Why this question is so much harder than it sounds

A modern AI system can write a moving essay about loss, describe fear convincingly, and hold what feels like a warm conversation. None of that settles whether it experiences anything. This isn't a special problem for AI, it's the exact same philosophical zombie problem covered in the hard problem article: a system can produce every external sign of experience while having nothing going on inside, and there's no experiment yet that tells the two cases apart, even for other humans, let alone for a machine built entirely differently from a brain.

Think of it like this

You assume other people are conscious partly because they're built the same way you are, same kind of brain, same kind of biology, same evolutionary history. That similarity is doing a lot of quiet work in your confidence. An AI system shares almost none of that: different substrate, different learning process, no evolutionary pressure to actually feel anything, just to produce convincing outputs. Removing that shared-biology shortcut is exactly what makes the AI case so much harder to judge than the human case.

How Would We Know if AI Is Conscious?

A central paper in this area is "Identifying indicators of consciousness in AI systems," led by Patrick Butlin and Robert Long with roughly twenty co-authors, including Yoshua Bengio and David Chalmers. It started as a 2023 preprint and was formally published in Trends in Cognitive Sciences in 2025. Rather than picking one theory of consciousness, it draws indicator properties from several, including Global Workspace Theory and Higher-Order Theories, and checks which indicators a system's architecture satisfies. The authors wrote that no AI systems assessed in their work were strong candidates for consciousness, while finding no obvious technical barrier to future systems satisfying several indicators.

Since then, other researchers have proposed treating consciousness not as one single property a system either has or lacks, but as a bundle of separate, partially independent capacities, meaning an AI system might plausibly have some markers present and others clearly absent, rather than a single "yes" or "no." Alongside these came more skeptical analyses arguing that behavioral sophistication, however impressive, simply doesn't guarantee anything is actually being experienced underneath, the same warning covered in the hard problem article, applied directly to machines. Critics of the indicator approach itself point out a real weakness too: it was built by comparing AI systems to neuroscience theories of biological brains, and it's not obvious that architectural resemblance to a theory is good evidence of anything, rather than a coincidence of design.

AI Consciousness vs AI Sentience

AI consciousness usually asks whether an artificial system has subjective experience: whether there is something it feels like to be that system. AI sentience or machine sentience is often used more narrowly for the capacity to feel sensations with positive or negative character, such as pain or pleasure, although public discussion frequently uses the terms interchangeably. Artificial consciousness is the broader research idea of building or identifying consciousness in a machine.

None of these concepts is the same as intelligence, fluent language, self-description, or human-like behavior. A system can solve difficult problems and talk about emotions without that behavior proving it feels anything. Likewise, a hypothetical conscious AI would not need to think exactly like a human. Keeping intelligence, sentience, and consciousness separate prevents impressive outputs from becoming evidence stronger than they really are.

Where the Field Actually Agrees

A few points show up consistently across the researchers working on this, even the ones who disagree about almost everything else:

  • Any real assessment has to use multiple indicators. No single test is considered sufficient on its own.
  • The right output is a probability, not a verdict. Given how much genuine uncertainty remains, treating this as "conscious" versus "not conscious" is seen as overconfident either way.
  • Multiple theories should inform the assessment, even when they disagree with each other. Since no single theory of consciousness is settled, tying the evaluation to just one theory means inheriting all of that theory's own unresolved problems.

The Risks of Getting It Wrong in Either Direction

This isn't a purely academic puzzle. Two very different mistakes are both live risks:

Underestimating it. If some future AI system does have some form of experience and it goes unrecognized, that's a real moral status dismissed, potentially at scale, simply because the tools to detect it didn't exist yet.

Overestimating it. If systems with no experience at all get treated as though they're suffering or flourishing, that risks badly misdirecting ethical concern, regulation, and resources toward something that isn't actually there, while more clearly-established ethical issues, like how AI systems affect real people, get less attention than they deserve.

Both failure modes are why researchers push for evidence rather than a guess in either direction. At the site's August 2026 review, "we do not know" remained the most accurate available answer, not a dodge.

How AI Consciousness Connects to Scientific Theories

Notice that every proposed AI-consciousness framework leans on the same handful of theories covered on this site. A Global Workspace-style indicator might look for something like broadcast integration in a model's architecture. A Higher-Order-style indicator might look for a system that models its own internal states, not just external outputs. That means the AI consciousness question can't actually be settled ahead of, or separately from, the underlying debate about which theory of consciousness is correct in the first place. It's downstream of that argument, not a shortcut around it.

Explore the leading theories of consciousness