PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.


In this work, we survey skepticism regarding AI risk and show parallels with other types of scientific skepticism. We start by classifying different types of AI Risk skepticism and analyze their root causes. We conclude by suggesting some intervention approaches, which may be successful in reducing AI risk skepticism, at least amongst artificial intelligence researchers.
AI Risk Skepticism
Roman V. Yampolskiy
Computer Science and Engineering
University of Louisville
In this work, we survey skepticism regarding AI risk and show parallels with other types of
scientific skepticism. We start by classifying different types of AI Risk skepticism and analyze
their root causes. We conclude by suggesting some intervention approaches, which may be
successful in reducing AI risk skepticism, at least amongst artificial intelligence researchers.
Keywords: AI Risk, AI Risk Skepticism, AI Risk Denialism, AI Safety, Existential Risk.
1. Introduction to AI Risk Skepticism
It has been predicted that if recent advancement in machine learning continue uninterrupted,
human-level or even superintelligent Artificially Intelligent (AI) systems will be designed at some
point in the near future [1]. Currently available (and near-term predicted) AI software is subhuman
in its general intelligence capability but it is already capable of being hazardous in a number of
narrow domains [2], mostly with regard to privacy, discrimination [3, 4], crime automation or
armed conflict [5]. Superintelligent AI, predicted to be developed in the longer term, is widely
anticipated [6] to be far more dangers and is potentially capable of causing a lot of harm including
an existential risk event for the humanity as a whole [7, 8]. Together the short-term and long-term
concerns are known as AI Risk [9].
An infinite number of pathways exists to a state of the world in which a dangerous AI is unleashed
[10]. Those include mistakes in design, programming, training, data, value alignment, self-
improvement, environmental impact, safety mechanisms, and of course intentional design of
Malevolent AI (MAI) [11-13]. In fact, MAI presents the strongest; some may say undeniable,
argument against AI Risk skepticism (not to be confused with “skeptical superintelligence” [14]).
While it may be possible to argue that a particular pathway to dangerous AI will not materialize
or could be addressed, it seems nothing could be done against someone purposefully designing a
dangers AI. MAI convincingly establishes potential risks from intelligent software and makes
denialist’s point of view scientifically unsound. In fact, the point is so powerful, the authors of
[11] were contacted by some senior researchers who expressed concern about impact such a
publication may have on the future development and funding of AI research.
More generally, much can be inferred about the safety expectations for future intelligent systems
from observing abysmal safety and security of modern software. Typically, users are required to
click “Agree” on the software usage agreement, which denounces all responsibility from software
developers and explicitly waves any guarantees regarding reliability and functionality of the
provided software, including commercial products. Likewise, hardware components for the
Internet of Things (IoT) notoriously lack security
in the design of the used protocol. Even in
principle, sufficient levels of safety and security may not be obtainable for complex software
products [15, 16].
Currently a broad consensus
exists in the AI Safety community, and beyond it, regarding
importance of addressing existing and future AI Risks by devoting necessary resources to making
AI safe and beneficial, not just capable. Such consensus is well demonstrated by a number of open
signed by thousands of leading practitioners and by formation of industry coalitions
with similar goals. Recognizing dangers posed by an unsafe AI a significant amount of research
[17-24] is now geared to develop safety mechanisms for ever-improving intelligent software, with
AI Safety research centers springing up at many top universities such as MIT
, Berkeley
, and Cambridge
, companies
and non-profits
Given tremendous benefits associated with automation of physical and cognitive labor it is likely
that funding and effort dedicated to creation of intelligent machines will only accelerate. However,
it is important to not be blinded by the potential payoff, but to also consider associated costs.
Unfortunately, like in many other domains of science, a vocal group of skeptics is unwilling to
accept this inconvenient truth claiming that concerns about the human-caused issue of AI Risk are
just “crypto-religious” [25] pseudoscientific “mendacious FUD” [26] alarmism [27] and
luddite [28], “vacuous”, “nonsense” [29], “fear of technology, opportunism, or ignorance”,
“anti-AI”, “hype”, “comical”, “so ludicrous that it defies logic”, “magical thinking”, “techno-
panic” [30], “doom-and-gloom”, “Terminator-like fantasies” [31], “unrealistic”, “sociotechnical
blindness”, “AI anxiety” [32], technophobic”, “paranoid” [33]
, neo-fear [34] and “mental
masturbation” [35] by “fearmongers” [30], AI doomsayers [36], “AI Dystopians”, “AI
Apocalypsarians”, - a “Frankenstein complex” [37]. They accuse AI Safety experts of being
“crazy”, “megalomaniacal”, “alchemists”, and “AI weenies”, performing “parlor tricks” to spread
their “quasi-sociopathic”, “deplorable beliefs”, about the “nerd Apocalypse”, caused by their
“phantasmagorical AI” [38].
Even those who have no intention to insult anyone have a hard time resisting such temptation:
The idea that a computer can have a level of imagination or wisdom or intuition greater than
Comment on article by Steven Pinker
humans can only be imagined, in our opinion, by someone who is unable to understand the nature
of human intelligence. It is not our intention to insult those that have embraced the notion of the
technological singularity, but we believe that this fantasy is dangerous …” [39]. Currently,
disagreement between AI Risk Skeptics [40] (interestingly Etzioni was an early AI Safety leader
[41]) and AI Safety advocates [42] is limited to debate, but some have predicted that in the future
it will become a central issue faced by humanity and that the so-called “species dominance debate”
will result in a global war [43]. Such a war could be seen as an additional implicit risk from
progress in AI.
AI risk skeptics dismiss or bring into doubt scientific consensus of the AI Safety community on
superintelligent AI risk, including the extent to which dangers are likely to materialize, severity of
impact superintelligent AI might have on humanity and universe, or practicality of devoting
resources to safety research. [44]. A more extreme faction, which could be called AI Risk
, rejects any concern about AI-Risk, including from already deployed systems or soon to
be deployed systems.
For example, 2009 AAAI presidential panel on Long-Term AI Futures tasked with review and
response to concerns about the potential for loss of human control of computer-based intelligences,
concluded: “The panel of experts was overall skeptical of the radical views expressed by futurists
and science-fiction authors. Participants reviewed prior writings and thinking about the possibility
of an “intelligence explosion” where computers one day begin designing computers that are more
intelligent than themselves. They also reviewed efforts to develop principles for guiding the
behavior of autonomous and semi-autonomous systems. Some of the prior and ongoing research
on the latter can be viewed by people familiar with Isaac Asimov's Robot Series as formalization
and study of behavioral controls akin to Asimov’s Laws of Robotics. There was overall skepticism
about the prospect of an intelligence explosion as well as of a “coming singularity,” and also about
the large-scale loss of control of intelligent systems. [45].
Denialism of anthropogenic climate change has caused dangerous delays in governments
exercising control and counteraction. Similarly, influence of unchecked AI Risk denialism could
be detrimental for the long term flourishing of human civilization as it questions importance of
incorporating necessary safeguards into intelligent systems we are deploying. Misplaced
skepticism has negative impact on allocating sufficient resources for assuring that developed
intelligent systems are safe and secure. This is why it is important to explicitly call-out instances
of AI risk denialism, just as it is necessary to fight denialism in other domains in which it is
observed, such as history, healthcare and biology. In fact, in many ways the situation with
advanced AI risk may less forgiving. Climate change is comparable to soft takeoff [46], in which
temperature is gradually rising by a few degrees over a 100-year period. An equivalent to
superintelligence hard takeoff scenario would be global temperature rising by 100 degrees in a
2. Types of AI Risk Skeptics
I first used the term “AI risk denier” in a 2015 paper, and AGI risk skepticism in
a 2014 (co-authored) paper:
It is helpful to define a few terms and can be easily done by adopting the language addressing
similar types of science denialism
: AI risk denial is denial, dismissal, or unwarranted doubt that
contradicts the scientific consensus on AI risk, including its effects on humanity. Many deniers
self-label as "AI risk skeptics". AI risk denial is frequently implicit, when individuals or research
groups accept the science but fail to come to terms with it or to translate their acceptance into
action. While denying risk from existing intelligent systems is pure denialism, with respect to
future AIs, predicted to be superintelligent, it is reasonable to label such views as skepticism as
evidence is not as strong for risk from such systems and as they don’t currently exist and therefore
are not subject to empirical testing for safety. Finally, we also introduce the concept of an AI
safety skeptic, as someone who while accepting reality of AI risk doubts that safe AI is possible
to achieve either in theory or at least in practice.
In order to overcome AI Risk Skepticism it is important to understand its causes and the culture
supporting it. People who self-identify as AI Risk skeptics are very smart, ethical human beings
and otherwise wonderful people, nothing in this paper should be interpreted as implying otherwise.
Unfortunately, great people make great mistakes, and no mistake is greater than ignoring potential
existential risk from development of advanced AI. In this section, I review most common reasons
and beliefs for being an AI Risk Skeptic.
Non-Experts Non-AI-Safety researchers greatly enjoy commenting on all aspects of AI Safety. It
seems like anyone who saw Terminator thinks they have sufficient expertise to participate in the
discussion (on either side), but not surprisingly it is not the case. Not having formal training in the
area of research should significantly discount importance of opinion of such public intellectuals,
but it does not seem to be the case. By analogy, in discussions of cancer we listen to professional
opinion of doctors trained in treating oncological diseases but feel perfectly fine ignoring opinions
of business executives or lawyers. In AI Safety debates participants are perfectly happy to consider
opinions of professional atheists [47], web-developers [38] or psychologists [48], to give just some
Wrong Experts It may not be obvious but most expert AI Researchers are not AI Safety
Researchers! Many AI Risk Skeptics are very knowledgeable and established AI researchers, but
it is important to admit that having expertise in AI development is not the same as having expertise
in AI Safety and Security. AI researchers are typically sub-domain experts in one of many sub-
branches of AI research such as Knowledge Representation, Pattern Recognition, Computer Vision
or Neural Networks, etc. Such domain expert knowledge does not immediately make them experts
in all other areas of artificial intelligence, AI Safety being no exception. More generally, a software
developer is not necessarily a cybersecurity expert. It is easy to illustrate this by analogy with a
non-computer domain. For example, a person who is an expert on all things related to cement is
not inevitable an expert on the placement of emergency exits even though both domains have a lot
to do with building construction.
Professional Skeptics
Members of skeptic organizations are professionally predisposed to question everything and it is
not surprising that they find claims about properties of future superintelligent machines to fall in
their domain of expertise. For example, Michael Shermer, founder of Skeptics Society and
publisher of Skeptic magazine, has stated [47]: I'm skeptical. all such doomsday scenarios
involve a long sequence of if-then contingencies, a failure of which at any point would negate the
apocalypse. Similarly, those who are already skeptical about one domain of science, for example
theory of evolution, are more likely to also exhibit skepticism about AI Risk
Ignorant of Literature Regardless of background, intelligence or education many commentators
on AI Safety seem to be completely unaware of literature on AI Risk, top researchers in the field,
their arguments and concerns. It may no longer be possible to read everything on the topic due to
the share number of publications produced in recent years, but there really is no accuse for not
familiarizing yourself with top books [7, 8, 49] or survey papers [9] on the topic. It is of course
impossible to make a meaningful contribution to the discussion if one is not aware of what is
actually being discussed or is engaging with the strawman arguments.
Skeptics of Strawman
Some AI skeptics may actually be aware of certain AI Safety literature, but because of poor
understanding or because they were only exposed to weaker arguments for AI Risk, they find them
unconvincing or easily dismissible and so feel strongly justified in their skeptical positions.
Alternatively, they may find a weakness in one particular pathway to dangerous AI and
consequently argue against “fearing the reaper” [50].
With Conflict of Interest and Bias Lastly, we cannot ignore an obvious conflict of interest many
AI researchers, tech CEOs, corporations and others in the industry have with regards to their
livelihood and the threat AI Risk presents to unregulated development of intelligent machines.
History teaches us that we can’t count on people in the industry to support additional regulation,
reviews or limitations against their direct personal benefit. Tobacco company representatives, for
years, assured the public that cigarettes are safe, non-carcinogenic and non-addictive. Oil
companies rejected any concerns public had about connection between burning of fossil fuels and
global climate change, despite knowing better.
It is very difficult for a person whose success, career, reputation, funding, prestige, financial well-
being, stock options and future opportunities depend on unobstructed development of AI to accept
that the product they are helping to develop is possibly unsafe, requires government regulation,
and internal or even external review boards. As Upton Sinclair put it: “It is difficult to get a man
to understand something when his salary depends upon his not understanding it. They reasonably
fear that any initial concessions may lead to a significant “safety overhead” [51], reduced
competitiveness, slowdown in progress or a moratorium on development (à la human cloning) or
even an outright ban on future research. The conflict of interest developers of AI have with respect
to their ability to impartially assess dangers of their products/services is unquestionable and would
be flagged by any ethics panel. Motivated misinformation targeting lay people, politicians, and
public intellectuals may also come from governments, thought leaders and activist citizens
interested in steering debate in particular directions [52]. Corporations may additionally worry
about legal liability and overall loss of profits.
In addition to the obvious conflicts of interest, most people, including AI researchers, are also
subject to a number of cognitive biases making them underappreciate AI Risk. Those would
include Optimism Bias (thinking that you are at a smaller risk of suffering a negative outcome),
and Confirmation Bias (interpreting information in a way that confirms preconceptions).
Additionally, motivated reasoning may come into play, as Baum puts it [52]: Essentially, with
their sense of self-worth firmed up, they become more receptive to information that would
otherwise threaten their self-worth. As a technology that could outperform humans,
superintelligence could pose an especially pronounced threat to people’s sense of self-worth. It
may be difficult for people to feel good and efficacious if they would soon be superseded by
computers. For at least some people, this could be a significant reason to reject information about
the prospect of superintelligence, even if that information is true.”
3. Arguments for AI Risk Skepticism
In this section, we review most common arguments for AI risk skepticism. Russell has published
a similar list, which in addition to objections to AI risk concerns also includes examples of flawed
suggestions for assuring AI safety [53], such as: “Instead of putting objectives into the AI system,
just let it choose its own”, “Don’t worry, we’ll just have collaborative human-AI teams”, “Can’t
we just put it in a box?”, Can’t we just merge with the machines?” and “Just don’t put in ‘human’
goals like self-preservation”.
Importance of understanding denialists’ mindset is well-articulated by Russell: When one first
introduces [AI risk] to a technical audience, one can see the thought bubbles popping out of their
heads, beginning with the words “But, but, but . . .” and ending with exclamation marks. The first
kind of but takes the form of denial. The deniers say, “But this can’t be a real problem, because
XYZ.” Some of the XYZs reflect a reasoning process that might charitably be described as wishful
thinking, while others are more substantial. The second kind of but takes the form of deflection:
accepting that the problems are real but arguing that we shouldn’t try to solve them, either because
they’re unsolvable or because there are more important things to focus on than the end of
civilization or because it’s best not to mention them at all. The third kind of but takes the form of
an oversimplified, instant solution: “But can’t we just do ABC?” As with denial, some of the ABCs
are instantly regrettable. Others, perhaps by accident, come closer to identifying the true nature of
the problem. … Since the issue seems to be so important, it deserves a public debate of the highest
quality. So, in the interests of having that debate, and in the hope that the reader will contribute to
it, let me provide a quick tour of the highlights so far, such as they are.” [54].
In addition to providing a comprehensive list of arguments for AI risk skepticism, we have also
classified such objections into six categories (see Figure 1): Objections related to Priorities,
Technical issues, AI Safety, Ethics, Bias, and Miscellaneous ones. While research on types of
general skepticism exists [55], to the best of our knowledge this is the first such taxonomy
specifically for AI risk. In general, we can talk about politicized skepticism and intellectual
skepticism [56]. Politicized skepticism has motives other than greater understanding, while
intellectual skepticism aims for better comprehension and truth seeking. Our survey builds and
greatly expands on previous lists from Turing [57], Baum [56], Russell [53], and Ceglowski [38].
Too Far
Soft Takeoff is more likely and so we will have Time to Prepare
No Obvious Path to Get to AGI from Current AI
Short Term AI Concerns over AGI Safety
Something Else is More Important
AI Doesn’t Exist
Superintelligence is Impossible
Self-Improvement is Impossible
AI Can’t be Conscious
AI Can be Just a Tool
We can Always just Turn it Off
We Can Reprogram AIs if We Don’t Like What They Do
AI doesn’t have a Body an so can’t Hurt Us
If AI is as Capable as You Say, it Will not Make Dumb Mistakes
Superintelligence Would (Probably) Not Be Catastrophic
Self-preservation and Control Drives Don't Just Appear They Have to be Programmed In
An AI is not Pulled at Random from the Mind Design Space
AI Can’t Generate Novel Plans
AI Safety Can’t be Done Today
AI Can’t be Safe
Skepticism of Particular Risks
Skepticism of Particular Safety Methods
Skepticism of Researching Impossibility Results
Superintelligence is Benevolence
Let the Smarter Beings Win
Let’s Gamble
Malevolent AI is not worse than Malevolent Humans
AI Safety Researchers are Non-Coders
Majority of AI Researchers is not Worried
Anti-Media Bias
Keep it Quiet
Safety Work just Creates an Overhead Slowing Down Research
Heads in the Sand
So Easy it will be Solved Automatically
AI Regulation Will Prevent Problems
Other Arguments, …
Figure 1: Taxonomy of Objections to AI Risk
3.1 Priorities Objections
Too Far A frequent argument against work on AI Safety is that we are hundreds if not thousands
of years away from developing superintelligent machines and so even if they may present some
danger it is a waste of human and computational resources to allocate any effort to address
Superintelligence Risk at this point in time. Such position doesn’t take into account possibility that
it may take even longer to develop appropriate AI Safety mechanisms and so the perceived
abundance of time is a feature, not a bug. It also ignores a non-zero possibility of an earlier
development of superintelligence.
Soft Takeoff is more likely and so we will have Time to Prepare AI takeoff refers to the speed
with which an AGI can get to superintelligent capabilities. While hard takeoff is likely and means
that process will be very quick, some argue that we will face a soft takeoff and so will have
adequate time (years) to prepare [46]. While nobody knows the actual take off speed at this point,
it is prudent to be ready for the worst-case scenario.
No Obvious Path to Get to AGI from Current AI While we are making good progress on AI, it
is not obvious how to get from our current state in AI to AGI and current methods may not scale
[58]. This may be true, but this is similar to the “Too Far” objection and we definitely need all the
time possible to develop necessary safety mechanisms. Additionally, current state-of-the-art
systems [59], don’t seem to hit limits yet, subject to availability of compute for increasing model
size [60, 61].
Something Else is More Important Some have argued that global climate change, pandemics,
social injustice, and a dozen of other more immediate concerns are more important than AI risk
and should be prioritized over wasting money and human capital on something like AI Safety. But,
development of safe and secure superintelligence is a possible meta-solution to all the other
existential threats and so resources allocated to AI risk are indirectly helping us address all the
other important problems. Time wise it is also likely, that AGI will be developed before projected
severe impact from such issues as global climate change.
Short Term AI Concerns over AGI Safety Similar to the argument that something else is more
important, proponents claim that immediate issues with today’s AIs, such as algorithmic bias,
technological unemployment or limited transparency should take precedence over concerns about
future technology (AGI/superintelligence), which doesn’t yet exist and may not exist for decades
3.2 Technical Objections
AI Doesn’t Exist The argument is that current developments in Machine Learning are not progress
in AI, but are just developments in statistics, particularly in matrix multiplication and gradient
. Consequently, it is suggested that calls for regulation of AI are absurd. Of course, human
criminal behavior can be seen as interactions of neurotransmitters and ion channels, making their
criminalization questionable.
Superintelligence is Impossible If a person doesn’t think that superintelligence can ever be build
they will of course view Risk from Superintelligence with strong skepticism. Most people in this
camp assign a very small (but usually not zero) probability to the actual possibility of
superintelligent AI coming into existence [63-65], but if even a tiniest probability is multiplied by
the infinite value of the Universe the math seems to be against skepticism. Skeptics in this group
will typically agree that if superintelligence did exist it would have potential of being harmful.
“Within the AI community, a kind of denialism is emerging, even going as far as denying the
possibility of success in achieving the long-term goals of AI. It’s as if a bus driver, with all of
humanity as passengers, said, “Yes, I am driving as hard as I can towards a cliff, but trust me, we’ll
run out of gas before we get there!”” [54].
Self-Improvement is Impossible This type of skepticism concentrates on the supposed
impossibility of intelligence explosion, as a side-effect of recursive self-improvement [66], due to
fundamental computational limits [50] and software complexity [67]. Of course such limits are not
a problem as long as they are actually located above the level of human capabilities.
AI Can’t be Conscious Proponents argue that in order to be dangerous AI has to be conscious
[68]. As AI risk is not predicated on artificially intelligent systems experiencing qualia [69, 70], it
is not relevant if the system is conscious or not. This objection is as old as the field of AI itself, as
Turing addressed “The Argument from Consciousness” in his seminal paper [57].
AI Can be Just a Tool A claim that we do not need A(General)I to be an independent agent, it is
sufficient for them to be designed as assistants to humans in particular domains, such as GPS
navigation and so permit as to avoid dangers of fully independent AI
. It is easy to see that the
demarcation between Tool AI and AGI is very fuzzy and likely to gradually shift as capability of
the tool increases and it obtains additional capabilities
We can Always just Turn it Off A very common argument of AI risk skeptics is that any
misbehaving AI can be simply turned off, so we have nothing to worry about [71]. If skeptics
realize that modern computer viruses are a subset of very low capability malevolent AIs it becomes
obvious why saying “just turn it off” may not be a practical solution.
We Can Reprogram AIs if We Don’t Like What They Do Similar to the idea of turning AI off,
is the idea that we can reprogram AIs if we are not satisfied with their performance [72]. Such “in
production” correction is equally hard to accomplish as it can be shown to be equivalent to shutting
current AI off.
AI doesn’t have a Body and so can’t Hurt Us This is a common argument and it completely
ignores the realities of modern ultra-connected world. Given simple access to the internet, it is
easy to affect the world via hired help, digital currencies, internet of things, cyberinfrastructure or
even DNA synthesis [73].
If AI is as Capable as You Say, it Will not Make Dumb Mistakes How can superintelligence
not understand what we really want? This seems like a paradox [21], any system worthy of the
title “human-level” must have the same common sense as we do [74]. Unfortunately, an AI could
be a very powerful optimizer while at the same time not being aligned with goals of humanity [75,
Superintelligence Would (Probably) Not Be Catastrophic Not quite benevolent, but
superintelligence would not be very dangerous by default, or at least the dangers would not be
catastrophic [77] or its behavior would be correctable in time and is unlikely to be malevolent if
not explicitly programmed to be [78]. Some of the ideas in [77] are analyzed in the highly relevant
paper on modeling and interpreting expert disagreement about AI [79].
Self-preservation and Control Drives Don't Just Appear They Have to be Programmed In
LeCun has publicly argued that “the desire to control access to resources and to influence others
are drives that have been built into us by evolution for our survival. There is no reason to build
these drives into our AI systems. Some have said that such drives will spontaneously appear as
sub-goals of whatever objective we give to our AIs. Tell a robot "get me coffee" and it will destroy
everything on its path to get you coffee, perhaps figuring out in the process how to prevent every
other human from turning it off. We would have to simultaneously be extremely talented engineers
to build such an effective goal-oriented robot, and extremely stupid and careless engineers to not
put any obvious safeguards into its objective to ensure that it behaves properly.
This dismisses
research which indicates that such AI drives do appear due to game theoretic and economic reasons
An AI is not Pulled at Random from the Mind Design Space Kruel has previously argued that
“[a]n AI is the result of a research and development process. A new generation of AI’s needs to be
better than other products at “Understand What Humans Mean” and “Do What Humans Mean” in
order to survive the research phase and subsequent market pressure.” [81]. Of course, being better
doesn’t mean being perfect of event great, almost all existing software is evidence of very poor
quality of the software research/development process.
AI Can’t Generate Novel Plans As originally stated by Ada Lovelace: “The Analytical Engine
has no pretensions whatever to originate anything. It can do whatever we know how to order it to
perform. It can follow analysis; but it has no power of anticipating any analytical relations or truths.
Its province is to assist us to making available what we are already acquainted with.” [82]. Of
course, numerous counterexamples from modern AI [83] systems provide a counterargument by
existence. This doesn’t stop modern scholars from making similar claims, specifically arguing that
only humans can have curiosity, imagination, intuition, emotions, passion, desires, pleasure,
aesthetics, joy, purpose, objectives, goals, telos, values, morality, experience, wisdom, judgment,
and even humor. [39]. Regardless of ongoing work [84], most AI safety researchers are not
worried about deadly superintelligence not having a superior sense of humor.
3.3 AI Safety Related Objections
AI Safety Can’t be Done Today Some people may agree with concerns about superintelligence
but argue that AI Safety work is not possible in the absence of a superintelligent AI on which to
run experiments [38]. This view is contradicted by a significant number of publications produced
by the AI Safety community in recent years, and the author of this article (and his co-authors) in
particular [8, 19, 85-90].
AI Can’t be Safe Another objection to doing AI Safety work is based on publications showing
that fundamental aspects of the control problem [91], such as containment [92], verification [16],
or morality [93] are simply impossible to solve and so such research is a wasted effort. Solvability
of the control problem in itself is one of the most important open questions in AI Safety, but not
trying is the first step towards failure.
Skepticism of Particular Risks Even people troubled by some AI Risks may disagree about
specific risks they are concerned about and may disagree on safety methods to implement, which
ones are most likely to be beneficial and which ones are least likely to have undesirable side effects.
This is something only additional research can help resolve.
Skepticism of Particular Safety Methods AI companies may be dismissive of effectiveness of
risk mitigation technology developed by their competitors in the hopes of promoting and
standardizing their own technology [56]. Such motivated skepticism should be dismissed.
Skepticism of Researching Impossibility Results Doing work on theoretical impossibility results
[94-96] in AI safety may not translate to problems in practice, or to at least not be as severe as
predicted. However, such research may cause reductions in funding for safety work or to cause
new researchers to stay away from the field of AI safety, but this is not an argument against
importance of AI risk research in general.
3.4 Ethical Objections
Superintelligence is Benevolence Scholars observed that as humans became more advanced
culturally and intellectually they also became nicer, less violent, and more inclusive [48]. Some
have attempted to extrapolate from that pattern to superintelligent advanced AIs that they will also
be benevolent to us and our habitat [97] and will not develop their own goals which are not
programmed into them explicitly. However, superintelligence doesn’t imply benevolence [98],
which is directly demonstrated by the Bostrom’s Orthogonality Thesis [75, 76].
Let the Smarter Beings Win This type of skeptic doesn’t deny that superintelligent system will
present a lot of risk to humanity but argue that if humanity is replaced with a more advanced
sentient beings it will be an overall good thing. They give very little value to humanity and see
people as mostly having a negative impact on the planet and cognition in the universe. Similarly,
AI rights advocates argue that we should not foist our values on our mind children because it would
be a type of forced assimilation. Majority of AI researchers don’t realize that people with such
views are real, but they are and some are also AI researchers. For example, de Garis [43] has
argued that humanity should make room for superintelligent beings. Majority of humanity is not
on board with such self-destructive outcomes, perhaps because of a strong inherent pro-human
Let’s Gamble In the vast space of possible intelligences [99] some are benevolent, some are
neutral and others are malicious. It has been suggested that only a small subset of AIs are strictly
malevolent and so we may get lucky and produce a neutral or beneficial superintelligence by pure
chance. Gambling with future of human civilization doesn’t seem like a good proposition.
Malevolent AI is not worse than Malevolent Humans The argument is that it doesn’t matter
who is behind malevolent action, human actors or AI, the impact is the same [33]. Of course, a
more intelligent and so more capable AI can be much more harmful and is harder to defeat with
human resources, which are frequently sufficient to counteract human adversaries. AI is also likely
to be cognitively different from humans and so find surprising ways to cause harm.
3.5 Biased Objections
AI Safety Researchers are Non-Coders An argument is frequently made that since many top AI
Safety researchers do not write code, they are unqualified to judge AI Risk or its correlates
However, one doesn’t need to write code in order to understand the inherent risk of AGI, just like
someone doesn’t have to work in a wet lab to understand dangers of pandemics from biological
Majority of AI Researchers is not Worried To quote from Debhashi and Lappin - While it is
difficult to compute a meaningful estimate of the probability of the singularity, the arguments here
suggest to us that it is exceedingly small, at least within the foreseeable future, and this is the view
of most researchers at the forefront of AI research.[100]. Not only does this misrepresent actual
views of actual AI researchers [101, 102], it is also irrelevant, even if 100% of mathematicians
believed 2 + 2 = 5, it would still be wrong. Scientific facts are not determined by democratic
process, and you don’t get to vote on reality or truth.
Anti-Media Bias Because of how the media sensationalizes coverage of AI Safety issues, it is also
likely that many AI Researchers have Terminator-aversion, subconsciously or explicitly equating
all mentions of AI Risk with pseudoscientific ideas from Hollywood blockbusters. While, literal
Terminators are of little concern to the AI safety community, AI weaponized for military
purposes is a serious challenge to human safety.
Keep it Quiet It has been suggested that bringing up concerns about AI risk may jeopardize AI
research funding and bring on government regulation. Proponents argue that it is better to avoid
public discussions of AI risk and capabilities, which advanced AI may bring, as it has potential of
bringing on another AI “winter”. There is also some general concern about the reputation of the
field of AI [56].
Safety Work just Creates an Overhead Slowing Down Research Some developers are
concerned that integrating AI safety into research will create a significant overhead and make their
projects less competitive. The worry is that groups which don’t worry about AI risk will get to
human-level AI faster and cheaper. This is similar to cost cutting measure in software
development, where security concerns are sacrificed, to be the first to the market.
Heads in the Sand An objection from Turing’s classic paper [57] arguing that “The consequences
of machines thinking would be too dreadful. Let us hope and believe that they cannot do so.” And
his succinct response “I do not think that this argument is sufficiently substantial to require
refutation. [57]. In the same paper Turing describes and appropriately dismisses a number of
common objections to the possibility of machines achieving human level performance in thinking:
The Theological Objection, The Mathematical Objection, the Argument from Various Disabilities,
Lady Lovelace’s Objection, Argument from Continuity of the Nervous System, The Argument
from Informality of Behavior, and even the Argument from Extrasensory Perception [57].
3.6 Miscellaneous Objections
So Easy it will be Solved Automatically Some scholars think that the AI risk problem is trivial
and will be implicitly solved as a byproduct of doing regular AI research [103]. Same flawed logic
can be applied to other problems such as cybersecurity, but of course, they never get completely
solved, even with significant effort.
AI Regulation Will Prevent Problems The idea is that we don’t need to worry about AI Safety
because government regulation will intervene and prevent problems. Given how poorly legislation
against hacking, computer viruses or even spam has performed it seems unreasonable to rely on
such measures for prevention of AI risk.
Other Arguments
There are many other arguments by AI risk skeptics, which are so weak they are not worth
describing, but the names of arguments hint at their quality, for example: The arguments from
Wooly Definitions, Einstein’s Cat, Emus, Slavic Pessimism, My Roommate, Gilligan’s Island,
Transhuman Woodoo, and Comic Books [38]. Luckily, others have taken the time to address them
[104, 105] so we did not have to.
Russell provides examples of what he calls “Instantly regrettable remarks”, statements from AI
researchers which they are likely to retract after some retrospection [54]. He follows each one with
a refutation, but that seems unnecessary given low quality of the original statements:
“Electronic calculators are superhuman at arithmetic. Calculators didn’t take over the
world; therefore, there is no reason to worry about superhuman AI.
“Horses have superhuman strength, and we don’t worry about proving that horses are safe;
so we needn’t worry about proving that AI systems are safe.”
Historically, there are zero examples of machines killing millions of humans, so, by
induction, it cannot happen in the future.
No physical quantity in the universe can be infinite, and that includes intelligence, so
concerns about superintelligence are overblown.
“We don’t worry about species-ending but highly unlikely possibilities such as black holes
materializing in near-Earth orbit, so why worry about superintelligent AI?
While aiming for good coverage of the topic of AI risk skepticism we have purposefully stopped
short of analyzing every variant of the described main types of arguments as the number of such
objections continues to grow exponentially and it is not feasible or even desirable to include
everything into a survey. Readers who want to get deeper into the debate may enjoy the following
articles [106-114]/videos [115, 116]. In our future work we may provide additional analysis of the
following objections:
Bringing up concerns about AGI may actively contribute to the public misunderstanding
of science and by doing so contribute to general science denialism.
Strawman objections: "The thought that these systems would wake up and take over the
world is ludicrous." [29].
We will never willingly surrender control to machines.
While AGI is likely, superintelligence is not.
Risks from AI are minuscule in comparison to benefits (immortality, free labor, etc.) and
so can be ignored.
Intelligence is not a single dimension, so “smarter than humans” is a meaningless
Humans do not have general purpose minds, and neither will AIs.[117].
Emulation of human thinking in other media will be constrained by cost.[117].
Dimensions of intelligence are not infinite.[117].
Intelligences are only one factor in progress.[117].
You can’t control research or ban AI [54].
Malevolent use of AI is a human problem, not a computer problem [47].
“Speed alone does not bring increased intelligence” [118].
Not even exponential growth of computational power can reach the level of
superintelligence [119].
AI risk researchers are uneducated/conspiracy theorists/crazy/etc, so they are wrong.
AI has been around for 65 years and didn’t destroy humanity, it is unlikely to do so in the
AI risk is science fiction.
Just box it; just give it laws to follow; just raise it as a human baby; just …
AI is just a tool, it can’t generate its own goals because it is not conscious.
I don’t want to make important AI researchers angry at me and retaliate against me.
Narrow AI/robots can’t even do some basic thing, certainly they can’t present danger to
Real threat is AI being too dumb and making mistakes.
Certainly, many smart people are already working on AI safety they will take care of it.
Big companies like Google or Microsoft would never release a dangerous product or
service which may damage their reputation or reduce profits.
Smartest person in the world, [multiple names are used by proponents], is not worried about
it so it must not be a real problem.
4. Countermeasures for AI Risk Skepticism
First, it is important to emphasize that just like with any other product or service the burden of
proof [120] is on the developers/manufacturers (frequently AI Risk skeptics) to show that their AI
will be safe and secure regardless of its capability, customization, learning, domain of utilization
or duration of use. Proving that an intelligent agent in a novel environment will behave in a
particular way is a very high standard to meet. The problem could be reduced to showing that a
particular human or an animal, for example a Pit bull, is safe to everyone, a task long known to be
impractical. It seems to be even harder with much more capable agents, such as AGI. The best we
can hope for is showing some non-zero probability of safe behavior.
A capable AI researcher not concerned with safety is very dangerous. It seems that the only
solution to reduce prevalence of AI risk denialism is education. It is difficult for a sharp mind to
study the best AI risk literature and to remain unconvinced of scientific merits behind it. The
legitimacy of risk from uncontrolled AI is undeniable. This is not fear mongering, we don’t have
an adequate amount of fear in the AI researcher community, an amount which would be necessary
to make sure that sufficient precautions are taken by everyone involved. Education is likewise
suggested as a desirable path forward by the skeptics, so all sides agree on importance of education.
Perhaps if we were to update and de-bias recommendations from the 2009 AAAI presidential panel
on Long-Term AI Futures to look like this: The group suggested outreach and communication to
people and organizations about the low likelihood of the radical outcomes, sharing the rationale
for the overall comfort [position] of scientists in this realm, and for the need to educate people
outside the AI research community about the promise of AI[45], we could make some progress
on AI risk denialism reduction.
The survival of humanity could depend on rejecting superintelligence misinformation [52]. Two
main strategies could be identified: those aimed at preventing spread of misinformation and those
designed to correct peoples’ understanding after exposure to misinformation. Baum reviews some
ways to prevent superintelligence misinformation, which would also apply to reducing AI Risk
skepticism [52]: educate prominent voices, create reputation costs, mobilize against institutional
misinformation, focus media attention on constructive debate, establish legal requirements. For
correcting superintelligence misinformation Baum suggests: building expert consensus and the
perception of thereof, address pre-existing motivations for believing misinformation, inoculate
with advance warnings, avoid close association with polarizing ideas, explain misinformation and
corrections [52].
Specifically for politicized superintelligence skepticism Baum suggests [56]: With this in mind,
one basic opportunity is to raise awareness about politicized skepticism within communities that
discuss superintelligence. Superintelligence skeptics who are motivated by honest intellectual
norms may not wish for their skepticism to be used politically. They can likewise be cautious about
how to engage with potential political skeptics, such as by avoiding certain speaking opportunities
in which their remarks would be used as a political tool instead of as a constructive intellectual
contribution. Additionally, all people involved in superintelligence debates can insist on basic
intellectual standards, above all by putting analysis before conclusions and not the other way
around. These are the sorts of things that an awareness of politicized skepticism can help with.
Baum also recommends [56] to: redouble efforts to build scientific consensus on
superintelligence, and then to draw attention to it”, engage with AI corporations to encourage
them to avoid politicizing skepticism about superintelligence or other forms of AI”, and “follow
best practices in debunking misinformation in the event that superintelligence skepticism is
politicized.“… Finally, the entire AI community should insist that policy be made based on an
honest and balanced read of the current state of knowledge. Burden of proof requirements should
not be abused for private gain. As with climate change and other global risks, the world cannot
afford to prove that superintelligence would be catastrophic. By the time uncertainty is eliminated,
it could be too late.” [56].
AI risk education research [121] indicates that most AI risk communication strategies are effective
[122] and are not counter-productive and the following “good practices” work well for introducing
general audiences to AI risk [121]:
1. Allow the audience to engage in guided thinking on the subject (“What do you think the
effects of human-level AI will be?”), but do not neglect to emphasize its technical nature
2. Reference credible individuals who have spoken about AI risk (such as Stephen
Hawking, Stuart Russell, and Bill Gates)
3. Reference other cases of technological risk and revolution (such as nuclear energy and
the Industrial Revolution)
4. Do not reference science-fiction stories, unless, in context, you expect an increase in the
audience’s level of engagement to outweigh a drop in their perceptions of the field’s
importance and its researchers’ credibility
5. Do not present overly vivid or grave disaster scenarios
6. Do not limit the discussion to abstractions (such as “optimization,” “social structures,”
and “human flourishing”), although they may be useful for creating impressions of
Recent research indicates that individual differences in the AI risk perception may be personality
[123] and/or attitude [124, 125] dependent but are subject to influence by experts [126] and choice
of language [127].
Healthy skepticism is important to keep scientists, including AI researchers honest. For example,
during early days of AI research it was predicted that human level performance will be quickly
achieved [128]. Luckily, a number of skeptics [129, 130] argued that perhaps the problem is not
as simple as it seems, bringing some conservativism to the overly optimistic predictions of
researchers and as a result improving quality of research actually being funded and conducted by
AI researchers. For a general overview of threat inflation Thierer’s work on technopanics [131] is
a good reference.
7. Conclusions
In this paper, we didn’t reiterate most of the overwhelming evidence for AI Risk concerns, as it
was outside of our goal of analyzing AI Risk skepticism. Likewise, we did not go in depth with
rebuttals to every type of objections to AI Risk. It is precisely because of skeptic attitudes from
the majority of mainstream AI researchers the fields of AI Safety was born outside of academia
[132]. Regardless, AI Risk skeptics need to realize that the burden of proof is not on AI Safety
researchers to show that technology may be dangers but on AI developers to establish that their
technology is safe at the time of deployment and throughout its lifetime of operation. Furthermore,
while science operates as a democracy (via majority of peer-reviewers), the facts are not subject
to a vote. Even if AI Safety researchers comprise only a small minority of the total number of AI
researchers that says nothing about the true potential of intelligent systems for harmful actions.
History is full of examples (continental drift [133], quantum mechanics [134]) in which a majority
of scientists held a wrong view right before a paradigm shift in thinking took place. Since, just like
AI Skeptics, AI Safety researchers also have certain biases, to avoid pro or con prejudice in
judgment it may be a good idea to rely on impartial juries of non-peers (scientists from outside the
domain) whose only job would be to evaluate evidence for a particular claim.
It is obvious that designing a Safe AI is a much harder problem than designing an AI and so will
take more time. The actual time to human level AI is irrelevant; it will always take longer to make
such an AI human friendly. To move the Overton window on AI Risk, AI Safety researchers have
to be non-compromising in their position. Perhaps a temporary moratorium on AGI (but not AI)
research similar to the one in place for human cloning needs to be considered. It would boost our
ability to engage in differential technological development [135-137] increasing our chances of
making the AGI safe. AI Safety research definitely needs to get elevated priority and more
resources including funding and human capital. Perhaps AI Safety researchers could generate
funding via economic incentives from developing safer products. It may be possible to market
“Safe AI Inside” government certification on selected progressively ever-smarter devices to boost
consumer confidence and sales. This would probably require setting up FDA for algorithms
[138, 139].
Scientific skepticism in general and skepticism about predicted future events is of course
intellectually defensible and is frequently desirable to protect again flawed theories [131].
However, it is important to realize that 100% proof is unlikely to be obtained in some domains and
so a Precautionary Principle (PP) [140] should be used to protect humanity against existential risks.
Holm and Harris, in their skeptical paper, define PP as follows [141]: When an activity raises
threats of serious or irreversible harm to human health or the environment, precautionary measures
that prevent the possibility of harm shall be taken even if the causal link between the activity and
the possible harm has not been proven or the causal link is weak and the harm is unlikely to occur.
To use a stock market metaphor, no matter how great a return on investment one is promised, one
should not ignore the possibility of losing the principal.
The author is grateful to Seth Baum for sharing a lot of relevant literature and providing feedback
on an early draft of this paper. In addition, author would like to acknowledge his own bias, as an
AI safety researcher I would benefit from flourishing of the field of AI safety. I also have a conflict
of interest, as a human being with a survival instinct I would benefit from not being exterminated
by uncontrolled AI.
1. Kurzweil, R., The Singularity is Near: When Humans Transcend Biology. 2005: Viking
2. Yampolskiy, R.V., Predicting future AI failures from historic examples. foresight, 2019.
21(1): p. 138-152.
3. Caliskan, A., J.J. Bryson, and A. Narayanan, Semantics derived automatically from language
corpora contain human-like biases. Science, 2017. 356(6334): p. 183-186.
4. Bolukbasi, T., et al. Man is to computer programmer as woman is to homemaker? debiasing
word embeddings. in Advances in Neural Information Processing Systems. 2016.
5. Arkin, R., Governing lethal behavior in autonomous robots. 2009: CRC Press.
6. Fast, E. and E. Horvitz, Long-term trends in the public perception of artificial intelligence.
arXiv preprint arXiv:1609.04904, 2016.
7. Bostrom, N., Superintelligence: Paths, dangers, strategies. 2014: Oxford University Press.
8. Yampolskiy, R.V., Artificial superintelligence: a futuristic approach. 2015: cRc Press.
9. Sotala, K. and R.V. Yampolskiy, Responses to catastrophic AGI risk: a survey. Physica
Scripta, 2014. 90(1): p. 018001.
10. Yampolskiy, R.V. Taxonomy of Pathways to Dangerous Artificial Intelligence. in
Workshops at the Thirtieth AAAI Conference on Artificial Intelligence. 2016.
11. Pistono, F. and R.V. Yampolskiy. Unethical Research: How to Create a Malevolent
Artificial Intelligence. in 25th International Joint Conference on Artificial Intelligence
(IJCAI-16). Ethics for Artificial Intelligence Workshop (AI-Ethics-2016). 2016.
12. Vanderelst, D. and A. Winfield, The Dark Side of Ethical Robots. arXiv preprint
arXiv:1606.02583, 2016.
13. Charisi, V., et al., Towards Moral Autonomous Systems. arXiv preprint arXiv:1703.04741,
14. Corabi, J., Superintelligent AI and Skepticism. Journal of Evolution and Technology, 2017.
27(1): p. 4.
15. Herley, C., Unfalsifiability of security claims. Proceedings of the National Academy of
Sciences, 2016. 113(23): p. 6415-6420.
16. Yampolskiy, R.V., What are the ultimate limits to computational techniques: verifier theory
and unverifiability. Physica Scripta, 2017. 92(9): p. 093001.
17. Babcock, J., J. Kramar, and R. Yampolskiy, The AGI Containment Problem, in The Ninth
Conference on Artificial General Intelligence (AGI2015). July 16-19, 2016: NYC, USA.
18. Callaghan, V., et al., Technological Singularity. 2017: Springer.
19. Ramamoorthy, A. and R. Yampolskiy, Beyond mad? the race for artificial general
intelligence. ITU J, 2018. 1: p. 1-8.
20. Majot, A.M. and R.V. Yampolskiy. AI safety engineering through introduction of self-
reference into felicific calculus via artificial pain and pleasure. in Ethics in Science,
Technology and Engineering, 2014 IEEE International Symposium on. 2014. IEEE.
21. Yampolskiy, R.V., What to Do with the Singularity Paradox?, in Philosophy and Theory of
Artificial Intelligence (PT-AI2011). October 3-4, 2011: Thessaloniki, Greece.
22. Tegmark, M., Life 3.0: Being human in the age of artificial intelligence. 2017: Knopf.
23. Everitt, T., G. Lea, and M. Hutter, AGI safety literature review. arXiv preprint
arXiv:1805.01109, 2018.
24. Juric, M., A. Sandic, and M. Brcic, AI safety: state of the field through quantitative lens.
arXiv preprint arXiv:2002.05671, 2020.
25. Anonymous, Existential risk from artificial general intelligence - Skepticism, in Wikipedia.
Retrieved September 16, 2002: Available at:
26. Elkus, A., A Rebuttal to a Rebuttal on AI Values. April 27, 2016: Available at:
27. Doctorow, C., AI Alarmism: why smart people believe dumb things about our future AI
overlords. December 23, 2016: Available at:
28. Radu, S., Artificial Intelligence Alarmists Win ITIF’s Annual Luddite Award, in Information
Technology & Innovation Foundation. January 19, 2016: Available at:
29. Togelius, J., How many AGIs can dance on the head of a pin? October 30, 2020: Available
30. Atkinson, R.D., 'It's Going to Kill Us!'And Other Myths About the Future of Artificial
Intelligence. Information Technology & Innovation Foundation, 2016.
31. Brown, J.S. and P. Duguid, A response to Bill Joy and the doom-and-gloom technofuturists.
AAAS Science and Technology Policy Yearbook, 2001: p. 77-83.
32. Johnson, D.G. and M. Verdicchio, AI anxiety. Journal of the Association for Information
Science and Technology, 2017. 68(9): p. 2267-2270.
33. Lanier, J., The Myth of AI, in Edge. November 14, 2014: Available at:
34. Alfonseca, M., et al., Superintelligence cannot be contained: Lessons from Computability
Theory. Journal of Artificial Intelligence Research, 2021. 70: p. 65-76.
35. Voss, P., AI Safety Research: A Road to Nowhere. October 19, 2016: Available at:
36. Danaher, J., Why AI doomsayers are like sceptical theists and why it matters. Minds and
Machines, 2015. 25(3): p. 231-246.
37. McCauley, L. Countering the Frankenstein Complex. in AAAI spring symposium:
Multidisciplinary collaboration for socially assistive robotics. 2007.
38. Ceglowski, M., Superintelligence: The Idea That Eats Smart People, in Web Camp Zagreb.
October 29, 2016: Available at:
39. Braga, A. and R.K. Logan, The emperor of strong AI has no clothes: Limits to artificial
intelligence. Information, 2017. 8(4): p. 156.
40. Etzioni, O., No, the Experts Don’t Think Superintelligent AI is a Threat to Humanity, in MIT
Technology Review. September 20, 2016: Available at:
41. Weld, D.S. and O. Etzioni, The First Law of Robotics (a Call to Arms), in Twelfth National
Conference on Artificial Intelligence (AAAI). 1994. p. 1042-1047.
42. Dafoe, A. and S. Russell, Yes, We Are Worried About the Existential Risk of Artificial
Intelligence, in MIT Technology Review. November 2, 2016: Available at:
43. Garis, H.d. The Artilect War. 2005. ETC publications.
44. Babcock, J., J. Kramár, and R.V. Yampolskiy, Guidelines for artificial intelligence
containment, in Next-Generation Ethics: Engineering a Better Society, A.E. Abbas, Editor.
2019. p. 90-112.
45. Horvitz, E. and B. Selman, Interim Report from the AAAI Presidential Panel on Long-Term
AI Futures. August 2009: Available at
46. Yudkowsky, E. and R. Hanson, The Hanson-Yudkowsky AI-foom debate, in MIRI Technical
Report. 2008: Available at:
47. Shermer, M., Why artificial intelligence is not an existential threat. Skeptic (Altadena, CA),
2017. 22(2): p. 29-36.
48. Pinker, S., The better angels of our nature: Why violence has declined. 2012: Penguin Group
49. Yampolskiy, R.V., Artificial Intelligence Safety and Security. 2018: Chapman and
50. Benthall, S., Don't Fear the Reaper: Refuting Bostrom's Superintelligence Argument. arXiv
preprint arXiv:1702.08495, 2017.
51. Wiblin, R. and K. Harris, DeepMind’s plan to make AI systems robust & reliable, why it’s a
core issue in AI design, and how to succeed at AI research. June 3, 2019: Available at:
52. Baum, S.D., Countering Superintelligence Misinformation. Information, 2018. 9(10): p. 244.
53. Russell, S., Provably beneficial artificial intelligence. Exponential Life, The Next Step,
54. Russell, S., Human compatible: Artificial intelligence and the problem of control. 2019:
55. Aronson, J., Five types of skepticism. Bmj, 2015. 350: p. h1986.
56. Baum, S., Superintelligence skepticism as a political tool. Information, 2018. 9(9): p. 209.
57. Turing, A., Computing Machinery and Intelligence. Mind, 1950. 59(236): p. 433-460.
58. Alexander, S., AI Researchers on AI Risk. May 22, 2015: Available at:
59. Brown, T.B., et al., Language models are few-shot learners. arXiv preprint
arXiv:2005.14165, 2020.
60. Kaplan, J., et al., Scaling laws for neural language models. arXiv preprint arXiv:2001.08361,
61. Henighan, T., et al., Scaling Laws for Autoregressive Generative Modeling. arXiv preprint
arXiv:2010.14701, 2020.
62. Bundy, A., Smart machines are not a threat to humanity. Communications of the ACM,
2017. 60(2): p. 40-42.
63. Bringsjord, S., A. Bringsjord, and P. Bello, Belief in the singularity is fideistic, in Singularity
Hypotheses. 2012, Springer. p. 395-412.
64. Bringsjord, S., Belief in the singularity is logically brittle. Journal of Consciousness Studies,
2012. 19(7): p. 14.
65. Modis, T., Why the singularity cannot happen, in Singularity Hypotheses. 2012, Springer. p.
66. Yampolskiy, R.V., On the Limits of Recursively Self-Improving AGI. Artificial General
Intelligence: 8th International Conference, AGI 2015, AGI 2015, Berlin, Germany, July 22-
25, 2015, Proceedings, 2015. 9205: p. 394.
67. Bostrom, N., Taking intelligent machines seriously: Reply to critics. Futures, 2003. 35(8): p.
68. Logan, R.K., Can computers become conscious, an essential condition for the singularity?
Information, 2017. 8(4): p. 161.
69. Chalmers, D.J., The conscious mind: In search of a fundamental theory. 1996: Oxford
University Press.
70. Yampolskiy, R.V., Artificial Consciousness: An Illusionary Solution to the Hard Problem.
Reti, saperi, linguaggi, 2018(2): p. 287-318.
71. Hawkins, J., The Terminator Is Not Coming. The Future Will Thank Us. March 2, 2015:
Available at:
72. Kelly, K., Why I don't fear super intelligence (Comments section), in Edge. November 14,
2014: Available at:
73. Yudkowsky, E., Artificial Intelligence as a Positive and Negative Factor in Global Risk, in
Global Catastrophic Risks, N. Bostrom and M.M. Cirkovic, Editors. 2008, Oxford
University Press: Oxford, UK. p. 308-345.
74. Loosemore, R.P. The Maverick Nanny with a Dopamine Drip: Debunking Fallacies in the
Theory of AI Motivation. in 2014 AAAI Spring Symposium Series. 2014.
75. Armstrong, S., General purpose intelligence: arguing the orthogonality thesis. Analysis and
Metaphysics, 2013(12): p. 68-84.
76. Miller, J.D., R. Yampolskiy, and O. Häggström, An AGI Modifying Its Utility Function in
Violation of the Orthogonality Thesis. arXiv preprint arXiv:2003.00812, 2020.
77. Goertzel, B., Superintelligence: Fears, promises and potentials. Journal of Evolution and
Technology, 2015. 25(2): p. 55-87.
78. Searle, J.R., What your computer can’t know. The New York review of books, 2014. 9.
79. Baum, S., A. Barrett, and R.V. Yampolskiy, Modeling and interpreting expert disagreement
about artificial superintelligence. Informatica, 2017. 41(7): p. 419-428.
80. Omohundro, S.M., The Basic AI Drives, in Proceedings of the First AGI Conference, Volume
171, Frontiers in Artificial Intelligence and Applications, P. Wang, B. Goertzel, and S.
Franklin (eds.). February 2008, IOS Press.
81. Kruel, A., Four Arguments Against AI Risk. July 11, 2013: Available at:
82. Toole, B.A., Ada, the Enchantress of Numbers: Poetical Science. 2010: Betty Alexandra
83. Ecoffet, A., et al., First return, then explore. Nature, 2021. 590(7847): p. 580-586.
84. Binsted, K., et al., Computational humor. IEEE Intelligent Systems, 2006. 21(2): p. 59-69.
85. Majot, A.M. and R.V. Yampolskiy. AI safety engineering through introduction of self-
reference into felicific calculus via artificial pain and pleasure. in 2014 IEEE International
Symposium on Ethics in Science, Technology and Engineering. 2014. IEEE.
86. Brundage, M., et al., The malicious use of artificial intelligence: Forecasting, prevention,
and mitigation. arXiv preprint arXiv:1802.07228, 2018.
87. Aliman, N.-M., L. Kester, and R. Yampolskiy, Transdisciplinary AI Observatory
Retrospective Analyses and Future-Oriented Contradistinctions. Philosophies, 2021. 6(1): p.
88. Ziesche, S. and R. Yampolskiy. Introducing the concept of ikigai to the ethics of AI and of
human enhancements. in 2020 IEEE International Conference on Artificial Intelligence and
Virtual Reality (AIVR). 2020. IEEE.
89. Miller, J.D., R. Yampolskiy, and O. Häggström, An AGI Modifying Its Utility Function in
Violation of the Strong Orthogonality Thesis. Philosophies, 2020. 5(4): p. 40.
90. Williams, R.M. and R.V. Yampolskiy, Understanding and Avoiding AI Failures: A Practical
Guide. April 30, 2021: Available at:
91. Yampolskiy, R.V., On Controllability of AI. arXiv preprint arXiv:2008.04071, 2020.
92. Yampolskiy, R.V., Leakproofing Singularity-Artificial Intelligence Confinement Problem.
Journal of Consciousness Studies JCS, 2012.
93. Brundage, M., Limitations and risks of machine ethics. Journal of Experimental &
Theoretical Artificial Intelligence, 2014. 26(3): p. 355-372.
94. Yampolskiy, R.V., Unexplainability and Incomprehensibility of AI. Journal of Artificial
Intelligence and Consciousness, 2020. 7(02): p. 277-291.
95. Yampolskiy, R.V., Unpredictability of AI: On the Impossibility of Accurately Predicting All
Actions of a Smarter Agent. Journal of Artificial Intelligence and Consciousness, 2020.
7(01): p. 109-118.
96. Howe, W.J. and R.V. Yampolskiy, Impossibility of Unambiguous Communication as a
Source of Failure in AI Systems. 2020: Available at:
97. Waser, M.R., Wisdom Does Imply Benevolence, in First International Conference of IACAP.
July 4-6, 2011: Aarhus University. p. 148-150.
98. Fox, J. and C. Shulman, Superintelligence Does Not Imply Benevolence, in 8th European
Conference on Computing and Philosophy. October 4-6, 2010 Munich, Germany.
99. Yampolskiy, R.V., The Space of Possible Mind Designs, in Artificial General Intelligence.
2015, Springer. p. 218-227.
100. Dubhashi, D. and S. Lappin, AI dangers: Imagined and real. Communications of the ACM,
2017. 60(2): p. 43-45.
101. Grace, K., et al., When will AI exceed human performance? Evidence from AI experts.
Journal of Artificial Intelligence Research, 2018. 62: p. 729-754.
102. Müller, V.C. and N. Bostrom, Future progress in artificial intelligence: A survey of expert
opinion, in Fundamental issues of artificial intelligence. 2016, Springer. p. 555-572.
103. Khatchadourian, R., The Doomsday Invention, in New Yorker. November 23, 2015:
104. Cantor, L., Superintelligence: The Idea That Smart People Refuse to Think About. December
24, 2016: Available at:
105. Graves, M., Response to Cegłowski on superintelligence. January 13, 2017: Available at:
106. Kruel, A., Why I am skeptical of risks from AI. July 21, 2011: Available at:
107. Wilks, Y., Will There Be Superintelligence and Would It Hate Us? AI Magazine, 2017.
38(4): p. 65-70.
108. Kurzweil, R., Don’t fear artificial intelligence. Time Magazine, 2014: p. 28.
109. Smith, M., Address the consequences of AI in advance. Communications of the ACM, 2017.
60(3): p. 10-11.
110. Dietterich, T.G. and E.J. Horvitz, Rise of concerns about AI: Reflections and directions.
Communications of the ACM, 2015. 58(10): p. 38-40.
111. Agar, N., Don’t worry about superintelligence. Journal of evolution and technology, 2016.
26(1): p. 73-82.
112. Yampolskiy, R.V., The singularity may be near. Information, 2018. 9(8): p. 190.
113. Sotala, K. and R. Yampolskiy, Risks of the Journey to the Singularity, in The Technological
Singularity. 2017, Springer. p. 11-23.
114. Sotala, K. and R. Yampolskiy, Responses to the Journey to the Singularity. The
Technological Singularity, 2017: p. 25-83.
115. Booch, G., Don't Fear Superintelligent AI, in TED. November 2016: Available at:
116. Etzioni, O., Artificial Intelligence will empower us, not exterminate us, in TEDx. November
2016: Available at:
117. Kelly, K., The myth of a superhuman AI, in Wired. April 15, 2017: Available at:
118. Walsh, T., The singularity may never be near. AI Magazine, 2017. 38(3): p. 58-62.
119. Wiedermann, J., A Computability Argument Against Superintelligence. Cognitive
Computation, 2012. 4(3): p. 236-245.
120. Haggstrom, O., Vulgopopperianism. February 20, 2017: Available at:
121. Garfinkel, B., A. Dafoe, and O. Catton-Barratt, A Survey on AI Risk Communication
Strategies. August 8, 2016: Available at:
122. Alexander, S., AI Persuasion Experiment Results, in Slate Start Codex. October 24, 2016:
Available at:
123. Wissing, B.G. and M.-A. Reinhard, Individual differences in risk perception of artificial
intelligence. Swiss Journal of Psychology, 2018. 77(4): p. 149.
124. Li, J. and J.-S. Huang, Dimensions of artificial intelligence anxiety based on the integrated
fear acquisition theory. Technology in Society, 2020. 63: p. 101410.
125. Chen, Y.-N.K. and C.-H.R. Wen, Impacts of Attitudes Toward Government and
Corporations on Public Trust in Artificial Intelligence. Communication Studies, 2021. 72(1):
p. 115-131.
126. Neri, H. and F. Cozman, The role of experts in the public perception of risk of artificial
intelligence. AI & SOCIETY, 2019: p. 1-11.
127. Sharkey, L., An intervention to shape policy dialogue, communication, and AI research
norms for AI safety. October 1, 2017: Available at:
128. Muehlhauser, L., What should we learn from past AI forecasts? May 2016: Available at:
129. Dreyfus, H.L., What computers can't do; A critique of artificial reason. 1972: Harper & Row.
130. Searle, J., Minds, Brains and Programs. Behavioral and Brain Sciences, 1980. 3(3): p. 417-
131. Thierer, A., Technopanics, threat inflation, and the danger of an information technology
precautionary principle. Minn. JL Sci. & Tech., 2013. 14: p. 309.
132. Yudkowsky, E.S., Creating Friendly AI - The Analysis and Design of Benevolent Goal
Architectures. 2001: Available at:
133. Hurley, P.M., The confirmation of continental drift. Scientific American, 1968. 218(4): p.
134. Vardi, M.Y., Quantum hype and quantum skepticism. Communications of the ACM, 2019.
62(5): p. 7-7.
135. Bostrom, N., Existential risks: Analyzing human extinction scenarios and related hazards.
Journal of Evolution and technology, 2002. 9.
136. Ord, T., The precipice: existential risk and the future of humanity. 2020: Hachette Books.
137. Tomasik, B., Differential Intellectual Progress as a Positive-Sum Project. Center on Long-
Term Risk, 2013.
138. Tutt, A., An FDA for algorithms. Admin. L. Rev., 2017. 69: p. 83.
139. Ozlati, S. and R. Yampolskiy. The Formalization of AI Risk Management and Safety
Standards. in Workshops at the Thirty-First AAAI Conference on Artificial Intelligence.
140. O'Riordan, T., Interpreting the precautionary principle. 2013: Routledge.
141. Holm, S. and J. Harris, Precautionary principle stifles discovery. Nature, 1999. 400(6743):
p. 398-398.
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
A long tradition in philosophy and economics equates intelligence with the ability to act rationally—that is, to choose actions that can be expected to achieve one’s objectives. This framework is so pervasive within AI that it would be reasonable to call it the standard model. A great deal of progress on reasoning, planning, and decision-making, as well as perception and learning, has occurred within the standard model. Unfortunately, the standard model is unworkable as a foundation for further progress because it is seldom possible to specify objectives completely and correctly in the real world. The chapter proposes a new model for AI development in which the machine’s uncertainty about the true objective leads to qualitatively new modes of behavior that are more robust, controllable, and deferential to humans.
Full-text available
As AI technologies increase in capability and ubiquity, AI accidents are becoming more common. Based on normal accident theory, high reliability theory, and open systems theory, we create a framework for understanding the risks associated with AI applications. This framework is designed to direct attention to pertinent system properties without requiring unwieldy amounts of accuracy. In addition, we also use AI safety principles to quantify the unique risks of increased intelligence and human-like qualities in AI. Together, these two fields give a more complete picture of the risks of contemporary AI. By focusing on system properties near accidents instead of seeking a root cause of accidents, we identify where attention should be paid to safety for current generation AI systems.
Full-text available
Reinforcement learning promises to solve complex sequential-decision problems autonomously by specifying a high-level reward function only. However, reinforcement learning algorithms struggle when, as is often the case, simple and intuitive rewards provide sparse1 and deceptive2 feedback. Avoiding these pitfalls requires a thorough exploration of the environment, but creating algorithms that can do so remains one of the central challenges of the field. Here we hypothesize that the main impediment to effective exploration originates from algorithms forgetting how to reach previously visited states (detachment) and failing to first return to a state before exploring from it (derailment). We introduce Go-Explore, a family of algorithms that addresses these two challenges directly through the simple principles of explicitly ‘remembering’ promising states and returning to such states before intentionally exploring. Go-Explore solves all previously unsolved Atari games and surpasses the state of the art on all hard-exploration games1, with orders-of-magnitude improvements on the grand challenges of Montezuma’s Revenge and Pitfall. We also demonstrate the practical potential of Go-Explore on a sparse-reward pick-and-place robotics task. Additionally, we show that adding a goal-conditioned policy can further improve Go-Explore’s exploration efficiency and enable it to handle stochasticity throughout training. The substantial performance gains from Go-Explore suggest that the simple principles of remembering states, returning to them, and exploring from them are a powerful and general approach to exploration—an insight that may prove critical to the creation of truly intelligent learning agents. A reinforcement learning algorithm that explicitly remembers promising states and returns to them as a basis for further exploration solves all as-yet-unsolved Atari games and out-performs previous algorithms on Montezuma’s Revenge and Pitfall.
Full-text available
In the last years, artificial intelligence (AI) safety gained international recognition in the light of heterogeneous safety-critical and ethical issues that risk overshadowing the broad beneficial impacts of AI. In this context, the implementation of AI observatory endeavors represents one key research direction. This paper motivates the need for an inherently transdisciplinary AI observatory approach integrating diverse retrospective and counterfactual views. We delineate aims and limitations while providing hands-on-advice utilizing concrete practical examples. Distinguishing between unintentionally and intentionally triggered AI risks with diverse socio-psycho-technological impacts, we exemplify a retrospective descriptive analysis followed by a retrospective counterfactual risk analysis. Building on these AI observatory tools, we present near-term transdisciplinary guidelines for AI safety. As further contribution, we discuss differentiated and tailored long-term directions through the lens of two disparate modern AI safety paradigms. For simplicity, we refer to these two different paradigms with the terms artificial stupidity (AS) and eternal creativity (EC) respectively. While both AS and EC acknowledge the need for a hybrid cognitive-affective approach to AI safety and overlap with regard to many short-term considerations, they differ fundamentally in the nature of multiple envisaged long-term solution patterns. By compiling relevant underlying contradistinctions, we aim to provide future-oriented incentives for constructive dialectics in practical and theoretical AI safety research.
Full-text available
An artificial general intelligence (AGI) might have an instrumental drive to modify its utility function to improve its ability to cooperate, bargain, promise, threaten, and resist and engage in blackmail. Such an AGI would necessarily have a utility function that was at least partially observable and that was influenced by how other agents chose to interact with it. This instrumental drive would conflict with the strong orthogonality thesis since the modifications would be influenced by the AGI’s intelligence. AGIs in highly competitive environments might converge to having nearly the same utility function, one optimized to favorably influencing other agents through game theory. Nothing in our analysis weakens arguments concerning the risks of AGI.
This paper responds to Nick Bostrom’s suggestion that the threat of a human-unfriendly superintelligence should lead us to delay or rethink progress in AI. I allow that progress in AI presents problems that we are currently unable to solve. However, we should distinguish between currently unsolved problems for which there are rational expectations of solutions and currently unsolved problems for which no such expectation is appropriate. The problem of a human-unfriendly superintelligence belongs to the first category. It is rational to proceed on that assumption that we will solve it. These observations do not reduce to zero the existential threat from superintelligence. But we should not permit fear of very improbable negative outcomes to delay the arrival of the expected benefits from AI.
With the rapid development of artificial intelligence (AI), AI anxiety has emerged and is receiving widespread attention, but research on this topic is not comprehensive. Therefore, we investigated the dimensions of AI anxiety using the theoretical model of integrated fear acquisition and a questionnaire survey. A total of 494 valid questionnaires were recovered. Through a first-order confirmatory factor analysis (CFA), a factor model of AI anxiety was constructed, and eight factors of AI anxiety were verified. Then, a second-order CFA was applied to verify the adaptation of the factor structure of AI anxiety to fear acquisition. We identified four dimensions of AI anxiety and proposed a theory of AI anxiety acquisition that illustrates four pathways of AI anxiety acquisition. Each pathway includes two factors that cause AI anxiety. We conclude by analyzing the limitations of current AI anxiety research and proposing a broader research agenda for AI anxiety.