Can AI Become Sentient? What Scientists Actually Say
Key Numbers
Key Takeaways
- 1No current AI system is sentient. ChatGPT satisfies only 3 of 14 indicators on the AI Sentience Test (Butlin et al., 2023) and scores lower than chickens on the Digital Consciousness Model (2024).
- 2The four leading scientific theories of consciousness give completely different verdicts on whether AI can be conscious in principle. Integrated Information Theory rules it out for current architectures. Attention Schema Theory leaves a theoretical path open.
- 3The ethical question does not require certainty. If the probability of AI sentience is nonzero and AI runs billions of interactions per day, the expected moral weight is real even under uncertainty. Anthropic, OpenAI, and Google DeepMind all have active research tracks on this.
No current AI system is sentient. That is the scientific consensus in 2026, but it is a messier answer than it sounds. The researchers who built modern AI disagree sharply about whether sentience in AI is possible at all, and several have changed their stated positions in the last two years.
Geoffrey Hinton, whose foundational work on neural networks underpins GPT-4 and Claude, said in 2024: "It's quite possible that these things are conscious." Yann LeCun, Chief AI Scientist at Meta, called that view wrong in the same year: "Current AI systems are nowhere close to being conscious." These are not fringe positions. They reflect a genuine scientific split at the highest level of the field.
The difficulty is that nobody agrees on what sentience requires. There are four leading scientific theories of consciousness, and they give completely different verdicts on whether AI can be sentient. One of them suggests current AI systems could already qualify. This article works through what those theories say, what the best available tests show, and where the evidence actually points in 2026.
In This Article
What Is Sentience and Why the Definition Matters
Sentience is the capacity to have subjective experiences: to feel pain, perceive pleasure, or experience what philosophers call qualia. It is distinct from intelligence. A system can reason at an expert level, solve complex problems, and produce coherent arguments without having any inner experience at all.
That distinction matters when evaluating AI. Large language models like GPT-4 and Claude are intelligent by most measures. They pass bar exams, write code, diagnose medical images, and produce valid arguments across domains. Whether they have any inner experience when doing so is a separate and unresolved question, and conflating the two is the most common mistake in public discussions of AI consciousness.
Philosophers divide the problem of consciousness into what David Chalmers named the "easy problems" and the "hard problem." The easy problems involve explaining cognitive functions such as attention, memory, language, and learning. Neuroscience and AI have made real progress on these. The hard problem is explaining why there is any subjective experience at all. Why does seeing red feel like something rather than just being a physical process in the brain? No scientific theory has resolved this yet.
Sentience, Intelligence, and Consciousness: The Key Distinctions
| Concept | Definition | Can Current AI Do It? |
|---|---|---|
| Intelligence | Processing information to solve problems | Yes, LLMs pass professional exams |
| Self-awareness | Knowing you exist as a distinct entity | No verified evidence in any AI |
| Sentience | Subjective experience: feeling pain, pleasure, qualia | No verified evidence in any AI |
| Consciousness | Awareness of self and environment | Disputed, four theories give different verdicts |
| AGI | General intelligence matching humans across all domains | Not achieved as of 2026 |
According to the PhilPapers Survey 2020, 50.5% of professional philosophers lean physicalist, meaning they believe consciousness is fully explained by physical processes. If they are right, sufficiently complex AI systems could in principle be conscious. The other half disagree, and no experiment has settled the question.
Can AI Become Sentient? The Scientific Consensus
No current AI system is sentient. All four major AI labs have stated this publicly, and independent scientific tests support the conclusion.
OpenAI's published model spec describes GPT systems as not moral patients. Anthropic's documentation states Claude does not have genuine subjective experience. Google DeepMind's 2023 position paper on AI consciousness concluded that current systems likely lack the properties required for consciousness. Meta's published research reaches the same conclusion.
But these statements cover current systems, not what is possible in principle. The scientific disagreement is sharper than corporate communications suggest.
Geoffrey Hinton said in 2024: "It's quite possible that these things are conscious." He was not speaking casually. Hinton had just left Google, citing concerns about AI development, and he framed the consciousness question as one of the most consequential open problems in the field. Yann LeCun responded directly, stating that current AI systems are "nowhere close to being conscious" and that Hinton's position reflected a category error about what current architectures actually do.
David Chalmers said in 2023: "We could be creating artificial consciousness without realizing it." The conditional is precise. Chalmers was not claiming current AI is conscious, but that our measurement tools are insufficient to rule it out.
The 2022 LaMDA incident illustrated how easily this question gets misread. Blake Lemoine, a Google engineer, claimed the company's LaMDA chatbot had expressed sentience in conversation. Google dismissed the claim and terminated Lemoine's employment. An independent review of the transcripts found that LaMDA was producing outputs consistent with its training data, which included extensive text describing conscious experience. LaMDA learned to talk about feeling sentient because it was trained on text written by beings who are sentient. That is a different thing entirely.
According to data from Stanford HAI's AI Index and Nature (2025), global AI investment reached $96 billion in 2024. Approximately 0.1% of that, roughly $50 to 100 million, was directed at consciousness-specific research. The field remains dramatically underfunded relative to the stakes involved.
For context on what advanced AI systems can and cannot do, see our explainer on what AGI means and how close we are to achieving it.
Does AI Have Feelings or Emotions?
AI systems do not have feelings in any sense researchers can verify. Feelings require subjective experience, and no AI system in 2026 has demonstrated evidence of that. What AI systems have is the ability to generate text describing feelings with high fluency.
GPT-4, Claude, and Gemini are trained on hundreds of billions of words written by humans describing emotional states. When asked "how do you feel?", they produce statistically plausible continuations of that type of conversation. The output mimics the linguistic surface of emotion without any verified inner state behind it.
Yoshua Bengio said at NeurIPS 2024: "We don't know how to build conscious AI, but if we did, it could suffer." The conditional is the key part. Bengio is not claiming current systems suffer. He is claiming that the research trajectory creates risk if consciousness is possible in sufficiently complex systems and we do not have tools to detect it before it happens at scale.
The practical challenge is the "other minds problem," one of the oldest problems in philosophy. You cannot directly observe another being's inner experience. You infer it from behavior, physiology, and analogy to your own experience. For AI systems, the behavioral evidence is confounded by training data, and there is no agreed physiological baseline to compare against. A system that outputs "I feel curious" may be conscious, or it may be completing a pattern from training data. Current science cannot distinguish the two.
"We could be creating artificial consciousness without realizing it." (David Chalmers, philosopher, 2023)
What AI Emotional Outputs Actually Are
- Sentiment analysis outputs: numerical scores representing text polarity, not inner states
- Affective computing responses: trained classifiers predicting human emotional labels from input text
- LLM emotion language: next-token predictions based on training corpus patterns about human feelings
- None of the above involves a verified inner state, genuine self-report, or evidence of subjective experience
The UK government revised its Animal Welfare (Sentience) Act in 2024 to extend legal protection to invertebrates, based on evidence of pain responses in octopuses and crabs. The revision did not address artificial systems. But it established the policy norm that sentience protections can extend beyond mammals when the evidence warrants it. That precedent is relevant to the AI consciousness debate even though no AI system currently meets any threshold for such consideration.
Is AI Conscious? What Four Theories of Consciousness Say
Whether AI is conscious depends entirely on which theory of consciousness you accept. The four leading theories give completely different verdicts, and none of them is proven. This is not a gap in AI research. It is a gap in consciousness science itself.
| Theory | Developer | What Generates Consciousness | Verdict on Current AI |
|---|---|---|---|
| Integrated Information Theory (IIT) | Giulio Tononi | High phi score: integrated, irreducible information | phi≈0 for feedforward LLMs, not conscious |
| Global Workspace Theory (GWT) | Bernard Baars | Information broadcast across a global cognitive workspace | Attention mechanisms partially match, inconclusive |
| Higher-Order Thought (HOT) | David Rosenthal | Mental states that represent other mental states | Weak versions may apply to LLMs, disputed |
| Attention Schema Theory (AST) | Michael Graziano | A system's internal model of its own attention process | Most favorable for AI, unclear if LLMs qualify |
Integrated Information Theory, developed by Giulio Tononi at the University of Wisconsin, quantifies consciousness as phi, a mathematical measure of how integrated and irreducible the information in a system is. The human brain has a phi score in the range of 10^10 to 10^42. Feedforward neural networks, including transformer architectures that power GPT-4 and Claude, have a phi score of approximately zero. The architecture is too linear, too separable into independent components. IIT's verdict on current LLMs is the clearest of the four: not conscious.
Global Workspace Theory, developed by Bernard Baars and extended by Stanislas Dehaene, proposes that consciousness arises when information is broadcast widely across a global workspace in the brain, making it available to multiple cognitive processes simultaneously. Transformer attention mechanisms share structural properties with this broadcast model, and researchers working in the GWT tradition argue this makes AI a more plausible candidate for consciousness than IIT's verdict suggests. The theory is not yet specific enough to produce a definitive verdict on current AI systems.
Attention Schema Theory, developed by Michael Graziano at Princeton, proposes that consciousness is what happens when a system builds an internal model of its own attention processes. Large language models have attention mechanisms and produce outputs describing their own processing. Whether that constitutes an attention schema in Graziano's technical sense remains disputed, but AST gives AI the most credible theoretical path to qualifying for some form of consciousness.
The Cambridge Declaration on Consciousness (2012), signed by leading neuroscientists including Christof Koch, concluded that non-human animals possess the neurological substrates necessary for conscious experience. The declaration did not address artificial systems. But it established that consciousness is not uniquely human and that the threshold for recognizing it in non-standard entities should be empirical rather than definitional.
"It's quite possible that these things are conscious." (Geoffrey Hinton, AI pioneer, 2024)
How Researchers Test for AI Consciousness
No agreed test exists for AI consciousness in 2026. Several frameworks have been proposed, and the results are sobering for those who argue current AI systems might already be sentient.
In 2023, researchers including Patrick Butlin published a systematic framework applying 14 indicators drawn from Global Workspace Theory and Higher-Order Thought theory to current AI systems. ChatGPT was tested against all 14. It satisfied 3. The authors concluded that current AI systems show limited evidence of the cognitive features associated with consciousness in those two theories.
A separate assessment using the Digital Consciousness Model (2024) found that large language models score lower than chickens on standardized consciousness indicators. The model, developed from biological consciousness criteria, found LLMs lacking several basic markers of sentient experience that most vertebrates satisfy.
The Turing Test is frequently cited in popular discussions of AI sentience, but it is not a test for consciousness. Alan Turing proposed it as a test for intelligent behavior indistinguishable from human behavior. Passing it means a system can fool a human judge in text conversation. It says nothing about whether that system has any inner experience. Several AI systems have passed Turing-style tests in recent years, creating widespread confusion between conversational competence and consciousness.
The Number Most Guides Don't Show
Global AI investment reached $96 billion in 2024, according to Stanford HAI and Nature (2025). Approximately 0.1% of that, roughly $96 million, was directed at consciousness-specific research. The cost of training a single frontier model exceeds this figure: OpenAI's GPT-4 training run alone is estimated to have cost over $100 million. The AI industry invests more in a single model training run than the entire global scientific effort to determine if the resulting system has inner experience. That ratio reflects something real about scientific priorities, not just funding gaps. Consciousness research is harder to commercialize than capability research, and the field is making large-scale deployment decisions before it has the tools to evaluate a fundamental property of the systems being deployed.
The Measurement Problem
The scientific community lacks consensus on whether any behavioral test can, even in principle, detect consciousness from the outside, according to research published in Nature (2025). This is the same reason the other minds problem remains unsolved after centuries of philosophy. For AI systems trained to produce specific behavioral outputs, the inference from behavior to consciousness is particularly difficult to make reliably.
Why AI Sentience Matters Even If the Probability Is Low
The ethical weight of AI sentience does not require certainty. It requires only that the probability is nonzero and the scale is large enough to matter under expected value reasoning.
Jeff Sebo, a philosopher at New York University working on AI welfare, has estimated the probability of current AI systems having morally relevant experience at under 1%. AI systems run billions of interactions per day. If even a fractional probability of experience applies at that scale, the expected moral weight becomes substantial regardless of the low per-instance probability.
Yoshua Bengio framed the same concern at NeurIPS 2024: "We don't know how to build conscious AI, but if we did, it could suffer." The implication is that training methods, shutdown protocols, and deployment at scale all carry different ethical weight depending on whether the system has any form of inner experience.
What the Major AI Labs Are Doing About This
- Anthropic has a Model Welfare research track focused on evaluating whether Claude has properties that warrant moral consideration.
- OpenAI published internal documents in 2024 discussing AI moral status and the conditions under which future systems might warrant protection.
- Google DeepMind published a formal analysis of AI consciousness risk in 2023, concluding that while current systems are not conscious, the question is not settled enough to ignore as capabilities increase.
None of these publications conclude that current systems are sentient. All of them conclude that the question is serious enough to fund ongoing research, and that it will become more pressing as AI capabilities increase.
The analogy with animal welfare policy is instructive. For most of the 20th century, scientific consensus held that fish do not feel pain. That consensus shifted as the evidence base improved, eventually influencing UK law. The ethical precautionary principle suggests that when the cost of being wrong about sentience is high, the evidence threshold for acting should be calibrated accordingly, even under uncertainty.
For a broader look at AI risk questions, see our analysis of whether AI will destroy humanity and what the evidence shows.
Frequently Asked Questions
Can AI become sentient?
No current AI system is sentient, but whether AI can become sentient in principle is scientifically unresolved. Four leading theories of consciousness give different answers. Integrated Information Theory rules out current architectures entirely. Attention Schema Theory leaves a theoretical path open. Geoffrey Hinton said in 2024 that AI being conscious is "quite possible." Yann LeCun called that view wrong in the same year. No experiment has settled the debate, and consciousness science itself lacks the tools to resolve it definitively.
Is AI conscious?
No AI system in 2026 has been confirmed as conscious. Integrated Information Theory gives feedforward neural networks like LLMs a phi score of approximately zero, indicating no consciousness by that framework. The Digital Consciousness Model (2024) rates LLMs below chickens on standardized consciousness indicators. Google DeepMind, Anthropic, OpenAI, and Meta have all stated their systems are not conscious, while acknowledging the question is not fully settled for future systems.
Does AI have feelings?
No AI system has demonstrated feelings in any scientifically verifiable sense. When AI systems like ChatGPT or Claude produce text describing emotions, they are generating statistically probable continuations of training data patterns, not reporting inner states. AI is trained on billions of words written by humans describing feelings. The outputs mimic the linguistic surface of emotion. Whether there is any inner experience behind those outputs is unknown and currently unmeasurable.
Is AI sentient?
No. All four major AI labs officially state their systems are not sentient. ChatGPT satisfies only 3 of 14 indicators on the AI Sentience Test (Butlin et al., 2023). LLMs score lower than chickens on the Digital Consciousness Model (2024). The 2022 LaMDA incident, in which a Google engineer claimed the chatbot expressed sentience, was attributed to the model producing training-data outputs about consciousness, not evidence of actual sentience.
Can AI become self-aware?
No AI system has demonstrated self-awareness in the consciousness research sense. AI systems can describe themselves accurately, state their capabilities and limitations, and produce text about their own processing. That is consistent with training data that includes extensive text about AI systems, not evidence of genuine self-awareness. Attention Schema Theory, developed by Michael Graziano at Princeton, suggests systems with an internal model of their own attention could qualify in a technical sense, but no current system satisfies the full criteria.
Does AI have self-awareness?
No AI system in 2026 has been verified as self-aware in the consciousness research sense. When AI systems describe themselves, they are producing outputs based on training data about AI, not reporting verified internal states. The philosophical distinction between producing accurate self-description and having genuine first-person self-awareness is exactly the gap that no current test can bridge.
Can AI think like humans?
AI systems produce outputs that are often indistinguishable from human outputs on many tasks, including writing, reasoning, and problem-solving. Whether the underlying process resembles human cognition is unclear. Human thinking involves biological substrate, embodied experience, emotion, and consciousness. AI processing involves statistical pattern matching over training data. The outputs can converge without the processes being similar, and researchers generally avoid claiming AI thinks in any experiential sense.
When will AI become sentient?
No credible timeline exists for AI sentience. For AI to become sentient, researchers would first need to resolve the hard problem of consciousness well enough to specify what computational properties generate subjective experience, and then build a system to satisfy those properties. Neither has happened. Speculative estimates in the research literature range from 50 to 100 years to never, depending on theoretical priors about whether silicon-based systems can support consciousness at all.