Could an AI Actually Be Conscious? What Would Count as Evidence?

Suppose an AI tells you:

I am conscious.

What happens next?

Do you believe it?

Ask harder questions?

Inspect the code?

Look for brain-like processing?

Wait for it to beg not to be deleted?

The strange thing is that none of these steps gives us a simple answer.

Machine consciousness is difficult for the same reason animal consciousness, infant consciousness and disorders of consciousness are difficult:

experience is private.

The problem of evidence

You know your own consciousness directly.

You infer everyone else's.

Another human says they are in pain, and you believe them partly because:
- they behave like you;
- they have a nervous system like yours;
- they share your biological history;
- similar injuries cause similar responses.

With AI, those bridges weaken.

The system can speak like a person while being built very differently.

That means the usual shortcut from language to mind becomes less reliable.

One answer: derive indicators from consciousness theories

Patrick Butlin and a large interdisciplinary group proposed a systematic strategy.

Do not ask whether an AI "seems conscious."

Instead:
1. take leading scientific theories of consciousness;
2. identify the properties those theories associate with consciousness;
3. translate those properties into computational terms;
4. inspect AI systems for them. [S16]

This does not produce certainty.

It produces structured evidence.

A 2026 Trends in Cognitive Sciences paper develops the same indicator strategy further. [S17]

What kinds of indicators?

DarkBrain already has separate articles on the major theories, so we only need the machine-consciousness version here.

Recurrent processing Does information loop back through the system rather than only flowing forward once?

Some theories treat recurrent processing as important for conscious perception.

Global workspace Can information become globally available across multiple specialized processes?

The analogy is a stage:
many processes occur backstage, but selected information becomes widely broadcast.

Higher-order representation Can the system represent not only the world, but its own mental or informational states?

Predictive processing Does the system continuously generate and update predictions about incoming information?

Attention schema Does it build a simplified model of its own attention and use that model to control behavior?

None of these properties is a magic switch.

They are candidate clues.

Why current language models are difficult to classify

David Chalmers argues that current large language models have impressive capacities but also major obstacles under several mainstream theories of consciousness.

He highlights issues such as:
- limited recurrent processing;
- absence of a strong global workspace;
- limited unified agency. [S15]

But he does not conclude that machine consciousness is impossible.

His argument is almost the opposite.

If future systems gain richer memory, persistent goals, recurrent architectures, self-models, embodiment and integrated agency, the question becomes harder to dismiss.

Current systems are not established as conscious

The most careful current position is not:

AI is conscious.

It is also not:

AI can never be conscious.

The evidence does not justify either certainty.

The 2023 indicator report did not find a strong case that the systems it assessed were conscious. [S16]

A 2025 review similarly describes machine consciousness as unresolved and current machines as far from established consciousness. [S21]

That is an evidence statement.

Not a metaphysical verdict.

The substrate question

Here the debate becomes deeper.

Suppose you perfectly reproduce the relevant computation of a conscious brain.

Would the digital system be conscious?

One family of views says yes:
the right functional organization is what matters.

Another says perhaps not:
the physical and causal properties of biological systems may be essential.

Wanja Wiese argues that a simulation may fail to replicate consciousness if it lacks the right causal flow. [S19]

Farisco, Evers and Changeux emphasize evolved biological features while leaving open the possibility of alternative artificial consciousness. [S20]

A 2026 Trends article puts the conflict bluntly:
are only biological "meat machines" conscious, or can the right computation work in another substrate? [S22]

There is no consensus.

A machine can pass one test and fail another

This is why one grand "consciousness test" is unlikely to be enough.

A system might:
- pass conversational tests;
- have recurrent processing;
- model itself;
- lack stable agency;
- have no body;
- show global broadcasting;
- lack biologically plausible causal dynamics.

What should we conclude?

The likely future is not a binary switch.

It is a growing evidence profile.

The danger of both errors

Two mistakes are possible.

False positive We treat a nonconscious system as sentient because its language manipulates our social instincts.

False negative A genuinely conscious artificial system appears, and we dismiss its experience because "it is only code."

The 2026 indicator literature explicitly takes both possibilities seriously. [S17]

This is why machine-consciousness research is not merely science fiction.

It is a problem of inference under uncertainty.

DarkBrain assessment framework

Ask four questions.

Architecture Does the system contain mechanisms predicted by consciousness theories?

Integration Are those mechanisms connected into a unified cognitive system?

Agency Does the system maintain goals, memory and self-directed control across time?

Report Does it describe internal states in ways that track real internal processes rather than merely generating plausible language?

Even all four would not prove consciousness.

But they would make the question more serious.