory and causal logic—as well as their unique or firm-specific sources of data
and information—can shape the development or adoption of AI-related tools for executing strategy and
making decisions.
Second, there are ongoing opportunities to research—and develop taxonomies of—the respective
capabilities of humans versus AI when it comes to different types of tasks, problems, and decisions. There is
much excitement, hype, and fearmongering about the prospects of AI replacing humans tout court (cf. Grace
et al., 2024). However, in reality, there will likely be a division of labor between humans and AI—with each
focusing on the types of tasks, problems, and decisions that it is best suited for. There is an opportunity to
study how economic actors and organizations contingently “match” humans (and their cognitive capacities,
jobs, roles) versus algorithms (or AI-related tools) with the right tasks and decisions. At present, clearly AI is
remarkably well-suited for tasks and decisions that are repetitive, computationally intensive, and that directly
extrapolate from past data. A significant number of decisions made by humans are relatively routine and
amenable to algorithmic processing. AI will therefore undoubtedly play a key role in many areas of
management, especially where processes repeat, such as operations (Amaya and Holweg, 2024; Holmström, et
al., 2019). However, some decisions are more low-frequency and rare (Camuffo et al., 2022), and therefore not
amenable to AI. Here we anticipate that humans will continue to play a central role, given their ability to
engage in forward-looking theorizing and the development of causal logic beyond extant data. That said,
naturally there is a “sliding scale” (and interfaces) between routine and non-routine decision making. Even in
the context of rare-and-highly-impactful decision making, AI might play a role, perhaps in serving as an

40
additional “voice” or “sparring” partner when generating or considering various strategies. As we have
discussed in this paper, AI and humans have their respective strengths and limitations. Existing work tends to
compare AI and humans on the same benchmarks, rather than recognizing the respective strengths of each.
Studying the comparative capabilities of AI and humans—their respective capabilities, limitations, and
ongoing evolution—represents a significant opportunity for future work.
Third, our arguments point to perhaps more “foundational” questions about the very nature of
humans, particularly related to the purportedly computational nature of human cognition. While questions
about the nature of cognition might sound overly abstract and philosophical, they are critically important as
they have downstream consequences for the assumptions we make, the methods we employ, as well as how
and what we study. Here we echo Herbert Simon who argued that ‘‘nothing is more important in setting our
research agenda and informing our research methods than our view of the nature of the human beings whose
behavior we are studying’’ (1985: 303; emphasis added). So, what is the predominant view of human cognition
within AI and the cognitive sciences (and by extension, in economics and strategy)? The predominant view of
humans is that they are input-output devices engaged in information processing, akin to computers. In this
paper we have pointed out problems with the decades-old “computer metaphor” of the human mind, brain,
and cognition. The computer has served as a central, organizing metaphor of human cognition for well over
seven decades—from the work of Alan Turing and Herbert Simon to modern instantiations of artificial neural
networks, predictive processing, and the Bayesian brain (e.g., Cosmides and Tooby, 2013; Knill and Pouget,
2004; Goldstein and Gigerenzer, 2024; Kotseruba and Totsos, 2020; Russell and Norvig, 2022; Sun, 2023).
A generalized computational approach to cognition, however, does not take into consideration the
comparative nature of the organism under study, because humans, organisms, and machines are all seen as the
same—as “invariant” (see Simon, 1990; cf. Gershman et al., 2015; Simon, 1980). But there are significant
differences in cognition, and these differences deserve careful attention. For example, computers do not
meaningfully make decisions about which inputs might be relevant and which might not, nor can they
meaningfully identify a new input, while humans have control over which inputs they might select or
“generate” in the first place (Brembs, 2021; Felin and Koenderink, 2022; Yin, 2020). Human cognition, as we
have discussed, is a form of forward-looking theorizing and causal reasoning. Notice that we are not trying to
argue for some kind of human exceptionalism here, as these capacities are manifest—in different ways—across

41
biological organisms more broadly (Riedl, 1984; cf. Popper, 1991).
26
There are significant research
opportunities to study the endogenous and comparative factors that enable biological organisms and
economic agents to theorize, reason, and experiment—and to compare various forms of biological intelligence
with artificial and nonbiological forms (cf. Levin, 2024). Treating all cognition and intelligence as generalized
computation unnecessarily narrows the scope of theoretical and empirical work, and fundamentally misses the
rich and heterogeneous ways that intelligence manifests itself across systems. Furthermore, the interfaces
between biological and nonbiological forms of intelligence—as is manifest in the human use of technology
and tools in evolution (Felin and Kauffman, 2023)—provide intriguing opportunities for future work.
CONCLUSION
In this article we have focused on the differences in cognition between AI and humans. While AI-inspired
models of cognition continue to emphasize the similarities between machines and humans, we argue that AI’s
emphasis based on prediction (using past data) does not capture human cognition—that is, it cannot explain
the emergence of novelty, new knowledge, nor can it assist in decision making under uncertainty. Overall, we
grant that there are some parallels between AI and human cognition. But we specifically emphasize the
forward-looking nature of human cognition and how theory-based causal reasoning allows humans to
intervene in the world, to engage in directed experimentation, and develop new knowledge. Heterogeneous
beliefs and theories—data-belief asymmetries—enable the identification or generation of new data (for
example, through experimentation), rather than merely being reliant on prediction based on the past data. AI-
based computational models are necessarily built on data-belief symmetries. AI therefore cannot causally map
and project into or anticipate the future, as illustrated by LLMs. That said, our arguments by no means negate
or question many of the exciting developments within the domain of AI. We anticipate that AI will help
humans make better decisions across many domains, especially in settings that are characterized by routine and
repetition. However, decisions under uncertainty—given the emphasis on unpredictability, surprise, and the
new—provide a realm that is not readily amenable to data- or frequency-based prediction and associated
computation. Thus, we fundamentally question the notion that AI will (or should) replace human decision
26
For example, even “simple” organisms like Drosophila (fruit flies) exhibit novel and surprising behaviors—like initiating
activity, expectations, and problem solving—that cannot be explained by or reduced to environmental inputs, genetic
factors or neural processing (see Heisenberg, 2014).

42
making (e.g., Kahneman, 2018). We argue that humans—compared to computers and AI—have unique
cognitive capacities that center on forward-looking beliefs and theorizing: the ability to engage in novel causal
reasoning and experimentation.
REFERENCES
Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., & McGrew, B. 2023. Gpt-4 technical report. arXiv
preprint arXiv:2303.08774.
Agarwal, R., Bacco, F., Camuffo, A., Coali, A., Gambardella, A., Msangi, H., & Wormald, A. 2023. Does a theory-of-value
add value? Evidence from a randomized control trial with Tanzanian entrepreneurs. SSRN working paper.
Aggarwal, C. C. 2018. Neural networks and deep learning. Springer Publishing.
Agrawal, A., Gans, J., & Goldfarb, A. 2022. Prediction machines (updated and expanded): The simple economics of artificial intelligence.
Harvard Business Review Press.
Agrawal, A., Gans, J., & Goldfarb, A. 2022. Power and prediction: The disruptive economics of artificial intelligence. Harvard Business
Review Press.
Agrawal, A., McHale, J., & Oettl, A. 2023. Superhuman science: How artificial intelligence may impact innovation. Journal
of Evolutionary Economics, 33(5), 1473-1517.
Agrawal, A., McHale, J., & Oettl, A. 2024. Artificial intelligence and scientific discovery: A model of prioritized search.
Research Policy, 53(5), 104989.
Ajzen, I. 1991. The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2), 179-211.
Amaya, J. & Holweg, M. 2024. Using algorithms to improve knowledge work. Journal of Operations Management,
forthcoming.
Ananthaswamy, A. 2022. DeepMind AI topples experts at complex game Stratego. Nature.
Anderson, J. D. 2004. Inventing Flight: The Wright brothers and their predecessors. Johns Hopkins University Press.
Anderson, J. R. 1976. Language, memory, and thought. Psychology Press.
Anderson, J. R. 1990. The adaptive character of thought. Psychology Press.
Anglin, S. M. 2019. Do beliefs yield to evidence? Examining belief perseverance vs. change in response to congruent
empirical findings. Journal of Experimental Social Psychology, 82, 176-199.
Babina, T., Fedyk, A., He, A., & Hodson, J. 2024. Artificial intelligence, firm growth, and product innovation. Journal of
Financial Economics, 151, 103745.
Baker, B., Lansdell, B., & Kording, K. P. 2022. Three aspects of representation in neuroscience. Trends in Cognitive Sciences.
Bell, J. J., Pescher, C., Tellis, G. J., & Füller, J. 2024. Can AI help in ideation? A theory-based model for idea screening in
crowdsourcing contests. Marketing Science, 43(1), 54-72.
Bénabou, R., & Tirole, J. 2016. Mindful economics: The production, consumption, and value of beliefs. Journal of Economic
Perspectives, 30(3), 141-164.
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. 2021. On the dangers of stochastic parrots: Can language
models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency (pp. 610-623).

43
Bengio, Y., Hinton, G., Yao, A., Song, D., Abbeel, P., Harari, Y. N., Hadfield, G., Russell, S., Kahneman, D., &
Mindermann, S. 2023. Managing ai risks in an era of rapid progress. arXiv preprint arXiv:2310.17688.
Bengio, Y., Lecun, Y., & Hinton, G. 2021. Deep learning for AI. Communications of the ACM, 64(7), 58-65.
Bhatia, S. 2023. Inductive reasoning in minds and machines. Psychological Review.
Biber, D. 1991. Variation across speech and writing. Cambridge University Press.
Binz, M., & Schulz, E. 2023. Turning large language models into cognitive models. arXiv preprint arXiv:2306.03917.
Binz, M., Dasgupta, I., Jagadish, A. K., Botvinick, M., Wang, J. X., & Schulz, E. 2023. Meta-learned models of cognition.
Behavioral and Brain Sciences, 1-38.
Bratman M. 1987. Intention, Plans and Practical Reason. Harvard University Press: Cambridge, MA.
Bryan, K. A., Ryall, M. D., & Schipper, B. C. 2022. Value capture in the face of known and unknown unknowns. Strategy
Science, 7(3), 157-189.
Bordalo, P., Conlon, J. J., Gennaioli, N., Kwon, S. Y., & Shleifer, A. 2023. How people use statistics (No. w31631).
National Bureau of Economic Research.
Bory, P. 2019. Deep new: The shifting narratives of artificial intelligence from Deep Blue to AlphaGo. Convergence, 25(4),
627-642.
Brembs, B. 2021. The brain as a dynamically active organ. Biochemical and Biophysical Research Communications, 564, 55-69.
Brynjolfsson, E., & Hitt, L. M. 1998. Beyond the productivity paradox. Communications of the ACM, 41(8), 49-55.
Buchanan, B. G., & Shortliffe, E. H. 1984. Rule based expert systems: the mycin experiments of the stanford heuristic programming
project. Addison-Wesley Longman Publishing Co.
Buckner, C. J. 2023. From deep learning to rational machines: What the history of philosophy can teach us about the future of artificial
intelligence. Oxford University Press.
Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., & VanRullen, R. 2023. Consciousness in artificial
intelligence: insights from the science of consciousness. arXiv preprint arXiv:2308.08708.
Camuffo, A., Cordova, A., Gambardella, A., & Spina, C. 2020. A scientific approach to entrepreneurial decision making:
Evidence from a randomized control trial. Management Science, 66(2), 564-586.
Camuffo, A., Gambardella, A., & Pignataro, A. 2022. Microfoundations of low-frequency high-impact decisions. SSRN
working paper.
Carey, S., & Spelke, E. 1996. Science and core knowledge. Philosophy of Science, 63(4), 515-533.
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., & Xie, X. 2024. A survey on evaluation of large language
models. ACM Transactions on Intelligent Systems and Technology, 15(3), 1-45.
Chater, N. 2018. Mind is flat: The remarkable shallowness of the improvising brain. Yale University Press.
Chater, N., Felin, T., Funder, D. C., Gigerenzer, G., Koenderink, J. J., Krueger, J. I., & Todd, P. M. 2018. Mind,
rationality, and cognition: An interdisciplinary debate. Psychonomic Bulletin & Review, 25, 793-826.
Chater, N., Zhu, J. Q., Spicer, J., Sundh, J., León-Villagrá, P., & Sanborn, A. 2020. Probabilistic biases meet the Bayesian
brain. Current Directions in Psychological Science, 29(5), 506-512.
Christian, B., & Griffiths, T. 2016. Algorithms to live by: The computer science of human decisions. Macmillan.
Chollet, F. 2019. On the measure of intelligence. arXiv preprint arXiv:1911.01547.
Chomsky N. 1975. Reflections on language. Pantheon: New York.

44
Choudhary, V., Marchetti, A., Shrestha, Y. R., & Puranam, P. 2023. Human-AI ensembles: When can they work? Journal of
Management.
Clark, A. 2018. A nice surprise? Predictive processing and the active pursuit of novelty. Phenomenology and the Cognitive
Sciences, 17(3), 521-534.
Clifford, W. K. 2010. The ethics of belief and other essays. Prometheus Books.
Clough, D. R., & Wu, A. 2022. Artificial intelligence, data-driven learning, and the decentralized structure of platform
ecosystems. Academy of Management Review, 47(1), 184-189.
Coady, C. A. J. 1992. Testimony: A philosophical study. Oxford University Press.
Constant, A., Friston, K. J., & Clark, A. 2024. Cultivating creativity: predictive brains and the enlightened room problem.
Philosophical Transactions of the Royal Society B, 379(1895), 20220415.
Cosmides, L. and Tooby, J., 2013. Evolutionary psychology: New perspectives on cognition and motivation. Annual Review
of Psychology, 64, 201-229.
Crouch, T. D. 2002. A dream of wings: Americans and the airplane, 1875-1905. WW Norton & Company.
Csaszar, F.A., Ketkar, H.J. & Kim, H. 2024. AI and strategic decision making. Working paper presented at Bocconi
University. October 2023.
Csaszar, F. A., & Levinthal, D. A. 2016. Mental representation and the discovery of new strategies. Strategic Management
Journal, 37(10), 2031-2049.
Csaszar, F. A., & Steinberger, T. 2022. Organizations as artificial intelligences: The use of artificial intelligence analogies in
organization theory. Academy of Management Annals, 16(1), 1-37.
Dasgupta, I., Schulz, E., Tenenbaum, J. B., & Gershman, S. J. 2020. A theory of learning to infer. Psychological Review,
127(3), 412.
Davenport, T. H., & Kirby, J. 2016. Only humans need apply: Winners and losers in the age of smart machines. New York: Harper
Business.
Dewey, J. 1916. Essays in experimental logic. University of Chicago Press.
Du, N., Huang, Y., Dai, A. M., Tong, S., Lepikhin, D., Xu, Y., & Cui, C. 2022. Glam: Efficient scaling of language models
with mixture-of-experts. International Conference on Machine Learning (5547-5569). PMLR.
Ehrig, T., Felin, T., & Zenger, T. 2024. Causal reasoning and the scientific entrepreneur: Beyond bayes. In Bayesian
Entrepreneurship, edited by Ajay Agrawal, Arnaldo Camuffo, Alfonso Gambardella, Joshua Gans, Erin Scott & Scott Stern.
MIT Press.
Ehrig, T., & Schmidt, J. 2022. Theory-based learning and experimentation: How strategists can systematically generate
knowledge at the edge between the known and the unknown. Strategic Management Journal, 43(7), 1287-1318.
Ellis, K. 2024. Human-like few-shot learning via bayesian reasoning over natural language. Advances in Neural Information
Processing Systems, 36.
Feigenbaum, E. A. 1963. Artificial intelligence research. IEEE Transactions in Information Theory, 9(4), 248-253.
Felin, T., Gambardella, A., Novelli, E., & Zenger, T. 2024. A scientific method for startups. Journal of Management.
Felin, T., Gambardella, A., & Zenger, T. 2021. Value lab: a tool for entrepreneurial strategy. Management & Business Review.
Felin, T., Kauffman, S., Koppl, R., & Longo, G. 2014. Economic opportunity and evolution: Beyond landscapes and
bounded rationality. Strategic Entrepreneurship Journal, 8(4), 269-282.

45
Felin, T., & Kauffman, S. 2023. Disruptive evolution: Harnessing functional excess, experimentation, and science as
tool. Industrial and Corporate Change, 32(6): 1372-1392.
Felin, T., Kauffman, S., & Zenger, T. 2023. Resource origins and search. Strategic Management Journal, 44(6), 1514-1533.
Felin, T., Koenderink, J., & Krueger, J. I. 2017. Rationality, perception, and the all-seeing eye. Psychonomic Bulletin &
Review, 24, 1040-1059.
Felin, T., Koenderink, J., Krueger, J. I., Noble, D., & Ellis, G. F. 2021. The data-hypothesis relationship. Genome Biology,
22(1), 1-6.
Felin, T., & Koenderink, J. 2022. A generative view of rationality and growing awareness. Frontiers in Psychology, 13, 807261.
Felin, T., & Zenger, T. R. 2017. The theory-based view: Economic actors as theorists. Strategy Science, 2(4), 258-271.
Festinger, L., Riecken, H. W., & Schachter, S. 1956. When prophecy fails. University of Minnesota Press.
Franceschelli, G., & Musolesi, M. 2023. On the creativity of large language models. arXiv preprint arXiv:2304.00008.
Friedman, J. H., & Popescu, B. E. 2008. Predictive learning via rule ensembles. Annals of Applied Statistics, 2(3): 916-954
Friston, K., & Kiebel, S. 2009. Predictive coding under the free-energy principle. Philosophical Transactions of the Royal Society
B: Biological Sciences, 364(1521), 1211-1221.
Friston, K. J., Da Costa, L., Tschantz, A., Kiefer, A., Salvatori, T., Neacsu, V., & Buckley, C. L. 2023. Supervised structure
learning. arXiv preprint arXiv:2311.10300.
Friston, K. J., Ramstead, M. J., Kiefer, A. B., Tschantz, A., Buckley, C. L., Albarracin, M., & René, G. 2024. Designing
ecosystems of intelligence from first principles. Collective Intelligence, 3(1), 26339137231222481.
Gans, J. S., Stern, S., & Wu, J. 2019. Foundations of entrepreneurial strategy. Strategic Management Journal, 40(5), 736-756.
Gavetti, G., & Levinthal, D. 2000. Looking forward and looking backward: Cognitive and experiential search.
Administrative Science Quarterly, 45(1), 113-137.
Gennaioli, N., & Shleifer, A. 2018. A crisis of beliefs: Investor psychology and financial fragility. Princeton University Press.
Gershman, S. J., Horvitz, E. J., & Tenenbaum, J. B. 2015. Computational rationality: A converging paradigm for
intelligence in brains, minds, and machines. Science, 349(6245), 273-278.
Gigerenzer, G., & Goldstein, D. G. 2024. Herbert Simon on the birth of the mind-computer metaphor. In Elgar
Companion to Herbert Simon. Edward Elgar Publishing.
Gingerich, O., & MacLachlan, J. 2005. Nicolaus Copernicus: Making the earth a planet. Oxford University Press.
Girotra, K., Meincke, L., Terwiesch, C., & Ulrich, K. T. 2023. Ideas are dimes a dozen: Large language models for idea
generation in innovation. Available at SSRN 4526071.
Gordin, M. D. 2021. On the fringe: Where science meets pseudoscience. Oxford University Press.
Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., & Brauner, J. 2024. Thousands of AI authors
on the future of AI. arXiv preprint arXiv:2401.02843.
Gregory, R. W., Henfridsson, O., Kaganer, E., & Kyriakou, H. 2021. The role of artificial intelligence and data network
effects for creating user value. Academy of Management Review, 46(3), 534-551.
Griffin, D., & Tversky, A. 1992. The weighing of evidence and the determinants of confidence. Cognitive Psychology, 24(3),
411-435.
Griffiths, T. L., Chater, N., Kemp, C., Perfors, A., & Tenenbaum, J. B. 2010. Probabilistic models of cognition: Exploring
representations and inductive biases. Trends in Cognitive Sciences, 14(8), 357-364.

46
Gigerenzer, G. 2020. How to explain behavior? Topics in Cognitive Science, 12(4), 1363-1381.
Gilkerson, J., Richards, J. A., Warren, S. F., Montgomery, J. K., Greenwood, C. R., Kimbrough Oller, D. & Paul, T. D.
2017. Mapping the early language environment using all-day recordings and automated analysis. American Journal of Speech-
Language Pathology, 26(2), 248-265.
Goldman, A. I. 1999. Knowledge in a social world. Oxford University Press.
Goodfellow, I., Bengio, Y. & Courville, A., 2016. Deep learning. MIT Press.
Goyal, A., & Bengio, Y. 2022. Inductive biases for deep learning of higher-level cognition. Proceedings of the Royal Society A,
478(2266), 20210068.
Hahn, U., & Harris, A. J. 2014. What does it mean to be biased: Motivated reasoning and rationality. In Psychology of learning
and motivation (Vol. 61, pp. 41-102). Academic Press.
Hahn, U., Merdes, C., & von Sydow, M. 2018. How good is your evidence and how would you know? Topics in Cognitive
Science, 10(4), 660-678.
Halliday, M.A.K. 1989. Spoken and written language. Oxford University Press.
Harari, Y. N. 2018. 21 Lessons for the 21st Century. Random House.
Hart, B., and Risley, T. R. 2003. The early catastrophe: The 30 million word gap by age 3. American Educator, 27(1), 4-9.
Hasson, U., Nastase, S. A., & Goldstein, A. 2020. Direct fit to nature: an evolutionary perspective on biological and
artificial neural networks. Neuron, 105(3), 416-434.
Hebb, D. O. 1949. The organization of behavior: A neuropsychological theory. Psychology Press.
Heisenberg, M. 2014. The beauty of the network in the brain and the origin of the mind in the control of behavior. Journal
of Neurogenetics, 28(3-4), 389-399.
Hills, T. T., Todd, P. M., Lazer, D., Redish, A. D., & Couzin, I. D. 2015. Exploration versus exploitation in space, mind,
and society. Trends in Cognitive Sciences, 19(1), 46-54.
Hinton, G. E. 1992. How neural networks learn from experience. Scientific American, 267(3), 144-151.
Hinton, G.E. 2023. Will digital intelligence replace biological intelligence? University of Toronto Lecture. [Available online
here.]
Hinton, G. E., Osindero, S., & Teh, Y. W. 2006. A fast learning algorithm for deep belief nets. Neural Computation, 18(7),
1527-1554.
Hohwy, J. 2013. The predictive mind. OUP Oxford.
Hohwy, J. 2020. New directions in predictive processing. Mind & Language, 35(2), 209-223.
Holmström, J., Holweg, M., Lawson, B., Pil, F.K. and Wagner, S.M., 2019. The digitalization of operations and supply
chain management: Theoretical and methodological implications. Journal of Operations Management, 65(8), pp.728-734.
Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. 2019. The curious case of neural text degeneration. arXiv preprint
arXiv:1904.09751.
Hong, P., Ghosal, D., Majumder, N., Aditya, S., Mihalcea, R., & Poria, S. 2024. Stuck in the quicksand of numeracy, Far
from AGI Summit: Evaluating LLMs' mathematical competency through ontology-guided perturbations. arXiv preprint
arXiv:2401.09395.
Hu, S., Lu, C., & Clune, J. 2024. Automated design of agentic systems. arXiv preprint arXiv:2408.08435.
Jaeger, J., Riedl, A., Djedovic, A., Vervaeke, J., & Walsh, D. 2024. Naturalizing relevance realization: Why agency and
cognition are fundamentally not computational. Working paper.

47
James, W. 1967. The writings of William James: A comprehensive edition. University of Chicago Press.
Jia, N., Luo, X., Fang, Z., & Liao, C. 2024. When and how artificial intelligence augments employee creativity. Academy of
Management Journal, 67(1), 5-32.
Johnson-Laird, P. N. 1983. Mental models: Towards a cognitive science of language, inference, and consciousness. Harvard University
Press.
Joyce, J. M. 1999. The foundations of causal decision theory. Cambridge University Press.
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., & Hassabis, D. 2021. Highly accurate protein
structure prediction with AlphaFold. Nature, 596(7873), 583-589.
Kahneman, D. 2003. Maps of bounded rationality: Psychology for behavioral economics. American Eonomic Review, 93(5),
1449-1475.
Kahneman, D. 2011. Thinking fast and slow. Farrar, Straus & Giroux.
Kahneman, D., Rosenfield, A. M., Gandhi, L., & Blaser, T. 2016. Noise: How to overcome the high, hidden cost of
inconsistent decision making. Harvard Business Review, 94(10), 38-46.
Kahneman, D. 2018. A Comment on Artificial Intelligence and Behavioral Economics. In The Economics of Artificial
Intelligence: An Agenda (pp. 608-610). University of Chicago Press.
Kahneman, D., Sibony, O., & Sunstein, C. R. 2021. Noise: A flaw in human judgment. Hachette Publishing.
Karmiloff-Smith, A., & Inhelder, B. 1974. If you want to get ahead, get a theory. Cognition, 3(3), 195-212.
Kemp, A. 2023. Competitive advantages through artificial intelligence: Toward a theory of situated AI. Academy of
Management Review.
Keynes, J.M. 1921. A treatise on probability. London: Macmillan.
Kıcıman, E., Ness, R., Sharma, A., & Tan, C. 2023. Causal reasoning and large language models: Opening a new frontier
for causality. arXiv preprint arXiv:2305.00050.
Kim, H., Glaeser, E. L., Hillis, A., Kominers, S. D., & Luca, M. 2023. Decision authority and the returns to algorithms.
Strategic Management Journal.
Knill, D. C., & Pouget, A. 2004. The Bayesian brain: the role of uncertainty in neural coding and computation. Trends in
Neurosciences, 27(12), 712-719.
Korteling, J. H., van de Boer-Visschedijk, G. C., Blankendaal, R. A., Boonekamp, R. C., & Eikelboom, A. R. 2021.
Human-versus artificial intelligence. Frontiers in Artificial Intelligence, 4, 622364.
Kotseruba, I., & Tsotsos, J. K. 2020. 40 years of cognitive architectures: core cognitive abilities and practical
applications. Artificial Intelligence Review, 53(1), 17-94.
Kralik, J. D., Lee, J. H., Rosenbloom, P. S., Jackson Jr, P. C., Epstein, S. L., Romero, O. J., & McGreggor, K. 2018.
Metacognition for a common model of cognition. Procedia Computer Science, 145, 730-739.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. 2012. Imagenet classification with deep convolutional neural
networks. Advances in Neural Information Processing Systems, 25.
Kruglanski, A. W., Jasko, K., & Friston, K. 2020. All thinking is ‘wishful’ thinking. Trends in Cognitive Sciences, 24(6), 413-
424.
Kunda, Z. 1990. The case for motivated reasoning. Psychological Bulletin, 108(3), 480.
Kvam, P. D., & Pleskac, T. J. 2016. Strength and weight: The determinants of choice and confidence. Cognition, 152, 170-
180.

48
Laird, J. E., Newell, A., & Rosenbloom, P. S. 1987. Soar: An architecture for general intelligence. Artificial Intelligence, 33(1),
1-64.
Laird, J. E., Lebiere, C., & Rosenbloom, P. S. 2017. A standard model of the mind: Toward a common computational
framework across artificial intelligence, cognitive science, neuroscience, and robotics. AI Magazine, 38(4), 13-26.
Lake, B. M., & Baroni, M. 2023. Human-like systematic generalization through a meta-learning neural network. Nature,
623 (7985), 115-121.
Lakhotia, K., Kharitonov, E., Hsu, W. N., Adi, Y., Polyak, A., Bolte, B., & Dupoux, E. 2021. On generative spoken
language modeling from raw audio. Transactions of the Association for Computational Linguistics, 9, 1336-1354.
Lansdell, B. J., & Kording, K. P. 2019. Towards learning-to-learn. Current Opinion in Behavioral Sciences, 29, 45-50.
LeConte, J. 1888. The problem of a flying machine. Science Monthly 34, 69-77
LeCun, Y. 2017. A path to AI. Future of Life Institute.
LeCun, Y., Bengio, Y., & Hinton, G. 2015. Deep learning. Nature, 521(7553), 436–444.
Legg, S., & Hutter, M. 2007. Universal intelligence: A definition of machine intelligence. Minds and Machines, 17, 391-444.
Levin, M. 2024. AI: A bridge toward diverse intelligence and humanity’s future. Tufts University Working Paper.
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., & Kiela, D. 2020. Retrieval-augmented generation
for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459-9474.
Lu, C., Lange, R. T., Foerster, J., Clune, J., & Ha, D. 2024. The AI scientist: Towards fully automated open-ended
scientific discovery. arXiv preprint arXiv:2408.06292.
Macmillan-Scott, O., & Musolesi, M. 2024. (Ir)rationality and cognitive biases in large language models. Royal Society Open
Science, 11(6), 240255.
Mahowald, K., Ivanova, A. A., Blank, I. A., Kanwisher, N., Tenenbaum, J. B., & Fedorenko, E. 2024. Dissociating
language and thought in large language models. Trends in Cognitive Sciences.
Manning, B. S., Zhu, K., & Horton, J. J. 2024. Automated social science: Language models as scientist and subjects (No.
w32381). National Bureau of Economic Research.
Marr, D. 1982. Vision: A computational investigation into the human representation and processing of visual information. MIT Press.
Mastrogiorgio, A., Felin, T., Kauffman, S., & Mastrogiorgio, M. 2022. More thumbs than rules: is rationality an
exaptation? Frontiers in Psychology, 13, 805743.
McAfee, A., & Brynjolfsson, E. 2012. Big data: the management revolution. Harvard Business Review, 90(10), 60-68.
McCarthy J, & Hayes PJ. 1969. Some philosophical problems from the standpoint of artificial intelligence. Machine
Intelligence 4, 463–502
McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. 2007. A proposal for the Dartmouth summer research
project on artificial intelligence, August 31, 1955. AI Magazine, 27(4), 12-12.
McCoy, R. T., Yao, S., Friedman, D., Hardy, M., & Griffiths, T. L. 2023. Embers of autoregression: Understanding large
language models through the problem they are trained to solve. arXiv preprint arXiv:2309.13638.
McCoy, R. T., Smolensky, P., Linzen, T., Gao, J., & Celikyilmaz, A. 2023. How much do language models copy from their
training data? Evaluating linguistic novelty in text generation using raven. Transactions of the Association for Computational
Linguistics, 11, 652-670.
McCullough, D. 2015. The Wright Brothers. Simon and Schuster.

49
McIntosh, T. R., Susnjak, T., Liu, T., Watters, P., & Halgamuge, M. N. 2023. From google gemini to openai q*(q-star): A
survey of reshaping the generative artificial intelligence (ai) research landscape. arXiv preprint arXiv:2312.10868.
McCorduck, P. 2004. Machines who think: A personal inquiry into the history and prospects of artificial intelligence. CRC Press.
McClelland, J. L., & Rumelhart, D. E. 1981. An interactive activation model of context effects in letter perception: I. An
account of basic findings. Psychological Review, 88(5), 375.
McCulloch, W.S. & Pitts, W., 1943. A logical calculus of the ideas immanent in nervous activity. The Bulletin of Mathematical
Biophysics, 5, 115-133.
Melville, G. W. 1901. The engineer and the problem of aerial navigation. North American Review, 173(541), 820-831.
Merdes, C., Von Sydow, M., & Hahn, U. 2021. Formal models of source reliability. Synthese, 198, 5773-5801.
Miller, G. A. 1956. The magical number seven, plus or minus two: Some limits on our capacity for processing
information. Psychological Review, 63(2), 81.
Minaee, S., Mikolov, T., Nikzad, N., Chenaghlu, M., Socher, R., Amatriain, X., & Gao, J. 2024. Large Language Models: A
Survey. arXiv preprint arXiv:2402.06196.
Morris, M. R., Sohldickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., & Legg, S. 2023. Levels of AGI:
Operationalizing progress on the path to AGI. arXiv preprint arXiv:2311.02462.
Naveed, H., Khan, A. U., Qiu, S., Saqib, M., Anwar, S., Usman, M., & Mian, A. 2023. A comprehensive overview of large
language models. arXiv preprint arXiv:2307.06435.
Newcomb, S. 1901. Is the airship coming? McClure’s Magazine, 17, 432-435.s
Newell, A., Shaw, J. C., & Simon, H. A. 1959. Report on a general problem solving program. In International Conference on
Information Processing.
Newell, A. 1990. Unified theories of cognition. Harvard University Press.
New York Times, 1903 (October 9). Flying machines which do not fly.
Novelli, E., & Spina, C. 2022. When do entrepreneurs benefit from acting like scientists? A field experiment in the UK.
SSRN Working paper.
Parr, T., & Friston, K. J. 2017. Working memory, attention, and salience in active inference. Scientific Reports, 7(1), 14678.
Pearl, J., & Mackenzie, D. 2018. The book of why: the new science of cause and effect. New York: Basic Books.
Peirce CS. 1957. The logic of abduction. In Peirce’s Essays in the Philosophy of Science, Thomas V (ed). Liberal Arts Press: New
York; 195–205.
Perconti, P., & Plebe, A. 2020. Deep learning and cognitive science. Cognition, 203, 104365.
Pezzulo, G., Parr, T., & Friston, K. 2024. Active inference as a theory of sentient behavior. Biological Psychology, 108741.
Park, P. S., & Tegmark, M. 2023. Divide-and-conquer dynamics in AI-driven disempowerment. arXiv preprint
arXiv:2310.06009.
Pilgrim, C., Sanborn, A., Malthouse, E., & Hills, T. T. 2024. Confirmation bias emerges from an approximation to
Bayesian reasoning. Cognition, 245, 105693.
Pinker, S. 1994. The language instinct: How the mind creates language. William Morrow & Co.
Pinker, S. 2022. Rationality: What it is, why it seems scarce, why it matters. Penguin.
Polanyi, M. 1958. Personal knowledge. Routledge.

50
Polanyi, M. 1974. Genius in science. Cohen RS, Wartofsky MW, eds. Methodological and Historical Essays in the Natural and Social Science.
Boston Studies in the Philosophy of Science, Vol. 14 (Springer, Dordrecht), 57–71.
Poldrack, R. A. 2021. The physics of representation. Synthese, 199(1-2), 1307-1325.
Popper, K., 1991. All life is problem solving. Routledge.
Puranam, P., Stieglitz, N., Osman, M., & Pillutla, M. M. 2015. Modelling bounded rationality in organizations: Progress
and prospects. Academy of Management Annals, 9(1), 337-392.
Raisch, S., & Fomina, K. 2023. Combining human and artificial intelligence: Hybrid problem-solving in
organizations. Academy of Management Review.
Ramsey, F. P. 1929. General Propositions and Causality, in his Philosophical Papers, ed. D. H. Mellor. Cambridge: Cambridge
University Press, 1990, pp. 145–63.
Ramsey, F. P. 1931. The foundations of mathematics and other logical essays. Cambridge University Press.
Rao, H., & Greve, H. R. 2024. The plot thickens: A sociology of conspiracy theories. Annual Review of Sociology, 50.
Resnik, P. 2024. Large Language Models are Biased Because They Are Large Language Models. arXiv preprint
arXiv:2406.13138.
Rescorla, M. 2015. The computational theory of mind. Stanford Encyclopedia of Philosophy.
Riedl, R. 1984. Biology of knowledge: The evolutionary basis of reason. New York: Wiley.
Roli, A., Jaeger, J. and Kauffman, S.A. 2022. How organisms come to know the world: Fundamental limits on artificial
general intelligence. Frontiers in Ecology and Evolution, 1035-1050.
Rosenblatt, F. 1958. The perceptron: a probabilistic model for information storage and organization in the
brain. Psychological Review, 65(6), 386.
Rosenblatt, F. 1962. Principles of neurodynamics: Perceptrons and the theory of brain mechanisms. Washington, DC: Spartan Books.
Rudin, C. 2019. Stop explaining black box machine learning models for high stakes decisions and use interpretable models
instead. Nature Machine Intelligence, 1(5), 206-215.
Rumelhart, D. E., Smolensky, P., McClelland, J. L., & Hinton, G. 1986. Sequential thought processes in PDP models.
Parallel distributed processing: explorations in the microstructures of cognition, 2, 3-57.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. 1986. Learning representations by back-propagating
errors. Nature, 323(6088), 533-536.
Russell, S. J., & Norvig, P. 2022. Artificial intelligence a modern approach. London: Pearson.
Scheffer, M., Borsboom, D., Nieuwenhuis, S., & Westley, F. 2022. Belief traps: Tackling the inertia of harmful beliefs.
Proceedings of the National Academy of Sciences, 119(32), e2203149119.
Schwartenbeck, P., FitzGerald, T., Dolan, R., & Friston, K. 2013. Exploration, novelty, surprise, and free energy
minimization. Frontiers in Psychology, 4, 710.
Schwöbel, S., Kiebel, S., & Marković, D. 2018. Active inference, belief propagation, and the bethe approximation. Neural
Computation, 30(9), 2530-2567.
Shannon, C. E. 1948. A mathematical theory of communication. The Bell System Technical Journal, 27(3), 379-423.
Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., & Dean, J. 2017. Outrageously large neural
networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538.

51
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R. and Gal, Y., 2024. AI models collapse when trained
on recursively generated data. Nature, 631(8022), pp.755-759.
Simon, H. A. 1955. A behavioral model of rational choice. The Quarterly Journal of Economics, 99-118.
Simon, H. A. 1965. The shape of automation for men and management. New York: Harper & Row.
Simon, H. A. 1980. Cognitive science: The newest science of the artificial. Cognitive Science, 4(1), 33-46.
Simon, H. 1985. Human nature in politics: The dialogue of psychology with political science. American Political
Science Review, 79, 293-304.
Simon, H. 1985. Artificial intelligence: Current status and future potential. National Research Council Report – Office of Naval
Research.
Simon, H. A. 1990. Invariants of human behavior. Annual Review of Psychology, 41(1), 1-20.
Simon, H. A. 1996. The sciences of the artificial. MIT Press.
Simon, H. A., & Newell, A. 1958. Heuristic problem solving: The next advance in operations research. Operations Research,
6(1), 1-10.
Spelke, E. S., Breinlinger, K., Macomber, J., & Jacobson, K. 1992. Origins of knowledge. Psychological Review, 99(4), 605.
Srećković, S., Berber, A., & Filipović, N. 2022. The automated Laplacean demon: How ML challenges our views on
prediction and explanation. Minds and Machines, 1-25.
Sun, R. (Ed.). 2023. The Cambridge Handbook of Computational Cognitive Sciences. Cambridge University Press.
Tannen, D. 2007. Talking voices: Repetition, dialogue, and imagery in conversational discourse. Cambridge University Press.
Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman, N. D. 2011. How to grow a mind: Statistics, structure, and
abstraction. Science, 331(6022), 1279-1285.
Tervo, D. G. R., Tenenbaum, J. B., & Gershman, S. J. 2016. Toward the neural implementation of structure learning.
Current Opinion in Neurobiology, 37, 99-105.
Tinbergen, N., 1963. On aims and methods of ethology. Zeitschrift für Tierpsychologie, 20(4), 410-433.
Turing, A. M. 1948. Intelligent Machinery, in Mechanical Intelligence, Collected Works of A. M. Turing. D. C. Ince, ed. North
Holland, 1992, p. 107 -127.
Turing, A. M. 1950. Computing machinery and intelligence. Mind, 59 (236), 433-460.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., & Polosukhin, I. 2017. Attention is all you
need. Advances in Neural Information Processing Systems, 30.
Wright, O., & Wright, W. 1881-1940. Wilbur and Orville Wright papers. Library of Congress. Available online at
https://www.loc.gov/collections/wilbur-and-orville-wright-papers/about-this-collection/
Wuebker, R., Zenger, T., & Felin, T. 2023. The theory-based view: Entrepreneurial microfoundations, resources, and
choices. Strategic Management Journal.
Yanai, I., & Lercher, M. 2020. A hypothesis is a liability. Genome Biology, 21(1), 1-5.
Yin, H. 2020. The crisis in neuroscience. In The interdisciplinary handbook of perceptual control theory (pp. 23-48). Academic
Press.
Yiu, E., Kosoy, E., & Gopnik, A. 2023. Transmission versus truth, imitation versus innovation: What children can do that
large language and language-and-vision models cannot (yet). Perspectives on Psychological Science, 17456916231201401.

52
Zellweger, T., & Zenger, T. 2023. Entrepreneurs as scientists: A pragmatist alternative to the creation-discovery debate.
Academy of Management Review, 47(4), 696-699.
Zhou, H. Y., Yu, Y., Wang, C., Zhang, S., Gao, Y., Pan, J., & Li, W. 2023. A transformer-based representation-learning
model with unified processing of multimodal input for clinical diagnostics. Nature Biomedical Engineering, 1-13.
Zhou, P., Madaan, A., Potharaju, S. P., Gupta, A., McKee, K. R., Holtzman, A., & Faruqui, M. 2023. How far are large
language models from agents with theory-of-mind? arXiv preprint arXiv:2310.03051.
Zhu, J. Q., & Griffiths, T. L. 2024. Eliciting the priors of large language models using iterated in-context learning. arXiv
preprint arXiv:2406.01860.
APPENDIX 1
AI and Human Cognition: Some Further Background
The earliest attempts to develop machines that simulate human thought processes and
reasoning focused on general problem solving. Newell and Simon’s (1959) “general problem
solver” (GPS) represented an ambitious effort to (try to) solve any problem that could be
presented in logical form. GPS used means-ends analysis, a technique that compared a current
state to the desired state (or goal), identified the differences, and then applied operators
(actions) to reduce these differences. The early excitement associated with GPS and other AI
models—and their ability to mimic human intelligence and thought—was pervasive. As put by
Herbert Simon in 1958, “there are now in the world machines that think, that learn and create.
Moreover, their ability to do these things is going to increase rapidly until—in a visible future—
the range of problems they can handle will be coextensive with the range to which the human
mind has been applied” (Simon and Newell, 1958: 8).
Early models like GPS provided the foundations for general cognitive architectures like SOAR
and ACT-R (Anderson, 1990; Laird, Newell and Rosenbloom, 1987). The enthusiasm for these
general models of cognition and AI continues to this day. Kotseruba and Tsotsos (2020) offer
an extensive survey of over two hundred different “cognitive architectures” developed over the
past decades. The ultimate goal of all this research into cognition, they argue, “is to model the
human mind, eventually enabling us to build human-level artificial intelligence” (2020: 21).
However, while various cognitive architectures related to AI hope to be general—and to mimic
or even exceed human capability—their application domains have turned out to be extremely
narrow and specific in terms of the problems they actually solve. But despite limited success in
generalizing early models of AI (specifically, from the late 1950s to the 1990s), excitement about
the possibility of computationally modeling human cognition did not wane. Simon’s frequent
collaborator, Alan Newell, argued that “psychology has arrived at the possibility of unified
theories of cognition,” specifically where “AI provides the theoretical infrastructure for the
study of human cognition” (1990: 40). This unified approach builds on the premise that humans
share certain “important psychological invariants” with computers and artificial systems (Simon,
1990: 3). This logic has also been captured by such ideas as “computational rationality”
(Gershman et al., 2015).
To this day there are ongoing calls for and efforts to develop so-called “common model of
cognition”—or as put by others, a “standard model of the mind” based on AI (Laird et al.,
2017; cf. Kralik et al., 2018). The call for general models has been born out of a frustration with
the aforementioned proliferation of cognitive models that claim to be general, despite the fact
that these models are heterogeneous, and any given model is highly focused on solving very
specific tasks and problems. The effort to create a “meta”-model of cognitive AI—a model that
different proponents of various cognitive architectures could agree on—has so far led to the
identification of relatively generic elements. These models include basic elements like

53
perception (focused on incoming stimuli or observations of the state of the world), different
types of memory (and accompanied learning mechanisms), which in turn are linked to various
motor systems and behaviors (Laird et al., 2017).
Most of the above attempts to model the human mind and mimic human reasoning focused on
symbolic systems, so-called “good old-fashioned AI.” These approaches are an attempt to
model thinking and intelligence through the manipulation of symbols—which represent objects,
concepts, or states of the world—specifically through logical rules and the development of
heuristics. The symbolic approach to cognitive AI models the world using symbols, and then
uses logical operations to manipulate these symbols to solve problems. This represents a rule-
based and top-down approach to intelligence. It is top-down in the sense that it starts with a
high-level focus on understanding a particular problem domain and then breaking it down into
smaller pieces (rules and heuristics) for solving a specific task. Perhaps the most significant
applications in AI—between the 1950s and late 1980s—were based on these rule-based
approaches. One of the more successful applications of an AI-related problem solver was the
backward chaining expert system MYCIN, which was applied to the diagnosis of bacterial
infections and the recommendation of appropriate antibiotics for treatment (Buchanan and
Shortliffe, 1984). The goal of a system like this was to mimic the judgments of an expert
decision maker. The model was a type of inference engine that used various pre-programmed
rules and heuristics to enable diagnosis. In all, AI that is based on symbolic systems represents a
top-down approach to computation and information processing that seeks to develop a rule- or
heuristic-based approach to replicate how a human expert might come to a judgment or a
decision.
Another approach to AI and modeling the human mind—called subsymbolic—also builds on
the idea of information processing and computation, but it emphasizes bottom-up learning.
These models also see the mind (or brain) as an input-output device. But the emphasis is on
learning things “from scratch”—that is, learning directly from data. Vast inputs and raw data are
fed to these systems to recognize correlations and statistical associations, or in short, patterns.
The weakness of the aforementioned symbolic systems is that these approaches are only useful
for relatively static contexts which do not meaningfully allow for any form of dynamic, bottom-
up learning from data or environments.
The foundations of subsymbolic AI were laid by scholars seeking to understand the human
brain, particularly perception. Rosenblatt (1958, 1962; building on Hebb, 1949) proposed one of
the earliest forms of a neural network in his model of a “perceptron,” which is the functional
equivalent of an artificial neuron. Rosenblatt’s work on the perceptron aimed to replicate the
human neuron, which when coupled together would resemble human neural networks. Since
modern artificial neural networks—including convolutional, recurrent, autoencoders, generative
adversarial networks—build on this broad foundation (e.g., Aggarwal, 2018; LeCun, Bengio and
Hinton, 2015), it is worth briefly highlighting the general architecture of this approach. The
architecture of the multi-layer perceptron includes layers that resemble the sensory units (input
layer), association units (hidden layer), and response units (output layer) of the brain. This
structure is very much the foundation of modern neural networks (Hinton, 1992; Rumelhart et
al., 1986) and the basis for the radical advances made in areas such as AI image recognition and
computer vision (Krizhevsky, Sutskever and Hinton, 2012). While these models emerged
seemingly out of nowhere, it is important to understand that the foundations were laid decades
ago (Buckner, 2023).
The process of learning in a neural network—as specified by Rosenblatt—begins with stimuli
hitting the sensory units, generating a binary response that is processed by the association cells
based on a predetermined threshold. The association cells then send signals to the response
area, which determine the perceptron’s output based on the aggregated inputs from the

54
association cells. The perceptron’s learning mechanism is based on feedback signals between
the response units and the association units, allowing the network to learn and self-organize
through repeated exposure to stimuli. So-called Hebbian learning (Hebb, 1949)—which posits
the relatively cliché but important idea that “neurons that fire together, wire together”—was the
precursor to these types of feedback-based learning processes and many modern concepts of
neural network theory.
In the intervening decades, research on artificial neural networks has progressed radically from
simple classifiers to highly complex, multilayer non-linear models capable of sophisticated
feature learning and pattern recognition through weighting and updates using large datasets
(e.g., Aggarwal, 2018; Shazeer et al., 2017). Various forms and combinations of machine
learning types—for example: supervised, unsupervised, and reinforcement learning—have
enabled radical breakthroughs in image recognition and computer vision, recommender
systems, game play, text generation, and so forth. And commensurate interest in the interaction
between human neural networks and AI—various forms of learning—has continued within the
cognitive sciences. This includes work on learning the structure of the environment (Friston et
al., 2023; also see Hasson et al., 2020), meta learning (Lake and Baroni, 2023) or so-called
“meta-learned models of cognition” (Binz et al, 2023), as well as inductive reasoning by humans
and AI (Bhatia, 2023), and inferential learning (Dasgupta et al., 2020). Many of these models of
learning build on neural networks in various forms, as well as related approaches.
In all, in this Appendix we have sought to further highlight the deep connections between AI
and computational models of human cognition. AI and other cognitive systems are treated in
similar fashion, as information processing machines or input-output devices (Simon, 1990).
While there has been widespread emphasis on the similarities between machines and humans, in
the paper we explicitly focus on the differences and emphasize the importance of theory-based
causal reasoning in human cognition.