Can we detect consciousness in AI systems using theory-based indicators?

The report argues that assessing AI systems for "indicator properties" derived from scientific theories of consciousness is "the best method currently available for assessing whether AI systems are likely to be conscious." It surveys five theories — recurrent processing theory, global workspace theory, higher-order theories, predictive processing, and attention schema theory — deriving indicator properties "elucidated in computational terms that allow us to assess AI systems for these properties." Applying this method, the report's "analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators."

The method rests on two premises combined. First, computational functionalism as a working hypothesis: "it is necessary and sufficient for a system to be conscious that it has a certain functional organisation" that is computational, so consciousness is "multiply realisable" across substrates. Second, scientific theories identify which computational processes correlate with consciousness in the systems (mostly humans) they were built to explain. Combining these yields the "theory-heavy approach": checking "whether [AI systems] use computational processes which are similar to those described in scientific theories of consciousness," explicitly instead of looking for "theory-neutral behavioural signatures." The report is careful that functionalism does not license any substrate — quoting Michel and Lau, "Swiss cheese cannot implement the relevant computations" — so multiple realizability is bounded by which algorithms a substrate can actually implement.

This contrasts with Can any falsifiable theory of consciousness apply to LLMs?: Hoel's proximity argument rules out LLM consciousness formally, a priori, from any falsifiable theory given proximity to known non-conscious systems, whereas this report's piecemeal, empirical indicator-checking leaves the door open ("no obvious technical barriers"). The indicator properties here are architectural conditions theories deem relevant to actual consciousness, a different target from the behavioral cues that predict whether users perceive consciousness, as cataloged in What design features make users perceive AI as conscious? — one assesses internal computational organization against theory, the other catalogs external triggers of human perception, and the report's method is pointedly not about "theory-neutral behavioural signatures." It also bears on Does software intelligence exist independent of hardware and environment?: this report's computational functionalism isolates the consciousness-relevant properties at the level of algorithm, independent of hardware and environment, the same move that note argues is a dualist error when applied to intelligence — raising the question of whether that critique generalizes from intelligence to consciousness.

The report does not establish that any current AI system meets the indicators, only that none currently does by its own method, and that no "obvious" barrier blocks future systems — "obvious" leaves room for non-obvious barriers the surveyed theories may not capture, since both the theories and computational functionalism itself remain uncertain by the report's own admission. It also does not adjudicate which surveyed theory is correct, so the verdict is conditional on a method rather than a settled empirical fact: if the theory-heavy premises are wrong, both the "not conscious now" and "no barrier" conclusions could be too.