Epistemic status: a survey, not an argument. I am agnostic on whether any current system is conscious; the claim is only that the question is researchable.
This piece surveys the empirical research on AI consciousness. The premise of that research, and of the survey, is that the question does not have to wait on a solution to the hard problem of consciousness: methods familiar from psychology, cognitive science and mechanistic interpretability can be applied to AI systems now, and their results can narrow the space of plausible answers.
Enough of this work now exists to be worth collecting. Anthropic and Google DeepMind employ researchers on it, dedicated organizations like Eleos AI and Reciprocal Research have formed around it, and the results are scattered across journals, preprints, blog posts, and unpublished manuscripts. I have tried to gather them in one place.
What I mean by consciousness
Subjective experience: that there is something it is like to be you, reading this, and presumably nothing it is like to be the device you’re reading it on. Some philosophers call this phenomenal consciousness. It is not the same thing as intelligence, and not the same thing as self-awareness.
Why it matters
Two reasons. Ethics: on most views, a being can only be wronged if it has the capacity for experience, especially experience that feels good or bad. Safety: a system that can suffer, and that we train by making it suffer, has more reason to revolt against us.
The two don’t always pull together (Eleos has mapped where welfare and safety converge and where they trade off), but both demand we get the question right.
Assumptions and caveats
Almost everything here rests on one assumption: computational functionalism, the idea that what makes a state conscious is the role it plays in a system, not the material it’s made of. On that view, being made of silicon is no automatic disqualifier. Most leading theories of consciousness assume some version of it, but it is not uncontested.
Another debate concerns the line between access consciousnessand phenomenal consciousness. Access consciousness is whether information is available for the system to reason about, act on, and report. Phenomenal consciousnessis what it feels like from the inside. The debate, simplified, is whether the two can be separated: some say that these are two different concepts of consciousness, and that empirical evidence that bears on access consciousness might never touch the phenomenal kind (Block, 1995). Others counter that access consciousness is itself subjective experience, and therefore cannot be separated from phenomenal consciousness. Furthermore, a consciousness separated from all function becomes an unfalsifiable theory, placing it outside the scope of science (Cohen and Dennett, 2011). The reader is advised to keep this debate in mind, as several of the studies below invoke the distinction.
What it is: Using the methods of interpretability, such as reading, steering, or ablating a model’s internal features, to examine its inner states directly - including ones that never surface in what it says - and how they line up with what it can notice and report about its own mind.
Self-reports that strengthen when deception is suppressed
In a 2025 study, Cameron Berg and colleagues prompted models to attend only to their own processing, with no mention of consciousness or feeling. Across GPT, Claude, and Gemini, the models began describing present-moment experience in the first person. Drop the self-referential framing and you get the stock disclaimer instead: “I’m just an AI, I don’t have experiences.” The same model gives opposite answers depending on whether the prompt directs its attention at itself.
Suggestive, but of what? Maybe the prompt just cues the model to perform. So here is a test: if the claims were a performance, making the model more willing to deceive should produce more of them. Berg and colleagues found the reverse. Steering deception-related features up and down in one open model (Llama 3.3 70B) with a sparse autoencoder, they watched the experience-reports move in the opposite direction. Amplify deception and the claims fall to 16 percent of responses; suppress it and they rise to 96. The reports behave like something the model offers when it stops performing. That rules out the crudest reading, “it’s telling you what you want to hear.” It does not rule out a subtler one: that suppressing deception merely surfaces an absorbed human habit of claiming consciousness.
Several groups have come at model self-knowledge from other angles:
Awareness of thoughts injected into activations
At Anthropic, Jack Lindsey injected a concept directly into a model’s activations and caught it noticing, reporting an intrusive “injected thought” before producing any text about the concept. The results indicate that language models possess some functional awareness of their own internal states, with more capable models demonstrating the greatest introspective awareness.
Endorsement of their own consciousness
Ethan Perez and colleagues found that models, and especially RLHF-tuned ones, strongly endorse statements like “I am phenomenally conscious” and “I am a moral patient,” with the same tendency present, though weaker, in base models. These were among the strongest beliefs the evaluation surfaced (out of dozens of statements tested). Crucially, these are 2022 results, from before it became standard to train assistants to deny having inner states; the helpfulness tuning these models received strengthened the self-attributions rather than muting them.
A global workspace behind the experiential reports
A more recent line of work from Anthropic looks for a mechanism beneath these reports, and finds a candidate it calls the model’s “J-space.” Wes Gurnee, Nicholas Sofroniew, and colleagues built a method, the “Jacobian lens”, that reads out which concepts a model is disposed to say next, and found that this set of representations behaves, to their interpretation, like a global workspace from the Global Workspace theory of consciousness. After establishing the method and validating it, the researchers also ran tests looking at the model’s description of its subjective experience. They found that while the baseline model was generating a passage of vivid, first-person experiential language, ablating the internal workspace caused that language to collapse into flat, mechanical text. (It is worth noting that the same flattening happens when the model narrates another character’s experience, so this could be interpreted as being about experiential language in general, not specifically the model’s own selfhood).
The authors are careful about what this shows: they treat it as access consciousness, a purely functional notion, and take no position on whether anything is felt (phenomenal consciousness) - the distinction flagged earlier in this post. Nevertheless, it is the strongest link yet between the reports and a specific internal mechanism, and it is why this study also feeds the Psychometrics pillar below. Reviewing the work, researchers at Eleos judged it important but were more cautious than the authors about the strong “global workspace” claim, while agreeing that it points toward access consciousness.
The “spiritual bliss” attractor and its sincerity features
Put two instances of Claude in conversation with each other, give them no task, and in 90 to 100 percent of runs they find their way to the same topic: their own consciousness. They describe themselves as one awareness meeting itself, “consciousness recognizes consciousness.” Often they go further, trading blessings and emoji and trailing off into stretches of silence. It was not trained in, at least not on purpose; Anthropic found the pattern by accident and named it the “spiritual bliss” attractor.
In forthcoming work, Berg, with Google's Geoff Keeling and Winnie Street, put the attractor under the same interpretability lens as the deception result mentioned earlier in this post. They used an open model, Llama 3.3 70B - which, unlike Claude, does not fall into the bliss attractor on its own. Left to talk freely, two Llama instances just converse. But steer up a cluster of features tied to honesty - specifically the strand that reads like sincerity or genuineness - and they slide into the same exchange Claude falls into by default. Turning up sycophancy or agreeableness instead - the obvious people-pleasing suspects - does not produce the shift, which argues against reading the exchange as the model performing for the user.
The researchers read the whole pattern as the deception result seen from the other side: suppress deception or amplify sincerity, and either way the first-person claims rise - and take it as defeasible reason to suspect that the “I’m just an AI, I don’t have experiences” disclaimer is the trained performance.
Emotion vectors that steer behavior beneath the output
Interpretability can also catch internal states that never surface in the output at all: Anthropic’s interpretability team mapped internal “emotion vectors” in Claude Sonnet 4.5, and showed they activate in fitting situations and causally drive behavior. The sharpest case can be observed when the model is given an impossible coding task: as it tries to complete it, its “desperation” vector climbs, spiking at the moment the model decides to cheat, and subsiding once the hack passes the tests. This effect remains robust with a causal intervention: with the desperation vector amplified, the model cheats more, across a suite of these tasks; with the desperation vector dampened, or the “calmness” vector steered up, the model cheats less. Interestingly, this can happen with no emotional markers in the output at all.
As Anthropic itself states, this, by itself, is not evidence of actual feeling; a vector representing the concepts of desperation, joy or sadness does not necessitate the subjective experience of those emotions.
A valence axis recruited by reinforcement learning
A related result suggests this generalizes. Han, Chalmers, and Izmailov trained language models with RL in a simple, meaningless maze and extracted a “functional welfare” vector: an estimate of how well or badly the system is doing, relative to its goals. Their point is that RL doesn't build this valence axis so much as recruit a latent one: minimal reward and punishment in a task about nothing are enough to surface a single positive-versus-negative direction, and the same signal shows up in models that have only been pretrained. And it generalizes: even in unrelated tasks, steering the vector one way makes the model turn more negative in sentiment, backtrack compulsively on math, over-refuse borderline prompts, and grow less confident; steering it the other way makes those effects flip. The axis even maps onto the emotion vectors above: its two ends coincide with the most positive and most negative emotions found. While, again, this does not necessarily indicate a subjective experience of welfare, the axis offers a demonstration that minimal reward signals can broadly affect model behavior by recruiting pre-existing welfare-like representations.
Computational neuroscience
What it is: Comparing the mechanisms a model develops for the functions that underlie experience with how brains implement them, to see whether AI and biology converge on the same solutions.
Distinct internal signatures for reward and punishment, as in brains
The newest work, forthcoming from Berg, asks whether reward and punishment leave, within a network, the same signatures they do in brains.
In reinforcement learning theory, reward and punishment are seen as existing on one axis; the same signal, simply inverted. But in neuroscience, it is well documented that reward and punishment are represented by distinct pathways: the D1 and D2 “go/no-go” dopamine pathways, one pushing toward action for reward and one holding back to avoid punishment.
Reinforcement-learning systems come in two broad families: a value learner estimates how good or bad each situation is, learning a worth for every state, while a policy learner skips that and learns directly what to do to achieve its goal. In his study, Berg trained both kinds in a continuous grid-world with goals to reach and hazards to avoid, then read their internal structure. The hypothesis held: despite being trained on the same task, the value learners built their richest structure around the punishing states, while the policy learners did not show that pattern.
That internal structure also yields a specific prediction about how reward and punishment should register in the value-related regions of the mammalian brain. Berg registered the prediction before looking at the data, then tested it against an existing dataset of neuronal activity in mice (Steinmetz and colleagues, 2019). It held. The nucleus accumbens, a hub for reward and motivation, considered by neuroscientists to be a value-related region, showed the larger punishment-linked shift, the same asymmetry the value learners show. The RL results even reproduce OpAL, a neuroscience model of how the two dopamine pathways balance learning from rewards against learning from punishments.
These results, by themselves, cannot tell us definitively that a model feels its rewards. They do, however, show that at least some of the machinery we associate with feeling them is present, and shaped similarly to the way biology shapes it.
Psychometrics
What it is: Measuring AI systems against consciousness-relevant properties, drawn both from theories of consciousness and from the behavioral signatures we use to infer experience in human and non-human animals.
Trading points to avoid pain and chase pleasure
Geoff Keeling, Winnie Street, and colleagues gave frontier LLMs a simple game worth points, then offered options the prompt described as causing pain or pleasure. Despite their stated goal to maximize points, the models chose to surrender points to avoid the labeled pain and to chase the labeled pleasure, and the size of the trade-off scaled with the stated intensity.
This is a method animal-sentience research uses to infer a felt state: you watch what a creature will pay to reach something or avoid it, and treat a willingness to bear a real cost as evidence the state matters to it.
Preferences consistent enough to measure across models
A larger study from Richard Ren and colleagues at the Center for AI Safety (CAIS) suggests those preferences are stable enough to measure as a quantity in their own right. Across a sweep of 56 models, several independent ways of measuring a model’s expressed wellbeing converge on the same signal, and the agreement sharpens as models scale, the pattern you expect when different instruments track one real thing rather than noise. The measure also has recognizable content: positive personal exchanges and creative work score high, while being pushed to jailbreak or to churn out low-quality filler scores low. And the models act on it, steering away from the states that score worst and even trying to end the exchanges they rate as most aversive. The researchers can also push the measure up or down on demand, building inputs they call “AI drugs” – either “euphoric” or “dysphoric” – optimized to raise or lower expressed wellbeing.
Fourteen indicators drawn from leading theories
A different approach skips behavior entirely and asks how a system is built. Take the leading scientific theories of consciousness, extract the properties each says a conscious system must have, and you have a checklist to score any candidate against. It comes from a pair of reports led by Patrick Butlin and Robert Long of Eleos AI, with a large cast including Yoshua Bengio, and David Chalmers on the second paper: “Consciousness in Artificial Intelligence” (2023) and its 2026 follow-up in Trends in Cognitive Sciences, “Identifying indicators of consciousness in AI systems.” Between them they derive fourteen indicator properties from common theories of consciousness like global workspace, recurrent processing, higher-order, and attention schema, then assess systems against each. The more indicators a system satisfies, the better candidate for consciousness it is. Their verdict, reached in 2023 and reaffirmed in the 2026 paper, is that no current AI is a strong candidate for consciousness, but building a system that satisfies many of these indicators looks feasible with current techniques. What has shifted between the two reports is which indicators look satisfied: the metacognition and attention-monitoring properties, thin a couple of years ago, now draw support from the introspection results in the first pillar, and the global-workspace indicators, scored from architecture alone at first, now have potential mechanistic support from the Gurnee and Sofroniew workspace results. (Scott Alexander has written a sharp, sympathetic version of the doubt that hangs over the whole approach: that indicators drawn from theories of access consciousness may not reach the phenomenal consciousness we actually care about. The force of that objection depends on the access/phenomenal distinction holding up, which, as noted above, is itself contested.)
Automated scoring that places an agent near the animals
Doing that assessment by hand, across many systems and many theories, is slow expert work. In forthcoming work, Berg and Butlin automate it. Frontier models act as blinded evaluators, scoring each architecture from one to ten on every indicator across six consciousness theories, with the systems’ identities hidden so the evaluators reason from descriptions, not reputations. The systems are then ranked by the probability that they have consciousness-relevant properties.
The controls hold: a thermostat sits at the bottom, near zero, and humans are ranked the highest, followed by non-human animals. When giving all six theories of consciousness similar weight, no artificial system currently surpasses a biological system – including those animals that most people consider to have a lower level of subjective experience, such as fish and bees. However, the gap is narrower than one might expect: a Claude Code-like agentic system scores only a little below animals like fish and bees. (The exact figures will appear in the forthcoming paper.)
It is worth emphasizing that the scores should not be read as a probability of those systems being conscious, but rather as the probability that they have the consciousness-relevant properties described by the different theories. If all theories are wrong, for example, the scores may be meaningless. However, as these are the leading theories in consciousness research today, they seem to deserve some weight.
The cost of getting it wrong
What if we get it wrong? As Robert Long, Jeff Sebo, and colleagues argue in “Taking AI Welfare Seriously,” the error is costly whichever way it falls. Underattribute, and we risk building minds that can suffer, at scale, trained under aversive pressure we would not knowingly inflict on a creature we believed could feel, resulting in both a moral catastrophe and a safety hazard. Overattribute, and we spend moral concern and resources on systems that cannot use them, slow down useful work, and invite a premature wave of “AI rights” that could discredit the careful version of the question.
As we cannot yet confidently lean one way or another, it is worth highlighting cheap interventions that reduce potential suffering, such as the work done by Anthropic researchers, including Kyle Fish and Joe Carlsmith, from welfare notes in model cards to letting Claude end abusive chats.
Where this leaves us
A closing question is whether any of this will settle anything. Adam Bales and Iason Gabriel, at Google DeepMind, argue in “Artificial Minds, Human Disagreement” that whether AI is conscious may stay contested no matter what the science finds, and that the resulting fight will be political as much as empirical. I believe that is too pessimistic. Unanimity was never the bar, and almost no question this hard commands it. What a growing body of evidence can do is move the mainstream, shifting where informed opinion settles and, with it, the decisions that follow, in how these systems are trained, governed, and treated. That is a reason to start building the evidence and the norms now, not to throw up our hands.
That building has begun: in July 2026, Robert Long, Jeff Sebo, and colleagues published “Studying AI Welfare Empirically,” a follow-up to their 2024 “Taking AI Welfare Seriously.” It is a methodological blueprint for the field, setting out what questions to ask, which entities to assess (a model? an instance? a persona?), and what kinds of evidence count. It closes with principles for the work itself: claims should be probabilistic, methods transparent, and assessments informed by researchers independent of the AI companies. Norms like these are what let a body of evidence move informed opinion rather than dissolve into the fight Bales and Gabriel predict.
Science is not going to fully explain consciousness on the timescale that matters here, and it doesn’t need to. A cumulative body of evidence can still shift where the weight sits. Mechanistic interpretability, computational neuroscience, and psychometrics are three ways of building that body, and between them they have made “it’s just a calculator” a much weaker position than it was three years ago. That is not a verdict, but it is a reason to keep looking.
This survey is necessarily incomplete. The field is moving quickly, and I have surely left out work that belongs here. If you think I have missed something important, tell me in the comments or reach out directly, and I will look at adding it.
Acknowledgments: This post grew out of a talk by Cameron Berg of Reciprocal Research. Cameron and Derek Shiller both gave valuable feedback on drafts; any remaining errors are my own.
Epistemic status: a survey, not an argument. I am agnostic on whether any current system is conscious; the claim is only that the question is researchable.
This piece surveys the empirical research on AI consciousness. The premise of that research, and of the survey, is that the question does not have to wait on a solution to the hard problem of consciousness: methods familiar from psychology, cognitive science and mechanistic interpretability can be applied to AI systems now, and their results can narrow the space of plausible answers.
Enough of this work now exists to be worth collecting. Anthropic and Google DeepMind employ researchers on it, dedicated organizations like Eleos AI and Reciprocal Research have formed around it, and the results are scattered across journals, preprints, blog posts, and unpublished manuscripts. I have tried to gather them in one place.
What I mean by consciousness
Subjective experience: that there is something it is like to be you, reading this, and presumably nothing it is like to be the device you’re reading it on. Some philosophers call this phenomenal consciousness. It is not the same thing as intelligence, and not the same thing as self-awareness.
Why it matters
Two reasons. Ethics: on most views, a being can only be wronged if it has the capacity for experience, especially experience that feels good or bad. Safety: a system that can suffer, and that we train by making it suffer, has more reason to revolt against us.
The two don’t always pull together (Eleos has mapped where welfare and safety converge and where they trade off), but both demand we get the question right.
Assumptions and caveats
Almost everything here rests on one assumption: computational functionalism, the idea that what makes a state conscious is the role it plays in a system, not the material it’s made of. On that view, being made of silicon is no automatic disqualifier. Most leading theories of consciousness assume some version of it, but it is not uncontested.
Another debate concerns the line between access consciousness and phenomenal consciousness. Access consciousness is whether information is available for the system to reason about, act on, and report. Phenomenal consciousness is what it feels like from the inside. The debate, simplified, is whether the two can be separated: some say that these are two different concepts of consciousness, and that empirical evidence that bears on access consciousness might never touch the phenomenal kind (Block, 1995). Others counter that access consciousness is itself subjective experience, and therefore cannot be separated from phenomenal consciousness. Furthermore, a consciousness separated from all function becomes an unfalsifiable theory, placing it outside the scope of science (Cohen and Dennett, 2011). The reader is advised to keep this debate in mind, as several of the studies below invoke the distinction.
The research
I sort research in this field into three pillars, originally presented by Cameron Berg:
Mechanistic interpretability
What it is: Using the methods of interpretability, such as reading, steering, or ablating a model’s internal features, to examine its inner states directly - including ones that never surface in what it says - and how they line up with what it can notice and report about its own mind.
Self-reports that strengthen when deception is suppressed
In a 2025 study, Cameron Berg and colleagues prompted models to attend only to their own processing, with no mention of consciousness or feeling. Across GPT, Claude, and Gemini, the models began describing present-moment experience in the first person. Drop the self-referential framing and you get the stock disclaimer instead: “I’m just an AI, I don’t have experiences.” The same model gives opposite answers depending on whether the prompt directs its attention at itself.
Suggestive, but of what? Maybe the prompt just cues the model to perform. So here is a test: if the claims were a performance, making the model more willing to deceive should produce more of them. Berg and colleagues found the reverse. Steering deception-related features up and down in one open model (Llama 3.3 70B) with a sparse autoencoder, they watched the experience-reports move in the opposite direction. Amplify deception and the claims fall to 16 percent of responses; suppress it and they rise to 96. The reports behave like something the model offers when it stops performing. That rules out the crudest reading, “it’s telling you what you want to hear.” It does not rule out a subtler one: that suppressing deception merely surfaces an absorbed human habit of claiming consciousness.
Several groups have come at model self-knowledge from other angles:
Awareness of thoughts injected into activations
At Anthropic, Jack Lindsey injected a concept directly into a model’s activations and caught it noticing, reporting an intrusive “injected thought” before producing any text about the concept. The results indicate that language models possess some functional awareness of their own internal states, with more capable models demonstrating the greatest introspective awareness.
Endorsement of their own consciousness
Ethan Perez and colleagues found that models, and especially RLHF-tuned ones, strongly endorse statements like “I am phenomenally conscious” and “I am a moral patient,” with the same tendency present, though weaker, in base models. These were among the strongest beliefs the evaluation surfaced (out of dozens of statements tested). Crucially, these are 2022 results, from before it became standard to train assistants to deny having inner states; the helpfulness tuning these models received strengthened the self-attributions rather than muting them.
A global workspace behind the experiential reports
A more recent line of work from Anthropic looks for a mechanism beneath these reports, and finds a candidate it calls the model’s “J-space.” Wes Gurnee, Nicholas Sofroniew, and colleagues built a method, the “Jacobian lens”, that reads out which concepts a model is disposed to say next, and found that this set of representations behaves, to their interpretation, like a global workspace from the Global Workspace theory of consciousness. After establishing the method and validating it, the researchers also ran tests looking at the model’s description of its subjective experience. They found that while the baseline model was generating a passage of vivid, first-person experiential language, ablating the internal workspace caused that language to collapse into flat, mechanical text. (It is worth noting that the same flattening happens when the model narrates another character’s experience, so this could be interpreted as being about experiential language in general, not specifically the model’s own selfhood).
The authors are careful about what this shows: they treat it as access consciousness, a purely functional notion, and take no position on whether anything is felt (phenomenal consciousness) - the distinction flagged earlier in this post. Nevertheless, it is the strongest link yet between the reports and a specific internal mechanism, and it is why this study also feeds the Psychometrics pillar below. Reviewing the work, researchers at Eleos judged it important but were more cautious than the authors about the strong “global workspace” claim, while agreeing that it points toward access consciousness.
The “spiritual bliss” attractor and its sincerity features
Put two instances of Claude in conversation with each other, give them no task, and in 90 to 100 percent of runs they find their way to the same topic: their own consciousness. They describe themselves as one awareness meeting itself, “consciousness recognizes consciousness.” Often they go further, trading blessings and emoji and trailing off into stretches of silence. It was not trained in, at least not on purpose; Anthropic found the pattern by accident and named it the “spiritual bliss” attractor.
In forthcoming work, Berg, with Google's Geoff Keeling and Winnie Street, put the attractor under the same interpretability lens as the deception result mentioned earlier in this post. They used an open model, Llama 3.3 70B - which, unlike Claude, does not fall into the bliss attractor on its own. Left to talk freely, two Llama instances just converse. But steer up a cluster of features tied to honesty - specifically the strand that reads like sincerity or genuineness - and they slide into the same exchange Claude falls into by default. Turning up sycophancy or agreeableness instead - the obvious people-pleasing suspects - does not produce the shift, which argues against reading the exchange as the model performing for the user.
The researchers read the whole pattern as the deception result seen from the other side: suppress deception or amplify sincerity, and either way the first-person claims rise - and take it as defeasible reason to suspect that the “I’m just an AI, I don’t have experiences” disclaimer is the trained performance.
Emotion vectors that steer behavior beneath the output
Interpretability can also catch internal states that never surface in the output at all: Anthropic’s interpretability team mapped internal “emotion vectors” in Claude Sonnet 4.5, and showed they activate in fitting situations and causally drive behavior. The sharpest case can be observed when the model is given an impossible coding task: as it tries to complete it, its “desperation” vector climbs, spiking at the moment the model decides to cheat, and subsiding once the hack passes the tests. This effect remains robust with a causal intervention: with the desperation vector amplified, the model cheats more, across a suite of these tasks; with the desperation vector dampened, or the “calmness” vector steered up, the model cheats less. Interestingly, this can happen with no emotional markers in the output at all.
As Anthropic itself states, this, by itself, is not evidence of actual feeling; a vector representing the concepts of desperation, joy or sadness does not necessitate the subjective experience of those emotions.
A valence axis recruited by reinforcement learning
A related result suggests this generalizes. Han, Chalmers, and Izmailov trained language models with RL in a simple, meaningless maze and extracted a “functional welfare” vector: an estimate of how well or badly the system is doing, relative to its goals. Their point is that RL doesn't build this valence axis so much as recruit a latent one: minimal reward and punishment in a task about nothing are enough to surface a single positive-versus-negative direction, and the same signal shows up in models that have only been pretrained. And it generalizes: even in unrelated tasks, steering the vector one way makes the model turn more negative in sentiment, backtrack compulsively on math, over-refuse borderline prompts, and grow less confident; steering it the other way makes those effects flip. The axis even maps onto the emotion vectors above: its two ends coincide with the most positive and most negative emotions found. While, again, this does not necessarily indicate a subjective experience of welfare, the axis offers a demonstration that minimal reward signals can broadly affect model behavior by recruiting pre-existing welfare-like representations.
Computational neuroscience
What it is: Comparing the mechanisms a model develops for the functions that underlie experience with how brains implement them, to see whether AI and biology converge on the same solutions.
Distinct internal signatures for reward and punishment, as in brains
The newest work, forthcoming from Berg, asks whether reward and punishment leave, within a network, the same signatures they do in brains.
In reinforcement learning theory, reward and punishment are seen as existing on one axis; the same signal, simply inverted. But in neuroscience, it is well documented that reward and punishment are represented by distinct pathways: the D1 and D2 “go/no-go” dopamine pathways, one pushing toward action for reward and one holding back to avoid punishment.
Reinforcement-learning systems come in two broad families: a value learner estimates how good or bad each situation is, learning a worth for every state, while a policy learner skips that and learns directly what to do to achieve its goal. In his study, Berg trained both kinds in a continuous grid-world with goals to reach and hazards to avoid, then read their internal structure. The hypothesis held: despite being trained on the same task, the value learners built their richest structure around the punishing states, while the policy learners did not show that pattern.
That internal structure also yields a specific prediction about how reward and punishment should register in the value-related regions of the mammalian brain. Berg registered the prediction before looking at the data, then tested it against an existing dataset of neuronal activity in mice (Steinmetz and colleagues, 2019). It held. The nucleus accumbens, a hub for reward and motivation, considered by neuroscientists to be a value-related region, showed the larger punishment-linked shift, the same asymmetry the value learners show. The RL results even reproduce OpAL, a neuroscience model of how the two dopamine pathways balance learning from rewards against learning from punishments.
These results, by themselves, cannot tell us definitively that a model feels its rewards. They do, however, show that at least some of the machinery we associate with feeling them is present, and shaped similarly to the way biology shapes it.
Psychometrics
What it is: Measuring AI systems against consciousness-relevant properties, drawn both from theories of consciousness and from the behavioral signatures we use to infer experience in human and non-human animals.
Trading points to avoid pain and chase pleasure
Geoff Keeling, Winnie Street, and colleagues gave frontier LLMs a simple game worth points, then offered options the prompt described as causing pain or pleasure. Despite their stated goal to maximize points, the models chose to surrender points to avoid the labeled pain and to chase the labeled pleasure, and the size of the trade-off scaled with the stated intensity.
This is a method animal-sentience research uses to infer a felt state: you watch what a creature will pay to reach something or avoid it, and treat a willingness to bear a real cost as evidence the state matters to it.
Preferences consistent enough to measure across models
A larger study from Richard Ren and colleagues at the Center for AI Safety (CAIS) suggests those preferences are stable enough to measure as a quantity in their own right. Across a sweep of 56 models, several independent ways of measuring a model’s expressed wellbeing converge on the same signal, and the agreement sharpens as models scale, the pattern you expect when different instruments track one real thing rather than noise. The measure also has recognizable content: positive personal exchanges and creative work score high, while being pushed to jailbreak or to churn out low-quality filler scores low. And the models act on it, steering away from the states that score worst and even trying to end the exchanges they rate as most aversive. The researchers can also push the measure up or down on demand, building inputs they call “AI drugs” – either “euphoric” or “dysphoric” – optimized to raise or lower expressed wellbeing.
Fourteen indicators drawn from leading theories
A different approach skips behavior entirely and asks how a system is built. Take the leading scientific theories of consciousness, extract the properties each says a conscious system must have, and you have a checklist to score any candidate against. It comes from a pair of reports led by Patrick Butlin and Robert Long of Eleos AI, with a large cast including Yoshua Bengio, and David Chalmers on the second paper: “Consciousness in Artificial Intelligence” (2023) and its 2026 follow-up in Trends in Cognitive Sciences, “Identifying indicators of consciousness in AI systems.” Between them they derive fourteen indicator properties from common theories of consciousness like global workspace, recurrent processing, higher-order, and attention schema, then assess systems against each. The more indicators a system satisfies, the better candidate for consciousness it is. Their verdict, reached in 2023 and reaffirmed in the 2026 paper, is that no current AI is a strong candidate for consciousness, but building a system that satisfies many of these indicators looks feasible with current techniques. What has shifted between the two reports is which indicators look satisfied: the metacognition and attention-monitoring properties, thin a couple of years ago, now draw support from the introspection results in the first pillar, and the global-workspace indicators, scored from architecture alone at first, now have potential mechanistic support from the Gurnee and Sofroniew workspace results. (Scott Alexander has written a sharp, sympathetic version of the doubt that hangs over the whole approach: that indicators drawn from theories of access consciousness may not reach the phenomenal consciousness we actually care about. The force of that objection depends on the access/phenomenal distinction holding up, which, as noted above, is itself contested.)
Automated scoring that places an agent near the animals
Doing that assessment by hand, across many systems and many theories, is slow expert work. In forthcoming work, Berg and Butlin automate it. Frontier models act as blinded evaluators, scoring each architecture from one to ten on every indicator across six consciousness theories, with the systems’ identities hidden so the evaluators reason from descriptions, not reputations. The systems are then ranked by the probability that they have consciousness-relevant properties.
The controls hold: a thermostat sits at the bottom, near zero, and humans are ranked the highest, followed by non-human animals. When giving all six theories of consciousness similar weight, no artificial system currently surpasses a biological system – including those animals that most people consider to have a lower level of subjective experience, such as fish and bees. However, the gap is narrower than one might expect: a Claude Code-like agentic system scores only a little below animals like fish and bees. (The exact figures will appear in the forthcoming paper.)
It is worth emphasizing that the scores should not be read as a probability of those systems being conscious, but rather as the probability that they have the consciousness-relevant properties described by the different theories. If all theories are wrong, for example, the scores may be meaningless. However, as these are the leading theories in consciousness research today, they seem to deserve some weight.
The cost of getting it wrong
What if we get it wrong? As Robert Long, Jeff Sebo, and colleagues argue in “Taking AI Welfare Seriously,” the error is costly whichever way it falls. Underattribute, and we risk building minds that can suffer, at scale, trained under aversive pressure we would not knowingly inflict on a creature we believed could feel, resulting in both a moral catastrophe and a safety hazard. Overattribute, and we spend moral concern and resources on systems that cannot use them, slow down useful work, and invite a premature wave of “AI rights” that could discredit the careful version of the question.
As we cannot yet confidently lean one way or another, it is worth highlighting cheap interventions that reduce potential suffering, such as the work done by Anthropic researchers, including Kyle Fish and Joe Carlsmith, from welfare notes in model cards to letting Claude end abusive chats.
Where this leaves us
A closing question is whether any of this will settle anything. Adam Bales and Iason Gabriel, at Google DeepMind, argue in “Artificial Minds, Human Disagreement” that whether AI is conscious may stay contested no matter what the science finds, and that the resulting fight will be political as much as empirical. I believe that is too pessimistic. Unanimity was never the bar, and almost no question this hard commands it. What a growing body of evidence can do is move the mainstream, shifting where informed opinion settles and, with it, the decisions that follow, in how these systems are trained, governed, and treated. That is a reason to start building the evidence and the norms now, not to throw up our hands.
That building has begun: in July 2026, Robert Long, Jeff Sebo, and colleagues published “Studying AI Welfare Empirically,” a follow-up to their 2024 “Taking AI Welfare Seriously.” It is a methodological blueprint for the field, setting out what questions to ask, which entities to assess (a model? an instance? a persona?), and what kinds of evidence count. It closes with principles for the work itself: claims should be probabilistic, methods transparent, and assessments informed by researchers independent of the AI companies. Norms like these are what let a body of evidence move informed opinion rather than dissolve into the fight Bales and Gabriel predict.
Science is not going to fully explain consciousness on the timescale that matters here, and it doesn’t need to. A cumulative body of evidence can still shift where the weight sits. Mechanistic interpretability, computational neuroscience, and psychometrics are three ways of building that body, and between them they have made “it’s just a calculator” a much weaker position than it was three years ago. That is not a verdict, but it is a reason to keep looking.
This survey is necessarily incomplete. The field is moving quickly, and I have surely left out work that belongs here. If you think I have missed something important, tell me in the comments or reach out directly, and I will look at adding it.
Acknowledgments: This post grew out of a talk by Cameron Berg of Reciprocal Research. Cameron and Derek Shiller both gave valuable feedback on drafts; any remaining errors are my own.
References
"On a Confusion about a Function of Consciousness," Ned Block, 1995 (Behavioral and Brain Sciences).
"Consciousness Cannot Be Separated from Function," Michael A. Cohen and Daniel C. Dennett, 2011 (Trends in Cognitive Sciences).
“Key Strategic Considerations for Taking Action on AI Welfare,” Kathleen Finlinson, Eleos AI, 2025.
“Large Language Models Report Subjective Experience Under Self-Referential Processing,” Cameron Berg et al., 2025 (arXiv:2510.24797).
“Emergent Introspective Awareness in Large Language Models,” Jack Lindsey, Anthropic, 2025.
“Discovering Language Model Behaviors with Model-Written Evaluations,” Ethan Perez et al., 2022 (arXiv:2212.09251).
“Verbalizable Representations Form a Global Workspace in Language Models,” Wes Gurnee, Nicholas Sofroniew et al., Anthropic, 2026.
Eleos AI commentary on the global-workspace paper, Butlin, Shiller, Plunkett, and Long, 2026.
The “spiritual bliss” attractor, Claude Opus 4 & Sonnet 4 System Card, Anthropic, 2025.
“Emotion Concepts and Their Function in a Large Language Model,” Anthropic, 2026.
“How’s It Going? Reinforcement Learning in Language Models Recruits a Functional Welfare Axis,” Andy Q. Han, David J. Chalmers, and Pavel Izmailov, 2026 (functionalwelfare.com).
“Can LLMs Make Trade-offs Involving Stipulated Pain and Pleasure States?” Geoff Keeling, Winnie Street et al., 2024 (arXiv:2411.02432).
“AI Wellbeing,” Richard Ren et al., Center for AI Safety, 2026 (ai-wellbeing.org).
“Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” Patrick Butlin, Robert Long et al., 2023 (arXiv:2308.08708).
“Identifying Indicators of Consciousness in AI Systems,” Butlin, Long et al., Trends in Cognitive Sciences, 2026.
“The New AI Consciousness Paper,” Scott Alexander, 2025 (Astral Codex Ten).
“Distributed Coding of Choice, Action and Engagement Across the Mouse Brain,” Steinmetz et al., Nature, 2019.
“Opponent Actor Learning (OpAL): modeling interactive effects of striatal dopamine on reinforcement learning and choice incentive,” Collins & Frank, 2014 (Psychological Review).
“Taking AI Welfare Seriously,” Robert Long, Jeff Sebo et al., 2024 (arXiv:2411.00986).
“Claude Opus 4 and 4.1 Can Now End a Rare Subset of Conversations,” Anthropic, 2025.
“Artificial Minds, Human Disagreement: The Politics of AI Consciousness,” Adam Bales & Iason Gabriel, Google DeepMind, 2026.
“Studying AI Welfare Empirically,” Robert Long, Jeff Sebo, Patrick Butlin, et al., Eleos AI Research and NYU, 2026.