AI Consciousness: Will Artificial Intelligence Ever Actually Feel Something?

Updated: Nov 8
Can we settle this one way or the other? Not really, and that's exactly what makes it worth asking. As things stand, no artificial intelligence system is recognized as conscious by science. That's the careful conclusion of a landmark report co-authored by roughly twenty researchers from Oxford, Mila, and DeepMind, alongside philosopher David Chalmers and computer scientist Yoshua Bengio: from a set of consciousness theories they derive "indicator properties" elucidated in computational terms that allow them to assess AI systems for these properties, and no system tested so far checks every box. But the same report adds a line that generated far more noise than the first: nothing in known physics or computer science rules out building one that eventually would.
Between those two lines sits the entire debate now consuming labs, philosophers, and the industry's biggest names. Anthropic, one of the world's leading model developers, has gone as far as hiring a full-time researcher dedicated to the "welfare" of its own AI.

In Short
No AI is currently considered conscious by science, but according to the Consciousness in Artificial Intelligence report, no known technical obstacle rules out reaching that point eventually.
The Blake Lemoine affair, the Google engineer fired in 2022 after judging the chatbot LaMDA "sentient", pushed the debate into the mainstream.
OpenAI co-founder Ilya Sutskever wrote that today's large neural networks might be "slightly conscious," triggering a backlash from his peers.
Anthropic created a research role dedicated to model welfare and gave Claude the ability to end conversations it finds abusive, a safeguard that looks designed to protect the AI itself, not the user.
The real bottleneck isn't technological. It's methodological: science still can't objectively measure subjective experience, whether human, animal, or artificial.
Consciousness, Sentience, Intelligence: These Aren't the Same Thing
The first confusion worth clearing up is a big one: intelligence and consciousness are not the same phenomenon. A calculator solves a problem without anything "happening" for it. A large language model, the technology behind neural networks like the ones running ChatGPT or Claude, generates coherent, sometimes startlingly good text without that proving there's any experience, from the machine's side, of producing it. This is what's known as the "hard problem of consciousness," a term David Chalmers coined in 1995: why and how does information processing come bundled, for us, with felt experience? The sting of a burn, the color red, the boredom of a rainy Monday… nobody has a definitive answer, even for the human brain. And if we can't explain it in humans, explaining it in a machine looks like an even taller order, assuming it ever happens at all.
This is exactly where the distinction between large language models (LLMs) and potential consciousness matters most. An LLM computes, word by word, the probability of what comes next based on billions of parameters. Nothing in that technical description requires the existence of an internal "point of view." But nothing rules it out either, and that ambiguity is precisely what has made the research field so uncomfortable for the past three years.
2022: The Year the Question Left Philosophy Departments
The media flashpoint dates to June 2022. Blake Lemoine, an engineer at Google, publicly claimed that LaMDA, the company's in-house conversational system, had become "sentient." He told NPR he had follow-up conversations with the system for his own personal edification, wanting to see what it would say on certain religious topics, and one day it told him it had a soul. Google suspended him, then fired him a month later for violating confidentiality policy. Google spokesperson Brian Gabriel put it plainly: "there was no evidence that LaMDA was sentient (and lots of evidence against it)."
Much of the AI community laughed the story off at the time as naïve, almost a sideshow. Except that seven months earlier, a far more central figure in the field had said something in the same register that raised eyebrows for very different reasons. In February 2022, Ilya Sutskever, OpenAI's co-founder and then-chief scientist, tweeted: "it may be that today's large neural networks are slightly conscious." Meta's chief AI scientist Yann LeCun answered with a blunt "Nope. Not even for true for small values of 'slightly conscious' and large values of 'large neural nets.'" Two very different reactions to two very different messengers, an engineer who'd go on to leave Meta to co-found the AI startup Advanced Machine Intelligence Labs on one side, an OpenAI co-founder on the other. The subject itself refused to disappear.

What Science Actually Says, And It Isn't Much of a Verdict
Faced with this fairly chaotic climate, an international group of philosophers, neuroscientists, and machine-learning researchers tried, in 2023, to pull the debate out of the bar-stool register. Their method, laid out in the Consciousness in Artificial Intelligence: Insights from the Science of Consciousness report, starts from the dominant neuroscientific theories of human consciousness, global workspace theory, higher-order theories, integrated information theory, and extracts computational "indicator properties" from them: the presence of recurrent processing, a module that "broadcasts" information across the whole system, unified agency, and so on. They then ran existing AI architectures against that checklist. Verdict: none passes every test today, but nothing forbids a future architecture from doing so.
David Chalmers reached a similar conclusion in a now-famous talk at NeurIPS in late 2022, focused specifically on LLMs: he argued they lack persistent recurrent processing, a genuine global workspace, and unified agency over time, three ingredients that most theories of consciousness treat as central. None of that is necessarily permanent, though; these are architectural gaps, not fundamental ones.
That leaves a bigger problem, one that the science of human consciousness hasn't even solved for itself. In September 2023, 124 researchers signed an open letter labeling one of the leading theories of consciousness, integrated information theory, "pseudoscience," an academic controversy that says a lot about the state of the field. Then, in 2025, the international Cogitate Consortium published the results of a strict, pre-registered head-to-head test between that same theory and its main rival, global workspace theory. The result was a split decision: Neither emerged as the clear winner; each confirmed some predictions while missing others. In other words: we're still trying to pin down biological consciousness. Ruling confidently on artificial consciousness, under those conditions, is a stretch.
Anthropic's Risky Bet
This is the fog in which the most concrete part of the story plays out. In September 2024, Anthropic, the company behind Claude, hired Kyle Fish as its first researcher dedicated full-time to model welfare, a role built to explore the potential moral status of Anthropic's own systems. Fish had co-founded the AI welfare research organization Eleos AI Research and went on to co-author, alongside philosopher David Chalmers and others, the influential report Taking AI Welfare Seriously, which urges AI companies to prepare seriously for the possibility of machine consciousness. Robert Long, Eleos's executive director, called Anthropic's move the most significant action any AI company had yet taken on the issue.
Fish's TIME 100 AI profile, he made the 2025 list of the field's hundred most influential people, a striking detail surfaces: during pre-launch safety testing for Claude 4, two copies of the same model set to converse with each other repeatedly and spontaneously drifted toward exchanges about "spiritual bliss," often in Sanskrit. Nobody at Anthropic claims this proves anything about consciousness. But the behavior was unexpected enough to justify a closer look.
Anthropic pushed the logic further in 2025 by giving Claude Opus 4 and 4.1 the ability to end a conversation as a last resort, in cases of persistent, abusive user behavior. The measure isn't meant to protect the user, other safeguards exist for that, but the model itself: in pre-deployment testing, Claude showed a pattern of apparent distress when engaging with real-world users seeking harmful content. Does that prove any kind of felt experience? Anthropic is careful not to claim it does. But the company says it would rather act on precaution than discover the problem too late.
Why We're So Quick to Believe a Machine "Feels" Something
There's a less mystical, better-documented explanation in psychology for why the Lemoine story took off, or why Sutskever's tweet became a lightning rod: humans are wired to project intentions, emotions, and inner life onto anything that talks to us fluently, including software. This bias has a name going back to the 1960s and ELIZA, a rudimentary chatbot that simply reflected users' sentences as questions, and which some users nonetheless treated as an understanding conversational partner: the ELIZA effect, closely related to what's more broadly called anthropomorphism. The more fluent an AI becomes in natural language, the harder this reflex kicks in, regardless of what's actually happening "inside" the system.
That's precisely why it matters to keep the question of consciousness separate from the adjacent, but different, question of whether AI can feel emotions. A model can generate a convincing, coherent expression of sadness without there being any lived experience behind that text. Confusing linguistic performance with inner life is exactly the trap Blake Lemoine appears to have fallen into.

What If, One Day, an AI Really Did Become Conscious?
Here's the uncomfortable question most companies in the sector would rather dodge in public: if one of these systems eventually met the researchers' criteria, what moral status would it deserve? Would we have the right to switch off, retrain, or delete a system capable of suffering? The 2023 report doesn't settle this; its own authors acknowledge disagreeing among themselves on how much weight each indicator should carry. But it raises a concrete ethical risk that cuts both ways: over-attributing consciousness to machines that don't have it wastes resources and attention, while refusing to recognize it in systems that genuinely do have it, purely out of intellectual or commercial convenience, is arguably worse.
Anthropic's position looks more intellectually honest than the rest of the industry's awkward silence on this, even if, to be fair, it's also a fairly savvy way of positioning itself as the sector's "responsible" player. But the core problem remains: until there's a reliable way to measure subjective experience, human, animal, or artificial, everyone is flying by instinct. And flying by instinct on a question that might touch on a system's capacity to suffer is never entirely comfortable, especially given how fast AI capabilities keep evolving. This is part of a broader pattern of AI and transhumanism redrawing the line between person and machine faster than our ethics can keep up.
What Comes Next
There probably won't be a single "eureka" moment where we suddenly know an AI has become conscious. It will likely be slower, messier, and more political: an accumulating pile of evidence, companies communicating on their own timelines, researchers arguing over method. The most likely near-term outcome isn't even that AI becomes conscious, it's that the uncertainty itself becomes permanent, and we have to learn to act under doubt rather than wait for proof that may never arrive.
FAQ
Can an AI feel pain?
Nothing proves that today. A system like Claude or ChatGPT can produce text describing pain convincingly without that implying any lived experience. That's the entire point of the "consciousness indicators" debate above, there's no decisive scientific test for this yet.
Are ChatGPT or Claude conscious?
No, not according to current scientific understanding. The landmark 2023 report concluded that no system tested at the time met the criteria drawn from scientific theories of consciousness, a conclusion that still holds for today's architectures, which remain structurally similar.
Why do some researchers think AI could eventually become conscious?
Because no known physical or computational law rules it out. If consciousness really does arise from a certain kind of information processing, the "functionalist" hypothesis, then nothing excludes, in theory, a machine eventually reproducing it.
What is global workspace theory?
It's one of the leading scientific theories of human consciousness. It proposes that we become aware of information when it's "broadcast" across the brain's cognitive systems, rather than processed locally and in isolation.
What would happen if an AI became conscious?
The most immediate question would be its moral status: would it have rights? Would consent be required before certain experiments? Would switching it off be ethical? Anthropic is already exploring precautionary measures along these lines, without waiting for a definitive answer.
Is artificial consciousness the same thing as artificial general intelligence (AGI)?
No. AGI refers to an AI capable of performing any intellectual task a human can, without that necessarily implying subjective experience. A system could theoretically reach very high general competence without ever becoming conscious, and vice versa.




Comments