What happened: Nineteen researchers derived computationally specifiable indicator properties from the leading scientific theories and assessed AI architectures against them. The report states its footing first: it adopts computational functionalism — "the thesis that performing computations of the right kind is necessary and sufficient for consciousness" — "as a working hypothesis", calls it "a mainstream—although disputed—position", and gives pragmatic reasons. Its choice of evidence is argued rather than assumed: "The main alternative to this approach is to use behavioural tests for consciousness, but this method is unreliable because AI systems can be trained to mimic human behaviours while working in very different ways." Birch, a co-author, records the group's reasoning from inside — "given the gaming problem, what we need are deeper architectural indicators that the system is unable to game" — on the thought that "the system can't play-act with regard to its own architecture." The method weighs rather than sorts: systems "which possess more of the indicator properties are more likely to be conscious", on a rubric self-described as "provisional". Its verdict carries an author's erratum: an earlier abstract said the report "shows that there are no obvious barriers to building conscious AI systems", amended because "satisfying the indicators would not mean that such an AI system would definitely be conscious." The 2025 follow-up, now in Trends in Cognitive Sciences with Bayne and Chalmers added, is a guide to the method rather than a re-assessment, and it loosens all three fixings: the approach "can be used with other theories, including theories yet to be developed"; Box 2 demotes the assumption to "one version of the method we propose" and readmits IIT and biological substrate views as alternatives; and Box 3 concedes that the gaming problem "can also arise for computational markers", since a nonconscious system can be built to carry any feature short of sufficient. The mitigation offered is to check whether a system has "high computational similarity to biological systems that are known to be conscious."
Diagnostic reading: The programme declines the discriminator this corpus repeatedly says no one supplies — and declines it in its own voice, not under pressure. The erratum is the evidence: a report that retreats from shows to suggests, and from conscious AI systems to systems which satisfy these indicators, is not withholding a criterion it holds. Under ARG-02 v1.7's trilemma this is Path B taken openly — organisation does the discriminating, the substrate claim recedes, and functionalism's own line-drawing questions are inherited — and declaring the path is its honest form. ARG-02 already prices the class at §"What the Trilemma Addresses", where marker-based methodologies either weight their markers, re-posing the fork, or return no verdict; what the corpus had not done is name the strongest instance where it prices it. The symmetry acknowledgement is more than courtesy: the indicators are third-person and architectural, exactly the class FND-01 v2.4 §7.3 leaves admissible when it bars reports, and chosen for the same reason — behaviour can be produced without what it appears to report, the gaming problem received at ARG-03 v2.6. Two programmes from different axioms reach the same admissibility rule. Where they part is what the evidence then settles, and the sharpest statement of that is again a co-author's: Birch's Janus problem, "the same evidence can point in opposite directions, depending on one's background beliefs", with a second level — deadlock over "whether finding a tentatively consciousness-linked computational feature in AI even counts as evidence for consciousness or not" — and the concession that the 2023 group "assumed there could at least be agreement on that second-level matter." That is framework-dependence reached from inside the field's best-developed programme, by one of its builders; no refinement of an indicator reaches it, because what an indicator counts as evidence for is fixed before it is measured. The gaming concession then closes the loop in a way worth stating precisely: the fallback is computational similarity to known-conscious biological systems, and specifying the respects functionally is Path B by ARG-02's own fork — so the regress is internal, returning the programme to the line-drawing it already inherited rather than sliding into likeness-to-us. None of this is a defeat on the programme's own terms, and the architectural turn remains a real advance over marker-reading. The Janus problem cuts at 0,0 exactly as it cuts at biological naturalism, which is why it is recorded here as a shared condition and not a scored point. [Derived]
Both papers held; every quotation verified against the PDFs 2026-08-30 (working/BUTLIN_SOURCE_VERIFICATION_2026-08-30.md). Cite by version and date: the report is arXiv:2308.08708v3, 22 Aug 2023; the follow-up is the advance-online TiCS article, doi:10.1016/j.tics.2025.10.011, adding Bayne and Chalmers, with no volume or pagination yet assigned. Birch is quoted from the Centrist Manifesto v6, whose footnote 22 attaches his account to the 2023 report by name. The reports prohibition (FND-01 v2.4 §7.3) is untouched throughout.