Julian Jara-Ettinger’s research while affiliated with Yale University and other places

As AI systems become increasingly powerful, the need for safe AI has become more pressing. Humans are an attractive model for AI safety: as the only known agents capable of general intelligence, they perform robustly even under conditions that deviate significantly from prior experiences, explore the world safely, understand pragmatics, and can cooperate to meet their intrinsic goals. Intelligence, when coupled with cooperation and safety mechanisms, can drive sustained progress and well-being. These properties are a function of the architecture of the brain and the learning algorithms it implements. Neuroscience may thus hold important keys to technical AI safety that are currently underexplored and underutilized. In this roadmap, we highlight and critically evaluate several paths toward AI safety inspired by neuroscience: emulating the brain's representations, information processing, and architecture; building robust sensory and motor systems from imitating brain data and bodies; fine-tuning AI systems on brain data; advancing interpretability using neuroscience methods; and scaling up cognitively-inspired architectures. We make several concrete recommendations for how neuroscience can positively impact AI safety.

Human social cognition is built on our capacity to think about other minds. Do people have a general capacity to infer the contents of other minds from observable behavior? Or is this capacity specialized for inferring certain mental states, while struggling with others? Using a novel paradigm, we tested people’s capacity to infer another person’s desires, beliefs, or visual experience across three logically equivalent tasks. Despite the logical equivalence, people were less accurate, slower, and reported expending more effort when inferring the content of others’ visual experience compared to inferring others' beliefs and desires (Experiments 1 and 2; N = 120 U.S. adults and N = 60 U.S. adults). However, additional training rapidly improved people’s capacity to infer others’ visual experience (Experiment 3; N = 60 U.S. adults). Our results show that different kinds of mental state inferences produce different performance signatures, which may relate to how frequently we need to make such inferences in daily life.

The discovery of a new kind of experience can teach an agent what that kind of experience is like. Such a discovery can be epistemically transformative, teaching an agent something they could not have learned without having that kind of experience. However, learning something new does not always require new experience. In some cases, an agent can merely expand their existing knowledge using, e.g., inference or imagination that draws on prior knowledge. We present a computational framework, grounded in the language of partially observable Markov Decision Processes (POMDPs), to formalize this distinction. We propose that epistemically transformative experiences leave a measurable “signature” distinguishing them from experiences that are not epistemically transformative. For epistemically transformative experiences, learning in a new environment may be comparable to “learning from scratch” (since prior knowledge has become obsolete). In contrast, for experiences that are not transformative, learning in a new environment can be facilitated by prior knowledge of that same kind (since new knowledge can be built upon the old). We demonstrate this in a synthetic experiment inspired by Edwin Abbott’s Flatland, where an agent learns to navigate a 2D world and is subsequently transferred either to a 3D world (epistemically transformative change) or to an expanded 2D world (epistemically non-transformative change). Beyond the contribution to understanding epistemic change, our work shows how tools in computational cognitive science can formalize and evaluate philosophical intuitions in new ways.

People derive contrastive inferences when interpreting adjectives (e.g., inferring that ‘the short pencil’ is being contrasted with a longer one). However, classic eye-tracking studies revealed contrastive inferences with scalar and material adjectives, but not with color adjectives. This was explained as a difference in listeners’ informativity expectations, since color adjectives are often used descriptively (hence not warranting a contrastive interpretation). Here we hypothesized that, beyond these pragmatic factors, perceptual factors (i.e., the relative perceptibility of color, material and scalar contrast) and semantic factors (i.e., the difference between gradable and non-gradable properties) also affect the real-time derivation of contrastive inferences. We tested these predictions in three languages with prenominal modification (English, Hindi, and Hungarian) and found that people derive contrastive inferences for color and scalar adjectives, but not for material adjectives. In addition, the processing of scalar adjectives was more context dependent than that of color and material adjectives, confirming that pragmatic, perceptual and semantic factors affect the derivation of contrastive inferences.

How do humans build and navigate their complex social world? Standard theoretical frameworks often attribute this success to a foundational capacity to analyze other people’s appearance and behavior to make inferences about their unobservable mental states. Here we argue that this picture is incomplete. Human behavior leaves traces in our physical environment that reveal our presence, our goals, and even our beliefs and knowledge. A new body of research shows that, from early in life, humans easily detect these traces—sometimes spontaneously—and readily extract social information from the physical world. From the features and placement of inanimate objects, people make inferences about past events and how people have shaped the physical world. This capacity develops early and helps explain how people have such a rich understanding of others: by drawing not only on how others act but also on the environments they have shaped. Overall, social cognition is crucial not only to our reasoning about people and actions but also to our everyday reasoning about the inanimate world.

When deciding how to act in new situations, we expect agents to draw on relevant prior experiences. This expectation underlies many of our mental state inferences, allowing us to infer agents’ prior knowledge from their current actions. Across three experiments (n = 264 4- to 6-year-olds recruited at testing sites with diverse populations), we find that four-year-olds share this expectation; however, it is not until age six that children reliably use this expectation to infer what and how much others know. This work suggests that even young children have principled expectations for how ignorance will lead agents to act—and six-year-olds may already understand that knowledge is graded rather than binary, able to infer which of two agents knows more.

When reasoning about others’ knowledge, intuitively we consider not just action outcomes like success or failure, but also the probability of such outcomes under different knowledge states. Across three experiments (n = 240 North American four- to six-year-olds) we find that even four-year-olds understand that tasks with a lower probability of random success are harder—but not until age six do children use this information to gauge (Experiment 1) and infer (Experiments 2-3) what others know. These results suggest that, although basic probabilistic reasoning and representations of knowledge are well in place by age four, children do not integrate the two to make mental-state inferences until much later, pointing to an area of important developmental change in Theory of Mind.

Inferences about other people's knowledge and beliefs are central to social interaction. However, it is often not possible to tell what exactly other people know, because their behavior is consistent with a range of potential epistemic states. Nonetheless, in many of these situations we often have coarse intuitions about how much someone knows, despite being unable to pinpoint the exact content of their knowledge. Here we sought to explore this capacity in humans, by comparing their performance to a normative model capturing this kind of broad epistemic-state inference, centered on the expectation that agents maximize epistemic utilities. We evaluate our model in a graded inference task where people had to infer how much an agent knew based on the actions they chose (Experiment 1), and joint inferences about how much someone knew and how much they believed they could learn (Experiment 2). Critically, the agent's knowledge was always under-determined by their behavior, but the behavior nonetheless contained information about how much knowledge they possessed or believed they could gain. Our model captures nuanced patterns in participant judgments, revealing that people have a quantitative capacity to infer amorphous knowledge from minimal behavioral evidence.

Linguistic communication is an intrinsically social activity that enables us to share thoughts across minds. Many complex social uses of language can be captured by domain-general representations of other minds (i.e., mentalistic representations) that externally modulate linguistic meaning through Gricean reasoning. However, here we show that representations of others’ attention are embedded within language itself. Across ten languages, we show that demonstratives—basic grammatical words (e.g., “this”/“that”) which are evolutionarily ancient, learned early in life, and documented in all known languages—are intrinsic attention tools. Beyond their spatial meanings, demonstratives encode both joint attention and the direction in which the listener must turn to establish it. Crucially, the frequency of the spatial and attentional uses of demonstratives varies across languages, suggesting that both spatial and mentalistic representations are part of their conventional meaning. Using computational modeling, we show that mentalistic representations of others’ attention are internally encoded in demonstratives, with their effect further boosted by Gricean reasoning. Yet, speakers are largely unaware of this, incorrectly reporting that they primarily capture spatial representations. Our findings show that representations of other people’s cognitive states (namely, their attention) are embedded in language and suggest that the most basic building blocks of the linguistic system crucially rely on social cognition.

Moffett points to humans’ use of physical markers to signal group identity as crucial to human society. We characterize the developmental and cognitive bases of this capacity, arguing that it is part of an early-emerging, intuitive socio-physical interface which allows the inanimate world to encode rich social meaning about individuals’ identities, and the values of the society as a whole.