Home / Tests and Evidence

How to Test If AI Is Conscious: The Indicator Checklist

A row of seed pods on a wall, some glowing and some dark

How do you test if an AI is conscious? There is no single test that can settle it, because nobody can observe consciousness directly, even in other people. In 2026 researchers use three tools together: a checklist of indicators drawn from theories of consciousness, a way of turning those indicators into probabilities, and experiments that look inside the models instead of trusting what they say.

The indicator method was published in its current form in Trends in Cognitive Sciences in 2026, in Identifying indicators of consciousness in AI systems by Patrick Butlin, Robert Long and 18 co-authors. In September 2026 a second team, including Anil Seth, Murray Shanahan and Shane Legg, proposed a framework that combines such indicators into a single probability.

How to test if AI is conscious with indicators

The approach began with a 2023 report, Consciousness in Artificial Intelligence. Its 19 authors turned what leading scientific theories say consciousness requires into "indicator properties". The indicators came from five families of theory:

  • Recurrent processing theory: signals must loop back, not only pass forwards.
  • Global workspace theory: information must be broadcast to many processes at once, as explained in global workspace theory and AI.
  • Higher-order theories: a system must in some sense represent its own mental states.
  • Predictive processing: the system models its inputs and corrects its predictions.
  • Attention schema theory: the system keeps a model of its own attention.

The 2026 paper develops that method. It argues the approach works because some influential theories have consequences for AI that can be tested, and it treats the indicators as evidence that moves the odds rather than as a pass mark. Co-author Colin Klein summarised two points in June 2026: no existing AI system, ChatGPT included, is conscious, and the useful indicators are all structural, about how a system is built rather than how it behaves.

A lens revealing glowing veins inside a fallen leaf

From checklist to probability

The September 2026 framework, From cacophony to hierarchy, adds a step. It sorts theories by the level of description they treat as critical, from behaviour and computation through a system's physical causal structure to the whole organism in its environment, develops indicators for each level, and then combines trust in each theory with the evidence to give an overall probability.

Applied to current language models, its illustrative results ranged from below 1% to about 80% depending on the assumptions. The authors warn that this is not a measured range. Its lesson is that a test result means little until you say which theories you are relying on, a practice they call structured agnosticism.

Why asking the AI does not work

The obvious test, asking, is the weakest. Language models are trained on vast amounts of human writing about thoughts and feelings, so they can describe experience fluently whether or not they have any. 80,000 Hours makes the point in its review of the field: AI self-reports cannot be trusted, and behavioural, theory-based and architectural tests all have serious drawbacks. What chatbots do say about themselves is covered in is ChatGPT conscious.

A closer look: testing inside the model

Other tests examine the model's internal activity:

  • Introspection. In October 2025 Anthropic planted the pattern for a known concept in Claude's activity and asked whether it noticed. In those experiments Claude Opus 4.1 did so only about 20% of the time with the best method.
  • A workspace. In July 2026 Anthropic researchers found a workspace-like structure in Claude Sonnet 4.5 that matches many functional properties the global workspace indicator asks for. Removing it sharply weakened multi-step reasoning and translation.
  • Emotion representations. In April 2026 Anthropic reported emotion representations in the same model that change its behaviour.

Each team was explicit about the limit: these results describe functions, and none of them tells us whether the model experiences anything.

Checklist: reading a claim that an AI passed a test

  • Which test, from which theory? An indicator only counts if you accept the theory behind it.
  • Was the AI simply asked? Self-reports from models trained on human writing are weak evidence on their own.
  • Is a probability given, and on what assumptions? A number without its theoretical assumptions can mislead.
  • Function or experience? Detecting internal states is not the same as having experiences, as the Chinese room argument warns in a different way.

More in Tests and Evidence · Report a correction