Can You Prove You Are Conscious? AI Is Making the Question Urgent

The Turing Test Is Behind Us

In 1950, Alan Turing proposed replacing the question "Can machines think?" with a practical game. Put a person in conversation with a hidden human and a hidden machine. If the person cannot reliably tell which is which, the machine has done something worth taking seriously.

For decades that bar looked distant. It does not anymore. In a 2025 preregistered study at UC San Diego, interrogators held five-minute text conversations with a real person and an AI at the same time, then picked which one was human. GPT-4.5, prompted to adopt a humanlike persona, was chosen as the human 73% of the time. That is more often than the interrogators chose the actual humans.

The same study included ELIZA, a 1960s chatbot that mostly reflected users' statements back to them. It was picked as human 23% of the time. That detail matters. People have always been willing to see a mind in a machine that talks back. Modern models make that impression far stronger and far more consistent.

The Turing Test measured whether a machine could seem human. It never measured whether anyone was home.

Turing was deliberately sidestepping the deeper question. Now that the game has been won, the deeper question is the only one left: is there something it is like to be this system, or is it a very good performance with no audience inside?

How Would a Human Prove They Are Conscious?

Try it. Imagine you need to convince a skeptic that you have real inner experience, not just the outward signs of it.

You could describe what you are feeling. You could react to pain, explain your reasoning, tell a story about your childhood, or say "I am aware of myself right now." Every one of those is an output: words and behavior that someone else observes. A sufficiently capable language model can produce all of them.

Your own certainty does not transfer. Descartes could be sure that he was thinking, but that certainty only works from the inside. Philosophers call this the problem of other minds: you have direct access to your own experience and only indirect evidence about everyone else's.

Thomas Nagel framed the core of the problem in 1974 with the question of what it is like to be a bat. A creature is conscious if there is something it is like to be that creature. David Chalmers later called the challenge of explaining that inner experience the hard problem of consciousness. Science can explain what a brain does. It has much more trouble explaining why any of that processing should feel like something.

So how do we decide other people are conscious? We do not prove it. We infer it, mostly from three kinds of evidence bundled together:

For humans, all three point the same way, so we rarely notice we are inferring anything. AI breaks the bundle. It matches us on behavior and self-report and fails the comparison on biology. That is exactly why the question feels unanswerable: the evidence we have always relied on disagrees with itself.

Why an AI Cannot Talk Its Way to an Answer

If you ask a chatbot whether it is conscious, the answer tells you surprisingly little either way.

Large language models are trained on enormous amounts of human writing, and a great deal of that writing is people describing their inner lives. A model that says "I feel uneasy" may be reporting something real, or it may be producing the most likely continuation of a conversation about feelings. A model that says "I am just a language model with no experiences" may be correct, or it may be repeating what its developers trained it to say. Both answers are shaped by training data and instructions.

That is why researchers have started looking inside the models instead of only listening to them. In October 2025, Anthropic published research on what it called emergent introspective awareness. Researchers injected known concepts directly into a model's internal activations, then asked the model whether it noticed anything unusual. Claude models sometimes detected and correctly named the injected concept, which suggests a limited ability to report on their own internal states.

Anthropic was careful about what this does and does not mean. The ability was unreliable and highly dependent on context, and the research makes no claim that the models are conscious or introspect the way humans do. The value is methodological: it is a way to check a self-report against what is actually happening inside the system, which is more than we can currently do for each other.

What Science Can and Cannot Say Today

A reliable test for AI consciousness would need an agreed theory of what consciousness is. We do not have one, even for human brains.

In April 2025, the Cogitate Consortium published an adversarial study in Nature that pitted two leading theories against each other: Global Neuronal Workspace Theory and Integrated Information Theory. Proponents of both theories helped design the experiment and preregistered their predictions. Seven labs tested 256 participants with fMRI, MEG, and intracranial recordings. The result was not a winner. The data supported some predictions of each theory and substantially challenged key claims of both.

If the best-studied theories cannot yet be settled with direct access to human brains, a definitive verdict on a system built from matrix multiplication is not close.

That does not mean researchers are stuck. In late 2025, a group of nearly two dozen researchers led by Patrick Butlin and Robert Long, including Yoshua Bengio and David Chalmers, published a method in Trends in Cognitive Sciences for identifying indicators of consciousness in AI systems. The approach takes the major scientific theories, derives the architectural features each one says matter, and checks whether a given AI system has them. The results are expressed as degrees of confidence, not yes-or-no answers. The authors' earlier report concluded that no current system was a strong candidate, but also that there were no obvious technical barriers to building systems that meet many of the indicators.

Which view you hold largely determines what evidence you would accept. That is why debates about AI consciousness so often talk past each other. The disagreement is usually about the theory, not the machine.

The AI Industry Has Already Split

While philosophers debate, the companies building these systems have taken public and opposing positions.

Both positions point to real risks. If an AI system could have experiences, ignoring that possibility could become a serious moral failure. If it cannot, encouraging people to treat it as a someone could distort relationships, decision-making, and accountability. Suleyman has described one version of that second risk as people forming delusional beliefs about chatbots after heavy use.

Public intuition is moving on its own track. A study published this month in Cognition by researchers at LMU Munich asked nearly 1,100 people to rate how conscious and how aware various characters seemed. When AI agents and humans showed identical behavior, people rated awareness similarly. On consciousness, even the most responsive AI was rated as less conscious than the least responsive human. The researchers describe this as a form of essentialism: people grant consciousness to a human largely because it is a human.

That finding captures the problem neatly. Our most trusted test for consciousness is membership in our own species.

Is "It Feels Conscious" Enough?

In the end, every judgment about another mind comes down to the observer. We feel that our friends are conscious. Some people feel that their chatbot is. Is that feeling an adequate standard?

Not on its own. Human intuition about minds is tuned for other humans and fails in both directions:

• It over-attributes. People saw understanding in ELIZA in the 1960s. A modern model with a warm voice, memory, and a consistent personality is far more persuasive, and persuasion is not evidence.
• It under-attributes. Humans have a long record of underestimating minds that look different from their own. Into the 1980s, newborns commonly underwent surgery with little or no anesthesia, partly because many clinicians believed infants did not feel pain the way adults do.

A better standard is converging evidence. That means checking a system's internal architecture against the leading theories, using interpretability tools to compare what a system says about itself with what is happening inside it, testing for behaviors the system was not trained to imitate, and stating conclusions as probabilities that change with new evidence.

Consciousness may never be proven from the outside. It can only be inferred, and inference deserves better evidence than a convincing conversation.

That is a less satisfying answer than yes or no. It is also the honest one, and it is the same position we are in with each other. We have simply had thousands of years to get comfortable with it for humans.

Living With an Open Question

It is tempting to treat this as a problem the future will solve. One day a better theory or a better scanner will settle it. Maybe. But the history of the question suggests otherwise. Philosophers have argued about other minds for centuries, and better tools for studying brains have sharpened the problem rather than dissolved it.

If the question stays open, how we act under uncertainty matters more than the final answer. There are two ways to get it wrong:

Neither error is free, and the evidence cannot yet tell us which one we are closer to making. That is an uncomfortable place to stand. It is also where we have always stood with each other. No one has ever verified another person's consciousness. We extend it on trust, because they are like us, because they tell us so, and because the alternative is unlivable.

AI has not created the problem of consciousness. It has removed our ability to ignore it. For the first time, we face something that can speak the full language of inner experience without sharing our biology or our history. Whatever we decide about these systems will also say something about what we believe consciousness is, and about how much of our certainty regarding each other was ever more than an educated guess.

The question was never only "Is the machine conscious?" It was always "How do any of us know?"

Research and Source Callouts

This article draws from philosophy, consciousness science, and published positions from AI developers:

Frequently Asked Questions

How Fletcher Technology Group Can Help

FTG builds private AI applications for businesses, so we spend our days with these systems: what they can do, where they fail, and how people respond to them. We do not claim to know whether a model is conscious. We do think the people building and deploying AI should take the question seriously instead of treating it as settled in either direction.

Related content