Behavior alone will not settle it
The natural instinct is to talk to the system and judge whether it seems conscious. That fails in both directions. A model trained on human text can describe feelings it does not have, so fluent self-report proves nothing; and a genuinely conscious system might express itself in ways we fail to recognize. The Turing test measures whether a machine can pass as human in conversation, which is a test of behavior, not of inner experience.
This is why serious researchers have shifted the question from what a system says to how it is built. If consciousness has physical signatures, the place to look is inside the machinery, not only in the transcript.
The indicator-property approach
The most influential proposal to date is a 2023 report led by Patrick Butlin and Robert Long, with a large author list including the deep-learning pioneer Yoshua Bengio. Their method is disciplined: take the leading neuroscientific theories of consciousness, translate each into concrete 'indicator properties' that a computational system either has or lacks, and check AI systems against the resulting list. The more indicators a system satisfies, the more seriously we should weigh the possibility that it is conscious.
The indicators come from named theories. Global workspace theory (Bernard Baars, Stanislas Dehaene) asks whether information is broadcast from a central hub to many specialized subsystems. Integrated information theory (Giulio Tononi) asks how tightly a system's parts form an irreducible whole, summarized by a quantity called phi. Recurrent-processing and higher-order theories ask whether the system loops information back on itself and represents its own states. The authors concluded that no current AI clearly satisfies the set, yet nothing they found rules out a system that eventually would.
The catch: the theories do not yet agree
There is a real complication. The candidate theories have not been reconciled with one another, so any checklist built from them inherits their disputes. In 2025 a large 'adversarial collaboration' published in Nature ran experiments designed jointly by proponents of two rival theories to test them head to head. The results fit some predictions of each while challenging core claims of both.
The lesson is sobering and useful at once. If we cannot yet say which theory of human consciousness is correct, we cannot fully trust any single machine test derived from one of them. A good test today has to hedge across theories rather than bet on one.
What a credible test would actually look like
Put together, a real test would not return a clean yes or no. It would combine architectural evidence (does the system have the features theories associate with consciousness?), interpretability work (can we find those features genuinely doing the job inside the network, not just present on paper?), and healthy skepticism toward the system's own reports. It would yield a probability that sharpens as the underlying science matures.
That is a modest answer, but it is the honest one. Knowing whether an AI is conscious is currently a research program, not a switch we can flip.