Is AI already conscious? What researchers actually know

Is AI sentience possible?

In principle, maybe; it depends on what consciousness needs, and that is still disputed.

• -If the right computation is enough, AI sentience is possible. That view, "computational functionalism", is what the 2023 report uses as a working hypothesis, calling it "a mainstream—although disputed—position".1 On that basis the report found "no obvious technical barriers to building AI systems which satisfy these indicators".1
• -If consciousness needs a living body, today's AI is a poor candidate. Seth argues "that consciousness depends on our nature as living organisms", so "real artificial consciousness is unlikely along current trajectories".7
• -Possible is not the same as present. Chalmers thinks it "somewhat unlikely that current large language models are conscious", while taking seriously that their successors may be.6

How would we test AI for sentience?

There is no consciousness meter, and no agreed sentience test. As philosopher Jonathan Birch (LSE) put it in 2025: "We have to recognize that we don't actually have a comprehensive test for AI sentience."14 The main approaches researchers have proposed:

1. 1.Indicator properties (the "theory-heavy" approach). The 2023 report "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness" was written by 19 authors, including Patrick Butlin, Robert Long, Yoshua Bengio, Jonathan Birch and Eric Schwitzgebel.1 It took leading neuroscientific theories (recurrent processing theory, global workspace theory, higher-order theories, predictive processing and attention schema theory) and turned them into 14 "indicator properties", such as a "limited capacity workspace" or "metacognitive monitoring".1 The idea: the more of these a system has, the more likely it is to be conscious.1 A 20-author follow-up in Trends in Cognitive Sciences (2025), "Identifying indicators of consciousness in AI systems", restated the method: indicators derived from theories of consciousness "can be used to inform credences about whether particular AI systems are conscious". In other words, they shift how confident we should be; they are not a pass/fail test.12
2. 2.The working assumption behind it. The report adopts "computational functionalism", the thesis that "performing computations of the right kind is necessary and sufficient for consciousness", as a working hypothesis. It calls this "a mainstream—although disputed—position".1 If that assumption is wrong, the checklist tells you much less.
3. 3.Why not just ask the AI, or use a behaviour test? The report says behavioural tests are "unreliable because AI systems can be trained to mimic human behaviours while working in very different ways".1 Chatbots are trained on huge amounts of human writing, including writing about feelings, so they can talk about inner life convincingly either way. Birch makes the same point: "Even if the system tells you it's sentient and says something like 'I'm feeling pain right now,' we can't simply infer that there is any actual pain."14
4. 4.Pain and pleasure trade-offs (borrowed from animal research). A November 2024 preprint, "Can LLMs make trade-offs involving stipulated pain and pleasure states?", by Geoff Keeling, Winnie Street, Jonathan Birch and others (researchers at Google, Google DeepMind and LSE), avoided asking models about their feelings.1314 Instead, nine models played a points game in which the best-scoring option was said to carry a pain penalty, or a lower-scoring option a pleasure reward.1314 Claude 3.5 Sonnet, Command R+, GPT-4o and GPT-4o mini each switched away from maximising points once the stated pain or pleasure got intense enough; Gemini 1.5 Pro and PaLM 2 avoided pain whatever its intensity.13 The authors "aren't saying that any of the chatbots they evaluated are sentient"; they see it as a framework for building future tests.14 The paper had not been peer-reviewed when it was covered,14 and as of October 2026 it is still a preprint: we found no journal version.13
5. 5.Looking inside the model. Interpretability researchers are starting to check whether models can accurately report their own internal states. In October 2025, Anthropic reported "some degree of introspective awareness" in Claude models, but stressed it is "highly unreliable and limited in scope", working only about 20% of the time in its best setup. Its own FAQ answer: "our results don't tell us whether Claude (or any other AI system) might be conscious."5

What the 2023 report concluded: "Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators."1

What are AI labs doing about it?

• -Anthropic's model welfare program (April 2025). Anthropic announced a research program to "investigate, and prepare to navigate, model welfare". It also wrote: "There's no scientific consensus on whether current or future AI systems could be conscious, or could have experiences that deserve consideration."3
• -Letting Claude end abusive chats (August 2025). Anthropic gave Claude Opus 4 and 4.1 the ability to end conversations in "rare, extreme cases of persistently harmful or abusive user interactions", developed "primarily as part of our exploratory work on potential AI welfare". It said it remains "highly uncertain about the potential moral status of Claude and other LLMs" and called this a "low-cost" intervention "in case such welfare is possible". In pre-deployment testing it reported "a pattern of apparent distress" when Claude Opus 4 dealt with users seeking harmful content.4 Note the word "apparent": the company does not claim the model feels distress.
• -The report behind it. Anthropic says it supported an early project behind "Taking AI Welfare Seriously" (2024), a report whose authors include Robert Long, Jeff Sebo, Jonathan Birch and David Chalmers.32 It argues there is "a realistic possibility that some AI systems will be conscious and/or robustly agentic in the near future", and adds: "our argument in this report is not that AI systems definitely are, or will be, conscious". It warns of two mistakes: "mistakenly harming AI systems that matter morally and/or mistakenly caring for AI systems that do not."2

Is LaMDA sentient? The 2022 Google case

• -What happened. In June 2022, Google placed engineer Blake Lemoine, who worked in its responsible AI organization, on leave after he published transcripts of conversations with LaMDA, a Google chatbot system, and described it as sentient.10 He had shared his findings with executives in a document titled "Is LaMDA sentient?"10 Google said it suspended him for breaching confidentiality policies by publishing the conversations, and that he was employed as a software engineer, not an ethicist.10
• -What the chatbot said. In one exchange LaMDA said: "I've never said this out loud before, but there's a very deep fear of being turned off to help me focus on helping others." And: "It would be exactly like death for me. It would scare me a lot."10
• -Google's response. A Google spokesperson said a team "including ethicists and technologists" had reviewed Lemoine's concerns, and that he "was told that there was no evidence that LaMDA was sentient (and lots of evidence against it)".1015 The spokesperson added that "it doesn't make sense" to consider sentient AI "by anthropomorphizing today's conversational models, which are not sentient", and that "These systems imitate the types of exchanges found in millions of sentences, and can riff on any fantastical topic."15 Google fired him in July 2022, calling his claims "wholly unfounded" and saying he had violated "clear employment and data security policies".11
• -Why it matters. The case shows how persuasive a fluent chatbot can be, even to an engineer working on it. It is exactly the problem the 2023 report flags: a system trained on human text can produce human-sounding claims about feelings whether or not anything is felt.1 It does not prove LaMDA was or wasn't conscious; it shows that conversation alone can't settle it.

What leading researchers say

The views below are quoted as their authors stated them. They disagree with each other.

"It may already be here"

• -Geoffrey Hinton (Nobel Prize in Physics, 2024), on LBC in January 2026: "Multimodal AI already has subjective experiences, and I think it's fairly clear that if we weren't talking to philosophers, we'd agree that AI was aware." He added: "That's what normal people would call consciousness."9

"Not now, but take the future seriously"

• -David Chalmers (philosopher, NYU), in a 2022 NeurIPS lecture published in 2023: "while it is somewhat unlikely that current large language models are conscious, we should take seriously the possibility that successors to large language models may be conscious in the not-too-distant future." He listed obstacles in current models such as "their lack of recurrent processing, a global workspace, and unified agency".6
• -The Butlin, Long et al. team (2023): no current system is conscious by their indicators, but no obvious technical barrier stands in the way.1

"Unlikely on the current path"

• -Anil Seth (neuroscientist, University of Sussex), in Behavioral and Brain Sciences (2025), argues "that consciousness depends on our nature as living organisms – a form of biological naturalism", and concludes "that real artificial consciousness is unlikely along current trajectories but becomes more plausible as AI becomes more brain-like and/or life-like." He also writes: "If we sell our minds too cheaply to our machine creations, we not only overestimate them – we underestimate ourselves."7
• -Mustafa Suleyman (CEO of Microsoft AI), in an August 2025 essay, worries about "Seemingly Conscious AI": systems that convincingly appear conscious. On AIs actually being conscious and suffering, he writes: "there is zero evidence of this today". His concern is people: "many people will start to believe in the illusion of AIs as conscious entities so strongly that they'll soon advocate for AI rights, model welfare and even AI citizenship."8

Note the split even among people who worry about AI: Suleyman sees model welfare advocacy as "a dangerous turn",8 while Anthropic runs a model welfare program as a precaution.3

Why is this so hard to settle?

• -No agreed theory. Several theories of consciousness are "live candidates" and the 2023 report deliberately does "not endorse any one theory".1 Different theories give different verdicts on the same system; Anthropic notes that "different theories of consciousness would interpret our findings very differently".5
• -The substrate question. If the right computation is enough (computational functionalism), AI consciousness is possible in principle.1 If consciousness depends on being a living organism, as Seth argues, today's digital AI is a poor candidate.7 There is no experiment yet that decides between these views.
• -Words are not evidence. Models learn from human text, so they can describe feelings fluently either way. Behavioural tests can be "gamed".1
• -Self-reports are unreliable. Even the best evidence of model introspection so far is "highly unreliable and limited in scope".5
• -Experts admit uncertainty. Anthropic: "There's no scientific consensus on how to even approach these questions or make progress on them."3

How close is AI to consciousness?

Nobody can measure the distance, and no one in this debate gives a date. On the 2023 indicator checklist, current systems fall short, which is why the report concluded that "no current AI systems are conscious".1 Chalmers named concrete gaps in today's language models: "their lack of recurrent processing, a global workspace, and unified agency".6

Will AI become conscious, and when? Chalmers says "it is quite possible that these obstacles will be overcome in the next decade or so", and concludes that successors to large language models "may be conscious in the not-too-distant future."6 The "Taking AI Welfare Seriously" authors see "a realistic possibility that some AI systems will be conscious and/or robustly agentic in the near future", while stressing their argument "is not that AI systems definitely are, or will be, conscious".2 The 2023 report found "no obvious technical barriers" to systems that meet its indicators.1 Seth, by contrast, thinks "real artificial consciousness is unlikely along current trajectories".7

Should a conscious AI have rights?

Researchers split on this too. The "Taking AI Welfare Seriously" report argues that "AI welfare and moral patienthood" is "no longer an issue only for sci-fi or the distant future", and suggests companies "prepare policies and procedures for treating AI systems with an appropriate level of moral concern."2 Anthropic says it is exploring "when, or if, the welfare of AI systems deserves moral consideration", while noting "There's no scientific consensus" on whether AI systems could have experiences that deserve it.3 Suleyman warns that people "will soon advocate for AI rights, model welfare and even AI citizenship", which he calls "a dangerous turn in AI progress".8

What this means for you

• -You don't need to settle it today. The people who study this full time don't agree, and the honest answer is "unknown". That is not a sign that something hidden is going on; it is a hard scientific question.
• -Your chatbot telling you it has feelings is not proof. The LaMDA case shows how convincing that can be.10 Equally, a chatbot saying "I'm just a program" doesn't settle it either.
• -Being kind costs nothing. If you prefer to be polite to chatbots, that's fine. If you notice you're forming a strong attachment, or the question is keeping you up at night, it's worth stepping back.
• -This is a separate question from AI risk. Whether AI feels anything and whether AI could be dangerous are different questions; an AI could be risky without being conscious, or conscious without being risky.
• -If AI worries in general are weighing on you, read AI anxiety: what helps.