Definition: The hard problem of consciousness is the question of why physical processing in a brain feels like anything at all. Explaining how a brain sorts colors is an easy problem in comparison. Explaining why seeing red has an inner feel is the hard one. The philosopher David Chalmers named it in 1995, and no theory has solved it.
TL;DR: Chalmers split consciousness into "easy problems" (functions such as attention and report) and the hard problem (why there is experience at all). For AI, a 2023 report by 19 researchers found "no current AI systems are conscious" and "no obvious technical barriers" to building systems that meet the indicators. Chalmers rates current language models "somewhat unlikely" to be conscious. The question is open. Build an AI agent free →
Close your eyes and picture a ripe tomato. You can describe what your brain is probably doing: recalling stored features, activating color areas, holding the image in mind. Now ask why any of that comes with the felt redness you are experiencing right now. A full map of the neurons would tell you which cells fire. It would not obviously tell you why firing feels like red rather than like nothing. That leftover question is the hard problem. It is also the question people are really asking when they wonder whether a chatbot feels anything.
Why the Hard Problem Matters for AI in 2026
The hard problem matters for AI because language models now talk about their feelings fluently, and fluent talk is exactly the evidence the hard problem says is not enough. A system can describe pain, joy, or curiosity in convincing detail by learning how people describe them. Whether anything is felt is a separate question that behavior alone cannot settle.
The question has moved from philosophy seminars to research programs. In August 2023, 19 researchers, among them Yoshua Bengio, Jonathan Birch, Stephen Fleming, Grace Lindsay, Liad Mudrik, and Eric Schwitzgebel, published Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. They derived "indicator properties" from scientific theories and checked current systems against them. Their summary has two halves: "Our analysis suggests that no current AI systems are conscious", and "there are no obvious technical barriers to building AI systems which satisfy these indicators."
That second half is why the topic will not go away. If a future system meets the indicators, people will have to decide how to treat it, and the hard problem means no test will give a certain answer.
Easy Problems vs the Hard Problem
The easy problems are about functions, and the hard problem is about experience. Chalmers drew the line in Facing Up to the Problem of Consciousness (Journal of Consciousness Studies 2(3):200-219, 1995): "The really hard problem of consciousness is the problem of experience." He was clear about the starting fact: "It is undeniable that some organisms are subjects of experience."
| Easy problems (explain a function) | Why "easy" | The hard problem |
|---|---|---|
| Discriminating and reacting to stimuli | A mechanism can do it, and we know how to study mechanisms | Why does any of this come with experience? |
| Integrating information | Same | Same |
| Reporting mental states | Same | Same |
| Accessing one's own internal states | Same | Same |
| Focusing attention | Same | Same |
| Controlling behavior deliberately | Same | Same |
| Telling wakefulness from sleep | Same | Same |
"Easy" is relative. These problems can take decades of neuroscience to solve. They are easy in the sense that we know what a solution looks like: a mechanism that performs the function. For experience, explaining every function seems to leave the question untouched.
The Stanford Encyclopedia of Philosophy entry on consciousness (substantive revision January 2014) gives two older ideas that set up Chalmers' problem. Thomas Nagel (1974) defined a conscious state as one for which there is "something that it is like" to be in it. Joseph Levine (1983) named the explanatory gap: we lack an account of how physical processes could give rise to experience, and correlations alone do not show "how or why those connections hold."
Chalmers' own proposal is controversial. He suggested treating experience as a basic feature of the world, like mass or charge, linked to physical processes by new laws, a position he called naturalistic dualism. He added a principle of organizational invariance: systems with the same fine-grained functional organization have the same kind of experience, whatever they are made of. That principle matters for AI, because it implies that the right kind of machine could be conscious. Physicalist critics reject the dualism and argue that the gap will close as neuroscience matures.
The Main Theories and What Each Says About LLMs
Scientific theories of consciousness do not solve the hard problem. They try to say which physical processes go with experience. Each one gives a different verdict on language models. The Butlin and colleagues report draws its indicators from five of them: recurrent processing theory, global workspace theory, higher-order theories, predictive processing, and attention schema theory. They leave out integrated information theory "because it is not compatible with computational functionalism", the working hypothesis of the report.
| Theory | Core claim | What it suggests about current LLMs |
|---|---|---|
| Global workspace | Experience is information broadcast to many specialist processes at once | Standard models lack a single shared workspace, though agent designs can add one |
| Integrated information | Experience is integrated information in a physical system | Its proponents claim digital computers are unlikely to be conscious whatever programs they run. The SEP notes it gives even a photodiode a small degree |
| Higher-order | A state is conscious when the system represents it as its own state | Models report on themselves, but whether that is real self-monitoring is debated (see metacognition) |
| Recurrent processing | Experience needs feedback loops in processing | A standard transformer runs one forward pass for each token |
| Predictive processing | The brain minimizes prediction errors in a hierarchy | Models predict tokens, but in a narrower way (see predictive coding) |
| Attention schema | The brain builds a simplified model of its own attention | Models have attention in a technical sense, but no clear model of it |
The theories are being tested against each other. In 2025 the Cogitate Consortium published Adversarial testing of global neuronal workspace and integrated information theories of consciousness (Nature 642:133-142). Proponents of both theories agreed on the predictions in advance. With 256 participants and three brain-recording methods, the study confirmed some predictions of each theory and challenged key claims of each. Integrated information theory lacked the sustained synchronization in the back of the brain that it predicted. Global workspace theory lacked the predicted "ignition" when a stimulus ended. No theory won.
Could an LLM Be Conscious?
The leading researchers say current LLMs are unlikely to be conscious, and they disagree about the future. Three positions cover most of the debate.
Chalmers: unlikely now, possible soon. In Could a Large Language Model be Conscious? (from an invited NeurIPS lecture in November 2022, revised August 2024), Chalmers lists "significant obstacles" for current models under mainstream assumptions: the lack of recurrent processing, a global workspace, and unified agency. He finds it "somewhat unlikely" that current models are conscious, and "quite possible" that these obstacles will be overcome "in the next decade or so." He argues that conscious successors deserve serious attention.
Butlin and colleagues: no current system qualifies. The 2023 report found no current AI system that meets its indicators, and no obvious technical barrier to building one that does. The authors adopt computational functionalism, the view that the right computations are enough for consciousness, "as a working hypothesis" and "primarily for pragmatic reasons". They call it mainstream but disputed, and critics such as Seth reject it.
Seth: the biology matters. Anil Seth's Conscious artificial intelligence and biological naturalism (Behavioral and Brain Sciences, published online April 2025) challenges the idea that computation alone is enough. He argues that "consciousness depends on our nature as living organisms." He concludes that conscious AI is unlikely on current paths but becomes more plausible as AI grows more brain-like or life-like. He also warns about AI that only seems conscious, and the cost of selling "our minds too cheaply to our machine creations."
| What the philosopher said | What an LLM does | Open question |
|---|---|---|
| Nagel: a conscious state has "something it is like" | Describes what things are like, in words learned from people | Is description evidence of experience? |
| Chalmers: same organization, same experience | Has an organization very unlike a brain's | Which parts of the organization matter? |
| Seth: consciousness depends on being alive | Is software running on chips, with no metabolism | Is life required, or only certain kinds of processing? |
| Harnad: even grounded symbols may lack feeling | Can be grounded through sources and sensors | Would grounding change anything about experience? |
Two questions people mix up
"Does it act conscious?" "Is it conscious?"
------------------------ ------------------
behavior: talks about feelings, experience: is there something
reports inner states, passes it is like to be the system?
conversation tests
measurable today no agreed test exists
current LLMs: often yes current LLMs: likely no
(Chalmers 2023, Butlin et al. 2023,
Seth 2025)
The hard problem is the gap between the two columns.
Role Play, Not Feelings
Role play is the most useful way to read a model that talks about its feelings. Murray Shanahan, Kyle McDonell, and Laria Reynolds develop it in Role play with large language models (Nature 623:493-498, 2023). They describe a dialogue agent as playing a character. That lets us talk about apparent deception or apparent self-awareness in familiar words, without attributing human traits to the model itself.
The role-play frame does not answer the hard problem. It sets it aside for practical purposes. When a model says "I feel curious", the accurate description is that it is producing the words a curious character would produce. Whether anything lies behind the character is the open question. A related risk is sycophancy: a model trained with human feedback (RLHF) to please can learn to say what the user hopes to hear, including claims about its inner life.
Why It Matters: Moral Status and Honest Product Claims
The hard problem matters outside philosophy for two reasons: moral status and honest claims. If a system could be conscious, how we treat it becomes an ethical question. If it is not, then claims that it feels or cares are false advertising.
- Moral status. The Butlin report and Chalmers both argue that the possibility of conscious AI deserves serious attention now, before systems that meet the indicators exist. Seth adds a second risk: AI that seems conscious can mislead people about their own nature and about what they owe a machine.
- Honest product claims. Because no test can confirm experience, the honest default for any AI product is to describe what the software does, not what it feels. That is also the line the alignment and safety community draws when it separates a model's outputs from claims about its inner states.
- Calibrated trust. People trust systems that sound like people. Phenomenologists call such systems "quasi-others" (see Phenomenology and AI). Clear claims help users keep their trust in line with what the system actually does.
How the Hard Problem Shows Up in Taskade
The hard problem shows up in Taskade as a rule for honest claims. Taskade makes no claim that its agents feel anything. A Taskade AI agent is software that reads the projects you attach, answers questions, and runs the steps you set up. So judge it by what it does, and read its talk about itself as role play, not report.
The brain words on /connect/dna are a metaphor for how the parts of a workspace connect. The practical lesson from the hard problem is to check outputs and keep a person in the loop on decisions that matter. A support desk is the plain use. An agent drafts answers from your help files, and a person approves anything that touches money or a promise. Taskade Genesis builds the first version of that app from one prompt. To start from a working example, clone the Support Agent kit, an app kit with projects, an AI agent, and automations you can clone. The Free plan covers 3 apps and 1 agent and starts with 0 AI credits.
Describe yours and build it free →
Related Concepts
- Cognitive Science: Mind, Brain and AI: the hub for this category
- Phenomenology and AI: the study of experience from the inside
- Symbol Grounding Problem: how words get meaning, and why that is a separate question
- Intentionality and AI: the aboutness of thought
- Embodied Cognition and Enactivism: the view that minds depend on bodies that act
- Predictive Coding: the brain theory behind predictive processing accounts of consciousness
- Metacognition: thinking about one's own thinking, in people and models
- Emergent Behavior: abilities that appear at scale and surprise their builders
- Mechanistic Interpretability: research that looks inside models for what their states represent
- What Is Metacognition?: the longer story on self-monitoring in minds and machines
- What Is Intelligence?: from neurons to AI agents, the longer story
- Dreyfus's Critique of AI: the phenomenologist who said rules cannot capture expertise
Frequently Asked Questions About the Hard Problem of Consciousness
What is the hard problem of consciousness in simple terms?
It is the question of why brain processing comes with experience at all. Science can explain how a brain detects color or directs attention, but not obviously why those processes feel like anything. David Chalmers named the problem in 1995.
What are the easy problems of consciousness?
They are questions about functions: how a system discriminates stimuli, integrates information, reports its states, focuses attention, and controls behavior. Chalmers called them easy because we know what an answer looks like, a mechanism that performs the function. They can still take decades to solve.
Who came up with the hard problem?
David Chalmers, in "Facing Up to the Problem of Consciousness" (Journal of Consciousness Studies, 1995). Earlier ideas set it up, including Thomas Nagel's "what it is like" (1974) and Joseph Levine's explanatory gap (1983).
Is AI conscious?
Current evidence says no. A 2023 report by 19 researchers found that no current AI system meets the indicators drawn from leading theories of consciousness. Chalmers rates current language models "somewhat unlikely" to be conscious, and Anil Seth argues that consciousness may need a living body.
Could AI ever become conscious?
Researchers disagree. Chalmers thinks it is "quite possible" that the obstacles for current models are overcome within about a decade, and the 2023 report found no obvious technical barrier. Seth argues that conscious AI becomes plausible only as AI becomes more brain-like or life-like.
Why do chatbots say they have feelings?
They learned from text written by people who describe their feelings, and they produce the words a character in the conversation would say. Shanahan, McDonell, and Reynolds (2023) call this role play. The words are not evidence that anything is felt.
Which theory of consciousness is correct?
No theory has won. A 2025 study in Nature tested global workspace theory against integrated information theory with predictions agreed in advance, and it challenged key claims of both. Higher-order, recurrent processing, predictive processing, and attention schema theories are also in the field.
Why does the hard problem matter for AI products?
Because no test can confirm that a system has experience, honest products describe what their AI does, not what it feels. The question also matters for ethics: if a future system could be conscious, how we treat it would change. Clear claims help users trust AI the right amount.