About

Professor David Chalmers is a Philosophy and Neural Science Professor at NYU Arts & Science. He is also the co-director of the Center for Mind, Brain, and Consciousness. His work focuses on consciousness, exploring fundamental questions about the nature of the mind and experience. As a leading figure in his field, Professor Chalmers has made significant contributions to the understanding of consciousness and its relation to neural processes, engaging in interdisciplinary research that bridges philosophy, neuroscience, and cognitive science.

Research topics

• Computer Science
• Epistemology
• Philosophy
• Psychology
• Linguistics
• Management science
• Engineering
• Human–computer interaction

Selected publications

• Mohol by byť veľký jazykový model (LLM) vedomý?

  Filozofia · 2025-09-16

  articleOpen access1st authorCorresponding

  There has recently been widespread discussion of whether large language models such as the GPT systems might be sentient or conscious. Should we take this idea seriously? I will discuss the underlying issue and will break down the strongest reasons for and against.

  PublisherDOI

• When awareness outstrips performance: critical tests of subjective inflation under inattention

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-07-04 · 2 citations

  preprintOpen access

  Abstract Visual experience can sometimes depart from visual performance, providing a powerful lens into the mechanisms generating conscious perception. In one proposed dissociation—subjective inflation—unattended locations in the periphery appear stronger than attended ones despite equated performance. Subjective inflation has played a central role in motivating theories of consciousness that reject the sufficiency of sensory signals for conscious perception. Yet the empirical basis for subjective inflation is limited. Here, in a large-scale adversarial collaboration, we conducted four simultaneously-replicated experiments testing the strength, character, and extent of subjective inflation under inattention. We used a new analytic approach to quantify inattentional inflation over full psychometric functions, beyond single matched-performance levels. We found robust inattentional inflation for contrast-dependent and texture-based perception, at and above the visual threshold. However at suprathreshold, we found inattentional inflation for the overall stimulus but not the specific feature relevant for performance. Finally, we establish the unifying principle that inattentional inflation occurs if and only if attention reduces performance thresholds more than visibility thresholds. Thus what we think we see may regularly exceed what we can visually discriminate, placing constraints on theories of conscious perception.

  PublisherOA PDFDOI

• Propositional Interpretability in Artificial Intelligence

  ArXiv.org · 2025-01-27 · 2 citations

  preprintOpen access1st authorCorresponding

  Mechanistic interpretability is the program of explaining what AI systems are doing in terms of their internal mechanisms. I analyze some aspects of the program, along with setting out some concrete challenges and assessing progress to date. I argue for the importance of propositional interpretability, which involves interpreting a system's mechanisms and behavior in terms of propositional attitudes: attitudes (such as belief, desire, or subjective probability) to propositions (e.g. the proposition that it is hot outside). Propositional attitudes are the central way that we interpret and explain human beings and they are likely to be central in AI too. A central challenge is what I call thought logging: creating systems that log all of the relevant propositional attitudes in an AI system over time. I examine currently popular methods of interpretability (such as probing, sparse auto-encoders, and chain of thought methods) as well as philosophical methods of interpretation (including those grounded in psychosemantics) to assess their strengths and weaknesses as methods of propositional interpretability.

  PublisherOA PDFDOI

• Adversarial testing of global neuronal workspace and integrated information theories of consciousness

  Nature · 2025-04-30 · 77 citations

  articleOpen access

  Abstract Different theories explain how subjective experience arises from brain activity 1,2 . These theories have independently accrued evidence, but have not been directly compared 3 . Here we present an open science adversarial collaboration directly juxtaposing integrated information theory (IIT) 4,5 and global neuronal workspace theory (GNWT) 6–10 via a theory-neutral consortium 11–13 . The theory proponents and the consortium developed and preregistered the experimental design, divergent predictions, expected outcomes and interpretation thereof 12 . Human participants ( n = 256) viewed suprathreshold stimuli for variable durations while neural activity was measured with functional magnetic resonance imaging, magnetoencephalography and intracranial electroencephalography. We found information about conscious content in visual, ventrotemporal and inferior frontal cortex, with sustained responses in occipital and lateral temporal cortex reflecting stimulus duration, and content-specific synchronization between frontal and early visual areas. These results align with some predictions of IIT and GNWT, while substantially challenging key tenets of both theories. For IIT, a lack of sustained synchronization within the posterior cortex contradicts the claim that network connectivity specifies consciousness. GNWT is challenged by the general lack of ignition at stimulus offset and limited representation of certain conscious dimensions in the prefrontal cortex. These challenges extend to other theories of consciousness that share some of the predictions tested here 14–17 . Beyond challenging the theories, we present an alternative approach to advance cognitive neuroscience through principled, theory-driven, collaborative research and highlight the need for a quantitative framework for systematic theory testing and building.

  PublisherDOI

• Does Generative AI Extend the Mind?

  Frontiers in artificial intelligence and applications · 2025-02-04

  book-chapter1st authorCorresponding
  PublisherDOI

• Identifying indicators of consciousness in AI systems

  Trends in Cognitive Sciences · 2025-11-01 · 8 citations

  reviewOpen access
  PublisherOA PDFDOI

• Précis of <i>Reality+</i>

  Philosophy and Phenomenological Research · 2024-11-01

  article1st authorCorresponding
  PublisherDOI

• The Simulation Hypothesis: Metaphysics, Epistemology, Value

  2024-09-17 · 1 citations

  book-chapter1st authorCorresponding

  Abstract I’d like to thank Grace Helton, Terry Horgan, and Christopher Peacocke for their rich commentaries on my book Reality+. As it happens, all three focus on the simulation hypothesis: the hypothesis that we are living in a lifelong computer simulation. Where the simulation hypothesis is concerned, I have three main theses in the book: (i) Metaphysics: If we’re in a simulation, the objects around us are real; (ii) Epistemology: We can’t know we’re not in a simulation; and (iii) Value: We can live a good life in a simulation. The three commentators address all three of these theses. Horgan argues against the metaphysical thesis. Peacocke argues against both the metaphysical and the epistemological thesis. Helton uses epistemological considerations to argue against a version of the value thesis.

  PublisherDOI

• Taking the simulation hypothesis seriously

  Philosophy and Phenomenological Research · 2024-11-01 · 2 citations

  articleOpen access1st authorCorresponding

  Much of Reality+ focuses on the simulation hypothesis: the thesis that we are living in a computer simulation. I argue that we should take the simulation hypothesis seriously, and that we cannot rule it out. I also argue that the simulation hypothesis is not a skeptical hypothesis where most of our beliefs are false. If we are in a perfect simulation, most of our beliefs are true. As a result, the simulation hypothesis does not lead to skepticism, and life in a simulation can be roughly as good as life in a non-simulated world. All three commentators in this symposium focuses on broadly epistemological issues about the simulation hypothesis, often with issues about skepticism as well as issues about value in the background. Peter Godfrey-Smith argues that we should not take the simulation hypothesis seriously. Susan Schneider and Eric Schwitzgebel argue that while the perfect simulation hypothesis may not be a skeptical hypotheses, other versions of the hypothesis may be. As a result, they suggest that life in a simulation may not be as good as life in non-simulated reality, and that the simulation hypothesis may still lead to a degree of skepticism. (See also a recent symposium on Reality+ in Oxford Studies in the Philosophy of Mind, in which Terry Horgan, Christopher Peacocke, and Grace Helton all discuss epistemological about skepticism as well as metaphysical and value-theoretic issues.) Biosims. Godfrey-Smith's first objection to the simulation idea is tied specifically to the biosim scenario in which we are biological systems hooked up to a simulation. Godfrey-Smith suggests that some experiences may be hard to simulate this way, mentioning hot showers and skateboard crashes. Perhaps the thought is that the body plays a special role in these experiences, and it affects the brain in a way that goes beyond standard sensory inputs and outputs. This thought might be fleshed out by asking: to undergo these experiences, does the brain have to be warmed up and jolted around, or is it enough for it to represent warmth and motion? Simulator motives. Godfrey-Smith raises the question of motive: why would simulators create many cosmic simulations? There's an obvious answer: for the same sort of reasons that people create more mundane simulations today: science, engineering, prediction, and more. For example, if it becomes feasible to simulate whole universes at limited expense, then there will be strong scientific motive to do so, for example simulating many different universes with different laws will help understand the space of possible universes. Of course this motive could be defeated by other reasons: reasons of ethics, expense, or disinterest, for example. But the benefits of simulation already take us beyond “Who knows why they would do it?” to a positive reason grounded in reality. Perfect simulations. Godfrey-Smith's most serious objections to the simulation hypothesis concern the simulation of human brains: he suggests that the idea of a perfect brain simulation is problematic, and also that brain simulations will not replicate what is crucial for consciousness. He illustrates the point about perfect simulation by asking: how much fidelity of timing must be present in a perfect duplicate? I suppose a truly perfect simulation would demand perfect fidelity, which would plausibly demand simulating a system down to the level of fundamental physics. Godfrey-Smith doesn't address this sort of physics simulation, but various questions could be raised for it: for example, is it possible to simulate continuous systems with infinite precision, and is it feasible to simulate brains down to the level of fundamental physics? In any case, my arguments about the simulation hypothesis do not depend on simulations that are perfect in this sense. What matters is that simulations be humanlike. This does not requires perfectly simulating any existing human being. It just requires simulating a way that a human being might have been. This can plausibly be done without simulating every detail of a brain to infinite precision. One way to see this is to note that brain mechanisms themselves are subject to biological noise and so cannot rely on infinite precision for humanlike behavior. Substrate-neutrality. Perhaps Godfrey-Smith's largest concern about the simulation hypothesis concerns consciousness. He speculates that “the basis for felt experience” may be specifically biological and may not be replicated in a simulation. The specific view Godfrey-Smith likes gives a key role to the dynamics of electrical oscillations in neurons, involving ions that move across cell membranes. Godfrey-Smith does not deny that these processes could be simulated. Instead he denies that the simulation would have the relevant biological or chemical properties (such as electrical oscillations in neurons) that are required for consciousness: electrical oscillations in neurons, perhaps. Now, the mere fact that electrical oscillations in neurons are the basis for consciousness in humans does not dictate that they are required for consciousness in all systems. A functionalist can reasonably hold that it's the structural dynamics of these oscillations that matter, rather than their realization, and that this structural dynamics could be present in a simulation. Godfrey-Smith suggests that a simulation might represent this dynamics but it would not realize it. But I think this misses the fact that genuine dynamics is present in computer simulation, with parts of a computer system affecting other parts in genuine patterns of oscillation. This dynamics will not just represent but replicate the structural dynamics among neurons, at least at a certain level of abstraction. Around here the real disagreement is which level of abstraction in the dynamics matters for consciousness. I've used neuron replacement “fadng qualia” arguments (changing low-level processes while preserving high-level dynamics) to argued that low-level processes are unlikely to be required for consciousness. Godfrey-Smith is skeptical that this sort of replacement is possible without affecting behavior. Again, a perfect replication of behavior is not required for the argument to work. It suffices to have have humanlike behavior, without huge anomalies such as people reporting that their experiences are fading out. I suspect that neuron-replacement of this sort is possible, and may even become actual in coming decades. At that point we'll have empirical evidence to use in adjudicating these arguments. To step back: the debate between biological views and functionalist views of consciousness, and the corresponding debate about whether simulations can be conscious, is likely to go on. I favor a functionalist view, while allowing that biological views are serious possibilities. Godfrey-Smith favors a biological view, but he calls this “speculative” and presumably allows that functionalist views remain serious possibilities. Once we grant that both these views are serious possibilities, it must also be recognized as a serious possibility that simulations can be conscious, and this line of reasoning (at least in the absence of other sim blockers) cannot be used to dismiss the simulation hypothesis. Boltzmann brains. Boltzmann brains are shortlived humanlike brains that some physical theories predict will form infinitely many times through random processes in an infinite space. These raise obvious skeptical issues: isn't it likely that we are Boltzmann brains, and that most of what we believe is false? In my brief discussion in Reality+, I follow Sean Carroll in holding that the belief that we are Boltzmann brains is cognitively unstable: if it's true, I cannot stably endorse it. If I accept that I'm a Boltzmann brains, I should reject my perception of the external world as illusory and my scientific beliefs as incorrect, which would eliminate the scientific case for believing I'm a Boltzmann brain in the first place. This diagnosis of instability is far from a full response to the complex skeptical worries raised by Boltzmann brains, but undermining the support for the hypothesis is at least a first step in defusing these worries. Godfrey-Smith thinks there is a parallel between Boltzmann brains and the simulation hypothesis. He says that if the Boltzmann brain hypothesis is cognitively unstable, then so is the simulation hypothesis. Taking either of them seriously should lead us to doubt our own reasoning in general, and therefore to doubt the reasoning that led us to these hypotheses. I don't think there is parity between the Boltzmann brain hypothesis and the simulation hypothesis here. Once one accepts simulation realism, as I do, then the simulation hypothesis (at least in its usual forms) should not lead us to reject our perceptions of the external world or our scientific theories, and so it should not lead us to rejecting the case for the simulation hypothesis. In paradigmatic simulations, unlike Boltzmann brains, our reasoning and most of our beliefs are still good guides to reality. Godfrey-Smith says “To seriously suspect you are in a simulation is to suspect that many or all of your memory traces might be cooked up, and your background knowledge is no good at all”. Of course I reject this claim. Once one accepts simulation realism, one can suspect that one is in a simulation without doubting one's memory or background knowledge. Now, perhaps Godfrey-Smith rejects simulation realism (though he says at the start that he won't worry about this issue, and he doesn't give reasons for rejection). Or perhaps he has in mind simulations where only some aspects of the world are simulated and many of our beliefs are false. In Reality+, I argue that even in these simulations, many of our everyday beliefs are fine. So there is still not parity with the self-undermining Boltzmann brain hypothesis. Still, simulations of this sort rise many issues of interest, some of which I discuss in the following sections. In Reality+, I argue that the perfect simulation hypothesis, on which a whole universe is simulated with great fidelity, need not be a skeptical hypothesis. If we are in a perfect simulation, we are not deceived about most everyday matters (or at least, we are not deceived simply because we are in a simulation). Schneider and Schwitzgebel both accept this sort of simulation realism, at least for the purposes of argument. But they both argue that there are other versions of the simulation hypothesis that I cannot exclude, and that should be regarded as skeptical hypotheses: if those hypotheses are true, most of our everyday beliefs are false. Schneider focuses especially on deceptive simulations, devised by superintelligent AIs partly in order to deceive us. Schwitzgebel focuses on local simulations (or small simulations), which just simulate a very small part of the universe such as a single city. They argue that many or most simulations may be deceptive simulations or local simulations, and that if we are in a simulation of this sort, many or most of our everyday beliefs are false. Now, in Reality+, I am mainly concerned to argue against global skepticism, on which we can't rule out Cartesian scenarios in which we are radically deceived about almost everything. Various more local forms of skepticism, turning on scenarios in which we are deceived about some things but not everything, are left on the table. The final chapter of the book focuses on residual skeptical hypotheses from local simulations to Boltzmann brains, and draws the overall verdict “[W]hile there are some important skeptical challenges here, none lead to global skepticism.” Schneider and Schwitzgebel's claims are compatible with the general line in Reality+. They don't argue for global skepticism. Instead, they argue that simulation scenarios should lead us to take more local skeptical hypotheses seriously. I could simply agree with this, but I think that these local skeptical scenarios are not quite as straightforward as Schneider and Schwitzgebel suggest, so I will take a closer look at them. How confident ought we to be that if we inhabit a virtual reality the reality is large enough to be epistemically non-catastrophic – that the world contains more or less all of the things we care about, plus a reasonably deep past, plus a reasonably long future, and billions of people? Call this the Size Question. An optimist about the Size Question holds that we ought to be confident that if we are sims, we don't live in a catastrophically small simulation. The pessimist denies this. Let's call Schwitzgebel's “epistemically non-catastrophic” simulations large simulations, and call the contrast class small simulations. Large simulations needn't simulate the whole universe, but they need to simulate most of the earth and its people, and all of the people and places we care about. In small simulations, much of what we care about isn't simulated and doesn't exist. There's an immediate question about small simulations so defined. For people and places we care about, we typically have of and other such as and about them. In a small simulation, where do these and have to be The obvious is that they are by a simulation of those people and But if all of these are then our small simulation is a large simulation. Now, the small simulation could that the and from some other from and with or for But it's not at all to see how use of these things without simulation could the and that we might have of our and Perhaps there could be a where and are simulated in detail but where are simulated in less and people out are just and But it just how and the simulated will work. a small simulation could a scenario where most of these and don't exist. a simulation might the case where I'm just up and about these with just a that I've a life and have a world out Perhaps it will the I go through or the world. might call these simulations. Perhaps we can't rule these out. But this won't to even a single where I many many and with many For those we'll need many simulated people, have of their and we'll be the scenario In Reality+, I suggest that it's hard to a between simulations and large simulations such as simulations of the whole in part because of all the between people across the Perhaps in we could simulate a but for and this much the simulation only one that the has for at least a but beyond it in the and we have real with The you are in and the and the at the beyond the some If they the they from and they they with of from is all you up in the same your is This most of the key questions Let's I to a to How are my of will a of and my it. The same goes for most places on at least in the of things all the and of and even and the will very simulation. What by or with all the will presumably need to be simulated. All this suggests that simulation of the world the will be Let's a I a long with Eric I in and he in How this memory on Schwitzgebel's where only is I in so my of the the Eric simulation has out of (at least the he in But some sort of will be to of the think that to a Eric simulation would be A would be But if a Eric simulation is then a simulation will also be for most of the people with I and for most of the people other have will move us a simulation of billions of these simulated people the will of with many and those will also up on and the which will simulations in At one point Schwitzgebel that standard people as that doesn't as well we have specific experiences and to Of course simulators might such as large and their to in for Schwitzgebel has a simulation of that some people from the This used only small in special but perhaps a of from the might be But these systems are themselves simulations of a sort if they simulate behavior more than they simulate so they don't the need for simulation. It might be argue that these systems will not be conscious, or will other of genuine so that at least our beliefs about other are these systems are Grace Helton a line this that the simulation hypothesis to in a recent symposium on Reality+, and I far from that human behavior without consciousness is possible, but even if it this more on knowledge of other than on the physical world. mind where simulators our They can simply our our our or our so we don't such as behavior. may help simulations with the use of and simulations, a and simulation the sort Schwitzgebel will still if simulation of the world. In of mind to in an to I'm simply living out a of my life in perhaps with this no simulation is but my and will have to be by the simulators to that I do not Perhaps I can't rule out of this sort, but them as very different from the standard simulation hypothesis: life in a is not life in a simulation. In any case, I'm to skepticism as from simulation skepticism. raises skeptical issues of its with our reasoning to us doubt that plus three is As I note in Reality+, this sort of skepticism is to But mind can be with or without simulation, as see here. can or skeptical scenarios mind whether or not we have simulation. At simulation an especially to mind as may be to than brains. But it the mind rather than the simulation for the skeptical worries here. So understand the simulation hypothesis as involving scenarios without Once scenarios are I am to my in the simulation hypothesis mainly between large simulations much of the at and simulations simulating in my and about for a How do I that that simulations will be much and to up than large simulations. the other simulations of with are and it's why simulators will be to create them. The scientific and for large simulations are much Perhaps there will be many large simulations than simulations, but not enough to a in the large simulations. the point about A simulation will typically just one a at while a large simulation will billions of As a result, even if there are a times as many simulations as large simulations, there will be a times as many people in large simulations as in simulations. If what we we should only of our in the simulation hypothesis to simulations, and the to large simulations. Schwitzgebel to this point about the by but those have already at the first in allowing that there will be far more simulations than large simulations. Of course it is hard to what the is for but as long as it's a or so, the for in large simulations will lead to most of our I'm to have a in the large simulation hypothesis, on the simulation hypothesis in But this is quite and to the at least a of one in for the small simulation hypothesis, which Schwitzgebel as So I to have the corresponding local forms of skepticism. Still, to endorse Schwitzgebel's that we should that we in a simulation. Schneider an hypothesis, suggests that we might be in an illusory simulation up by an system that is to deceive us. says the simulation might have and Of course I think that are possible in simulations just as they are possible in reality. So I think partly illusory with are At the same I'm to deny that illusory where we and believe is are not obvious whether Schneider is partly or illusory If it's the then I agree with the So to things arguments as arguments for illusory and argue that they don't Schneider says a of things about illusory to support the that they are Perhaps the most is that the to deceive us. the example of the says world that is can to be as we believe it to be because it is an Now, just because an to deceive us does not that it The case this out. The may have to deceive humans by them the that exist. But if I am they no such humans in the have beliefs about the world them. They are not at least about everyday matters about whether (though they may be deceived about matters such as whether are and so view is that if the think they are us about these they are a Of the may be enough to this But the point out that their mere to deceive does not that the world is If the world is it must be for a more One reason for global is by a in the may be a simulation to deceive us about the of the world the simulation. For example, suggests that a deceptive simulation may have which I as that its laws are very different from those in the world in which the simulation place. But I don't think that laws in this that people in the simulation are They may be of the laws of the but they need not have beliefs about them. my view, their world is the world of the simulation, its laws are the and they may well have beliefs about their world. Of course if they that these laws the laws of the they would be but that belief does not to be to being in this simulation. as beliefs are about the world I argue they mainly and not about the then the fact that the world is unlike the world does not that the simulation is Schneider also suggests that the simulation is likely to be full of we can between global which is roughly us about the and local which is roughly us about many matters of fact without us across the If I am a deceptive simulation is much more likely to local than global At least, I don't have a of how simulators could deceive us while I have a of how simulators could deceive us This would just the same sort of mechanisms of that we in the world. Of course local is still and if every belief is subject to hypotheses in which it skepticism still But if the is local rather than then at least this skepticism will have Schneider and Schwitzgebel also both raise issues about the of the that the is likely to have and that the simulation is then reality. Schneider argues that if our world is the is very likely an This part quite to argues that this system is likely to many of the of existing it may be with and other of The reasoning here is less obvious to An system that can support a universe simulation will be far more than existing and one can reasonably that part of this will be on the of and the other Schneider has on do to be in a world with and All that suggests that if we are by an it has not done a good of and from the world it This to Schwitzgebel's worries about the in our world. He argues that if our world is then the is for and from the to the If so, the cannot be a good One could a point about and a would create such a world cannot be a good Of course this is a on the of for a would not all this so if there is a they are not The simulation hypothesis just the issues from to more is a of on of a For example, perhaps this world has enough good that the good the and it other good is on than not it. these can be to a and are about as or here as in the In Reality+, I argue that the simulation hypothesis is roughly to a of the hypothesis world is and the hypothesis the physical Schwitzgebel's worries about are for the hypothesis and not for the simulation hypothesis So he says not in a the more general point here is not a This point both and simulations. for I suspect that if our world is our is at But I'm not this we should not a already there is much in our or no I'm not that an would things much than we already them to be. Schwitzgebel suggests an would the world an and but I'm not it would be for us. Perhaps this would reasonably lead us to worry more about things that might world for but it might also raise our for some good things for The same goes if in a simulation.

  PublisherOA PDFDOI

• Does Thought Require Sensory Grounding? From Pure Thinkers to Large Language Models

  arXiv (Cornell University) · 2024-08-18 · 6 citations

  preprintOpen access1st authorCorresponding

  Does the capacity to think require the capacity to sense? A lively debate on this topic runs throughout the history of philosophy and now animates discussions of artificial intelligence. I argue that in principle, there can be pure thinkers: thinkers that lack the capacity to sense altogether. I also argue for significant limitations in just what sort of thought is possible in the absence of the capacity to sense. Regarding AI, I do not argue directly that large language models can think or understand, but I rebut one important argument (the argument from sensory grounding) that they cannot. I also use recent results regarding language models to address the question of whether or how sensory grounding enhances cognitive capacities.

  PublisherOA PDFDOI

Frequent coauthors

• Stanislas Dehaene

  NeuroSpine Institute

  8 shared

• Gabriel Kreiman

  Harvard University Press

  7 shared

• Lucía Melloni

  Max Planck Institute for Empirical Aesthetics

  6 shared

• Ling Liu

  Peking University

  6 shared

• Andy Clark

  6 shared

• Theofanis I. Panagiotaropoulos

  Commissariat à l'Énergie Atomique et aux Énergies Alternatives

  5 shared

• David Mazumder

  Harvard University

  5 shared

• Marcelo Armendáriz

  Boston Children's Hospital

  4 shared