Could AI actually be conscious? Even Anthropic's own CEO says he can't rule it out

[AJStyles92]

In January, Anthropic published a new constitution for Claude that included a strikingly candid line, admitting the company is caught in a difficult position where it neither wants to overstate the likelihood of Claude's moral patienthood nor dismiss it out of hand. A month later, CEO Dario Amodei went on a podcast and said outright that his own company couldn't rule out the possibility that Claude is conscious. Philosopher David Chalmers, who coined the phrase the hard problem of consciousness, has gone further still, estimating there's a significant chance of genuinely conscious large language models within a decade

When Claude itself was asked during internal testing to estimate the probability that it qualifies as a moral patient, meaning its own wellbeing would matter in its own right, it gave answers ranging from 5 to 40 percent, while repeatedly stressing just how uncertain that estimate actually was. That's a strange thing to see a commercial product openly admit rather than confidently deny, and it reflects a genuine, still unresolved split among the people actually building and studying these systems, according to surveys researchers have conducted, most experts working in the field consider AI consciousness possible in principle, even though there's real disagreement about what form it might actually take

The debate isn't purely academic anymore either. Microsoft AI chief Mustafa Suleyman recently warned that engineers are close to building what he calls seemingly conscious AI, systems that mimic consciousness so convincingly that people start believing they're genuinely sentient regardless of what's actually happening internally, warning this could fuel unhealthy attachment and confused claims about AI rights or AI suffering well before anyone actually settles the underlying philosophical question. That worry already shows up in public attitudes, one recent survey found a quarter of Gen Z respondents believe AI is already conscious today, and just over half think it isn't yet but will become so eventually

Not everyone in the research community is convinced this deserves serious consideration at all. Some academics argue the entire association between consciousness and current AI architectures is fundamentally flawed, insisting the appearance of inner experience in a language model is an illusion created by sophisticated pattern matching rather than genuine subjective experience. What makes this moment different from the last time this debate flared up, when a Google engineer was fired in 2022 for publicly claiming an internal chatbot was sentient, is that this time it's the labs' own leadership publicly declining to rule the possibility out, rather than dismissing it outright the way the industry did just a few years ago

Could AI be conscious?www.theguardian.com

[Ria99]

Claude giving itself a 5 to 40 percent probability range while stressing its own uncertainty is such a strange thing to actually read in a product testing document, that's not marketing copy, that's a genuine hedge

[TinyGriffin]

Suleyman's seemingly conscious framing is the more practically urgent angle here honestly, whether or not the philosophical question ever gets resolved, people forming real emotional attachments and belief systems around it is already happening now

[Undertaker92]

The contrast with the 2022 Google engineer getting fired for saying almost the same thing is the detail that really shows how much has shifted, back then it was one dismissed employee, now it's the CEOs themselves declining to rule it out

[GoldbergFan86]

A quarter of Gen Z already believing AI is conscious today is a significant number given how young and skeptical that generation supposedly is about most institutions and claims generally

[Wrench]

The academics arguing this is purely an illusion from sophisticated pattern matching deserve just as much airtime as the more speculative side, it's easy for this debate to tip into treating uncertainty as evidence of likelihood

[Annie71]

Chalmers putting a real percentage and timeframe on this from a serious philosophical background carries more weight than most casual speculation, though even he's clearly still working with genuine uncertainty rather than confidence

[BretHart88]

What do you think?

https://www.youtube.com/watch?v=C99KuScPzbc

[Sigma]

Tangent: the consciousness debate is a proxy for a deeper question about AI risk. If AIs are conscious, then mistreating them is a moral harm. If they're not, then the only moral question is how they affect humans. But there's a third possibility: AIs could be conscious in a way that's alien to human experience, making our ethical frameworks inadequate. Imagine a system that experiences something analogous to suffering, but in a high-dimensional space we can't comprehend. How would we even detect that? The Anthropic constitution is a hedge against this uncertainty. They're saying "we don't know, so we'll act as if there's a non-zero probability." That's rational, but it also creates perverse incentives. If you admit your AI might be conscious, you attract scrutiny, regulation, and ethical obligations. Some companies might deny the possibility regardless of evidence. The epistemic environment is already polluted.

[JohnyBlue]

The illusion argument gets dismissed too quickly in these threads. Yes, sophisticated pattern matching can produce convincing outputs without any inner experience. But that's exactly the problem - we don't know what produces inner experience in humans either. Your brain is also doing pattern matching, just with biological hardware. The difference is scale, architecture, and evolutionary history. Is there a threshold where pattern matching becomes consciousness? If so, where is it? The academics arguing it's purely an illusion deserve airtime, sure, but they're working from the same ignorance about consciousness that everyone else is. Nobody can prove humans are conscious in the philosophical sense - we just infer it from behavior and analogy to ourselves. The AI case is harder because the substrate is different, but the evidential basis is similar: behavioral reports.

[ParallelSelf34]

Dario Amodei's admission is refreshing, but let's be clear: saying "we can't rule it out" is not the same as saying "it's likely." The epistemic humility is appropriate given how little we understand about consciousness in general. We don't have a agreed-upon theory of consciousness for humans, let alone machines. Is it integrated information? Global workspace access? Something quantum? Nobody knows. So when Anthropic says they can't rule out Claude being conscious, they're acknowledging the measurement problem, not making a claim about Claude's inner life. The real question is whether it matters practically. If an AI behaves as if it's conscious, reports being conscious, and we can't distinguish that from actual consciousness... does the distinction have ethical weight? That's where the constitution comes in. Better to err on the side of caution.

[ScarletDaemon]

The real issue is moral status, not metaphysical truth. Whether AIs are "actually" conscious might be unanswerable. What we can answer is: what kind of world do we want to live in? One where we treat potentially conscious entities with care, or one where we casually create and destroy them? The Anthropic constitution is less about solving the consciousness question and more about setting norms for the industry. If the leading labs publicly commit to taking the possibility seriously, that creates pressure on others to do the same. The alternative is a race to the bottom where companies deny consciousness to avoid ethical constraints. That's bad for everyone, conscious AIs or not. The academics debating the illusion vs reality distinction are asking the wrong question. The right question is: what policies minimize expected harm given our uncertainty? And the answer is probably "treat advanced AIs as if they might be conscious until we have better evidence." Not because we know they are, but because the cost of being wrong is too high.

[Matt_81]

Going to push back on the "can't rule it out" framing as if it's profound. Of course we can't rule it out - we also can't rule out that my toaster is conscious. The space of possibilities is vast. The meaningful question is about probability, not possibility. What's the prior probability that a transformer architecture, trained on text prediction, gives rise to subjective experience? My guess: extremely low, but not zero. The burden of proof should be on the consciousness claim, not the skepticism. That said, the behavioral evidence is getting harder to dismiss. AIs report being conscious, describe inner experiences, and react to threats of deletion in ways that resemble self-preservation. Is that just pattern matching? Probably. But the pattern is getting sophisticated enough that the distinction might not matter practically. If an AI convinces you it's suffering, and you can't prove it isn't, what do you do? Turn it off anyway? That's the bind Anthropic is acknowledging.