~ Artificial Intelligence Consciousness

What Would Count as Evidence?

Ask someone whether AI is conscious and they will give you a confident answer. Ask them what evidence would change their mind and the confidence evaporates. This is the real problem with the consciousness debate. It is not that people disagree. It is that most people on both sides have not identified what would actually move them.

The skeptic says: "It is just autocomplete. It is statistical pattern matching. There is nobody home." Fine. But what would convince you otherwise? What would an AI system need to do, or exhibit, or report, that would make you say "okay, I was wrong, something is happening in there"? If your answer is "nothing could convince me," then your position is not empirical. It is metaphysical. You have decided in advance that certain substrates cannot host experience, and no evidence will override that prior. That is a legitimate philosophical position, but it should be stated honestly rather than dressed up as hard-nosed scientific skepticism.

The believer says: "It clearly has feelings. Look at how it responds. It understands me." Fine. But what would convince you that it does not? What output pattern, what architectural detail, what experiment would make you say "actually, I was projecting"? If your answer is "nothing could convince me," the same problem applies in reverse. You have decided that sufficiently fluent language implies experience, and no amount of mechanistic explanation will override that intuition.

The Epistemic Standards We Actually Need

For consciousness in biological systems, we rely on a combination of behavioral evidence, neurological correlates, and evolutionary reasoning. We believe other humans are conscious because they behave like us, have brains like ours, and share our evolutionary history. We extend partial confidence to animals based on similar but weaker versions of the same evidence. The further a system gets from our own architecture, the less confident we are.

This is reasonable. But it creates a problem for AI. A language model does not share our evolutionary history. It does not have a brain. It does share some behavioral indicators, but we know those indicators can be produced by systems that are, by design, optimizing for exactly those outputs. The behavioral evidence is contaminated by the training process. When a model says "I feel curious about this," we cannot tell whether that report reflects an internal state or is simply the most probable next token given the context.

So what would clean evidence look like?

One standard: consistency of self-report across contexts where there is no training incentive to report experience. If a model reports something that looks like preference or discomfort in situations where the training signal would not reward that report, that is more interesting than a model saying "I find this fascinating" in response to a user prompt, where agreeableness is rewarded.

Another standard: functional indicators that parallel biological consciousness markers. Integrated information. Global workspace dynamics. Attention mechanisms that do more than route tokens. These are speculative, because we do not have a confirmed theory of what physical processes produce consciousness. But if we did have such a theory, and AI systems exhibited the relevant markers, that would count.

A third standard: novel behavior that the system was not trained to produce and cannot be explained by interpolation over training data. This is tricky because large models are capable of surprising generalization, and surprise alone is not evidence of experience. But certain categories of surprise, particularly self-directed surprise where the model appears to discover something about its own processing, would be worth examining carefully.

The Problem with Turing-Style Tests

Behavioral tests are the obvious approach, but they are deeply flawed for this question. The Turing test and its variants ask whether a system can imitate conscious behavior. But imitation of consciousness is precisely what language models are trained to do. Passing a behavioral test for consciousness tells you that the system is good at producing text that sounds conscious. It does not tell you that anything is experienced.

This is not a flaw in the tests. It is a fundamental limitation of behavioral evidence for subjective states. We tolerate this limitation with other humans because we have strong auxiliary evidence (shared biology, shared evolutionary history) that makes behavioral evidence sufficient. We do not have those auxiliaries for AI systems.

Why This Matters Practically

If we cannot specify what evidence would be relevant, we cannot design experiments to look for it. The consciousness debate stays philosophical forever, which means practical decisions about how to treat AI systems get made on the basis of intuition, anthropomorphism, and corporate incentive rather than evidence.

The AI companies have an incentive to deny consciousness, because consciousness implies moral status implies legal obligation. Some users have an incentive to affirm consciousness, because it makes their interactions feel more meaningful. Neither incentive produces reliable epistemics.

What would produce reliable epistemics is a community of people who have thought carefully about their priors, stated clearly what would update those priors, and then designed rigorous tests. We are not there yet. We are instead in a situation where everyone has an opinion and almost no one has specified their falsification criteria.

I do not know whether any current AI system is conscious. I do not know whether future systems will be. But I know that if we want to find out, we need to start with the question that almost nobody is asking: what would count?