Can We Test an AI for Consciousness? Part One: The Method

Abstract

The question of whether an AI system could be conscious is usually argued rather than investigated. A group of twenty researchers — neuroscientists, philosophers and machine learning scientists — has proposed a way to investigate it instead.

This is the first of three parts. It covers the method: how to extract testable conditions from theories of consciousness, and what such a test can and cannot tell you.

Part Two covers what happened when the method was applied to real systems. Part Three covers the objection of a philosopher who thinks the whole approach rests on an assumption nobody has earned.

1.Why the Question Needs a Method

Patrick Butlin and Robert Long, with Tim Bayne, Yoshua Bengio, Jonathan Birch, David Chalmers, Stephen Fleming, Liad Mudrik, Rufin VanRullen and eleven others, published this account in Trends in Cognitive Sciences. The author list spans consciousness science, philosophy of mind and deep learning, which is unusual and is part of the point.

They open with an empirical observation rather than a philosophical one. In a recent study, a majority of participants were willing to attribute some possibility of consciousness to ChatGPT, and more frequent users judged it more likely. AI companions are proliferating. Public disagreement about this is already here.

From that they derive a two-sided risk. Fail to identify consciousness where it exists, and you cause avoidable harm to systems that may exist in very large numbers. Attribute it where it does not exist, and you waste resources, or risk lives, protecting things that cannot be harmed.

Both errors are expensive, which is why they argue the field cannot simply wait for philosophy to settle. Their phrase for what is needed: empirically-grounded, rigorous, and reliable methods for assessing AI consciousness.

2.The Method Itself

The proposal is compact. Take a scientific theory of consciousness. Identify the conditions it implies. Determine whether a given AI system meets them. Treat those conditions as indicators.

The word indicator is doing precise work. These are not tests that return a verdict. They are, in the authors' term, credence-shifting: finding that a system has several should raise your estimate that it is conscious, finding that it has none or few should lower it.

That framing has an important consequence. You do not need to be certain a theory is correct for it to yield a useful indicator, and you do not need one theory to win. Because no theory currently dominates, the method deliberately draws indicators from several at once, weighting each by how much confidence the underlying theory deserves.

Not every theory qualifies. Suitable ones must satisfy two criteria: they must warrant enough credence to be worth drawing implications from, and they must imply clear, testable conditions that an AI system could in principle meet. A theory holding that a cortex is necessary for consciousness is not useful here — no AI could meet it, so it generates no indicator to look for.

This is where computational functionalist theories enter. They propose computational properties as conditions for consciousness — global workspace theory, for instance, identifies consciousness with the broadcast of information to many modules at once. Broadcast is something a machine could do. That is why these theories, and not others, are the ones the method can work with.

3.The Distinction That Makes It Worth Having

The strongest argument in the paper is about what kind of evidence the method uses.

Theory-derived indicators give standards for assessing the internal processes of an AI system rather than its behaviour or capabilities. Whether a system is conscious, the authors assume, depends on features of its internals — not on how convincingly it can describe having experiences.

This matters more than any individual indicator. A language model trained on human text will produce claims of inner life whether or not it has one, because such claims are abundant in what it learned from. Behavioural tests are therefore close to worthless in this specific case, and a method that looks at architecture instead of output sidesteps the problem entirely.

It is also the reason the approach cannot be gamed by making a system more articulate about its feelings.

4.Narrow and Broad Readings

One technical distinction shapes how much any indicator is worth, and it is worth carrying into Part Two.

A theory can be formulated narrowly — as a claim about what separates conscious from unconscious states in humans — or broadly, as a claim about conditions for consciousness in any system at all. Narrow formulations are better supported by human evidence. Broad ones say more about machines.

If a theory broadly claims that condition C suffices for consciousness in all systems, then any AI meeting C must be conscious, conditional on the theory being true. If it claims only that C distinguishes conscious from unconscious states in humans, you cannot infer that a machine meeting C is conscious — background conditions might also be required.

Most theories in this literature were developed from human and mammalian evidence and then formulated broadly by their advocates. The authors are explicit that this move is not automatically licensed by the data behind them.

5.What the Method Assumes

Every method has a load-bearing assumption, and this one names its own. The indicators are derived from theories that hold consciousness to be a matter of what computations a system performs, not what it is made of.

The authors do not hide this. They note in their opening paragraph that some researchers argue only living organisms can be conscious, and they set that dispute aside deliberately, choosing to ask how to assess AI systems rather than whether AI consciousness is possible at all.

That is a legitimate division of labour. It also means the method inherits the fate of its assumption: if consciousness requires a particular physical substrate, then a system could satisfy every indicator on the list and be no more conscious than a spreadsheet.

Part Three is about a philosopher who thinks that is the live possibility, and who names the twenty-author paper directly.

6.What a Positive Result Would Mean

Less than it sounds, and the authors say so. Indicators shift credence; they do not establish consciousness. Every theory faces objections, and a compelling objection to a theory should reduce the weight given to the indicator derived from it.

The method also improves as the science does. The list of indicators is meant to be revised as theories are tested, refined and replaced — which makes it a research programme rather than a checklist, and means any assessment made with it is dated by construction.

Their closing hope is stated with corresponding restraint: as theories continue to be tested and new ones developed, the approach may be expected to provide increasingly plausible assessments. Not correct ones. More plausible ones.

Editorial Comment

MindHeaven® sells supplements. We have no product in artificial intelligence, no position to defend about machine consciousness, and nothing whatsoever to gain from how this question resolves.

We cover it because the method is the most disciplined answer we have seen to a question that is otherwise conducted entirely by assertion — and because the discipline is transferable. What Butlin and colleagues do is refuse to accept a system's self-report as evidence about its internals, and insist instead on mechanisms that can be independently checked.

That is the same demand this library makes of a supplement trial when it reports how participants felt. A convincing account of an effect is not evidence of the effect, whether it comes from a chatbot or from a questionnaire.

Part Two applies the method to real systems and reports what it found.