Can We Test an AI for Consciousness? Part Two: What the Assessment Found

Abstract

Part One covered a method for assessing AI systems for consciousness. This part covers what happened when nineteen researchers actually ran it, on real systems, and wrote down the answer.

The result: no current AI system is conscious, and no obvious technical barrier prevents building one that satisfies the indicators. Those two findings are usually reported separately, by people with opposite intentions.

The most instructive thing in the report is not either finding. It is a footnote in which the authors correct a single sentence of their own abstract for promising more than their evidence supported.

1.What Was Actually Done

The report — "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness" — was released in 2023 by Patrick Butlin and Robert Long with seventeen colleagues, and remains a preprint. Part One's peer-reviewed paper is the method; this is the application.

The preprint status matters and we flag it as we would any other. This document has been enormously influential, is cited in the peer-reviewed literature, and has not itself been through peer review.

They surveyed recurrent processing theory, global workspace theory, computational higher-order theories, predictive processing and attention schema theory. From each they derived indicator properties stated in computational terms — specific enough that you can look at an architecture and check.

The resulting list runs to fourteen properties. It begins with recurrent processing: whether input modules use algorithmic recurrence, and whether they generate organised, integrated perceptual representations. It includes four conditions from global workspace theory, starting with whether the system has multiple specialised modules capable of running in parallel. It ends with agency and embodiment — learning from feedback and selecting outputs to pursue goals.

2.The Theory That Was Left Out

One exclusion shapes everything that follows, and the authors state it in a single sentence: they do not consider integrated information theory, because it is not compatible with computational functionalism.

That is honest and it is also structurally decisive. Integrated information theory is the most prominent theory in consciousness science whose formalism implies that a conventional digital computer cannot be conscious regardless of what it computes — a point we covered separately in our article on IIT 3.0 and 4.0.

So the framework excludes, by construction, the major theory that would answer its central question in the negative. This is not concealed and it is not illegitimate — a method built on functionalist premises cannot use a non-functionalist theory. But it means the assessment is conditional in a way the headline finding does not convey.

Read carefully, the report does not say that no current AI is conscious. It says that on theories compatible with computational functionalism, no current AI is conscious.

3.The Systems, and the Verdict

They assessed Transformer-based large language models and the Perceiver architecture against global workspace theory; DeepMind's Adaptive Agent, a reinforcement learning agent operating in a three-dimensional virtual environment; a system trained to control a virtual rodent body; and PaLM-E, a model connecting language to robotic control.

The choice is revealing. These are not the systems that sound most conscious in conversation. They are the ones whose architectures come closest to the properties the theories point at — which is what Part One's insistence on internals rather than behaviour requires.

Some indicators are already clearly satisfied. Algorithmic recurrence — the first condition on the list — is met by existing systems. Agency, in the sense of learning from feedback and selecting outputs to pursue goals, is arguably met.

The verdict is that no current system possesses enough of them, and that where individual properties are missing there is generally no deep obstacle to building them in. Standard machine learning methods could produce systems with individual properties from the list; what would need experimentation is combining several into one functioning system.

4.The Footnote

The abstract's final sentence originally read that the analysis shows there are no obvious barriers to building conscious AI systems. It now reads that there are no obvious technical barriers to building AI systems which satisfy these indicators.

The change is recorded in a footnote on the first page, with the reason given: to better reflect the messaging of the report — that satisfying these indicators may be feasible, but that satisfying them would not mean such a system would definitely be conscious.

We think this is the most admirable paragraph in the document, and we want to be precise about why.

The original sentence was not false. It was a summary that permitted a stronger reading than the analysis supported, on the single sentence most likely to be quoted without the rest. Someone noticed, and rather than leaving it, they published the correction attached to the claim.

Almost nobody does this. The normal fate of an over-reaching abstract is that it circulates, gets cited, and becomes the thing everyone knows the paper said.

5.What the Answer Is Worth

Stated at its correct width: as of 2023, no AI system examined possessed enough of the indicator properties derived from five functionalist theories of consciousness to be considered a serious candidate — and the properties it lacked appear buildable.

The limits are considerable. The assessment is dated, and AI architectures have moved. It is conditional on functionalism. It excludes the major dissenting theory by design. And indicators shift credence rather than settling anything, so even a system satisfying all fourteen would be a candidate rather than a case.

What it does deliver is a floor under the discussion. Before this, the question was argued between people asserting that machines obviously could be conscious and people asserting that they obviously could not. This produced a list of things to look for, applied it, and reported a result that disappointed both camps.

Part Three takes up the objection the report cannot answer from inside: that its founding assumption may simply be wrong.

Editorial Comment

MindHeaven® makes no claim about machine consciousness and sells nothing connected to it. The finding here is that no system was found to be conscious, which is neither good nor bad news for anyone selling anything.

We are writing about the footnote as much as the finding, because the discipline it shows is the discipline this library is trying to practise. An abstract sentence that permits a stronger reading than the data supports is the most consequential kind of error a paper can make, precisely because it is the sentence that travels.

We have made that error ourselves and corrected it publicly this week, in our series on L-theanine with caffeine. Seeing a group of twenty researchers do the same thing to a single word in their own abstract is a reasonable standard to be held to.

One more thing worth stating plainly, because this topic invites the opposite. Nothing in this report suggests that any system you can currently talk to is conscious. Its authors examined that question with more care than anyone arguing about it online, and their answer was no.