Patrick Butlin, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, Axel Constant, et al. (19 authors) (2023), arXiv preprint (report) — ARXIV-2308.08708
Key takeaways
- Assuming computational functionalism, the report derives a rubric of 14 “indicator properties” from leading neuroscientific theories of consciousness (RPT, GWT, computational HOTs, AST, PP, agency and embodiment). Applying it, the authors find that no current AI system is a strong candidate for consciousness, and that there are no obvious technical barriers to building systems that satisfy many of the indicators.
Seeded from the abstract and a skim, before the reading in NOTE-052. What follows is what the work says about itself.
Whether current or near-term AI could be conscious is a question of growing scientific and public concern. The authors argue for a rigorous, empirically grounded approach: assess AI systems in detail against the best-supported neuroscientific theories of consciousness. They survey recurrent processing theory, global workspace theory, higher-order theories, predictive processing and attention schema theory, and derive computationally stated indicator properties from them. They use these to assess several recent AI systems and discuss how future systems might implement them. Their conclusion is that no current AI system is conscious, but that nothing obvious prevents building systems that satisfy the indicators.
Standing in the record
Filed from the survey of 2026-09-25 of work the anthology set aside as out of scope (tier B): 90 seconds of active reading over 5 sessions in the papers-feed tracker. It was filed Deferred, unread. NOTE-052 is the close reading of 2026-09-25, and it placed the work: Active — the reference statement of the indicator-property (“theory-heavy”) approach to AI consciousness, worth reading for §1–§2 (method and theory survey) and Tables 1–2. Its §3.2 system verdicts are architecture-dated and cover only GWT and agency/embodiment. The 2026 Trends in Cognitive Sciences paper by an overlapping 20-author group may now be the citable version; I have not read it, so this should not yet be marked Superseded.
When seeded, before the reading in NOTE-052, the entry gave a priority for a deeper reading: medium — A widely cited boundary item that bears on anthology scope. The owner’s time is low (t=90 s over 5 sessions, i.e. repeatedly opened). The executive summary is captured, but the 88-page body is not.
What the seeding asked a deeper reading to check (NOTE-052 is that reading):
- The survey’s flag: its subject is AI systems, so it may belong in the anthology. The skim suggests it is an assessment framework (a rubric), not a training practice or a model-behaviour theory. If the anthology has a topic for evaluating properties of AI systems, this is a candidate for it. Otherwise it is a reading-list note under consciousness.
- A deeper reading should check the §3.2 LLM analysis in detail, since that is the part that ages fastest. It should also check whether the later journal paper by an overlapping author group (reportedly in Trends in Cognitive Sciences, 2025; unverified) supersedes the report as the citable version.
Access when seeded: I read the arXiv abstract page (v1 17 Aug 2023, v3 22 Aug 2023; cs.AI, cs.CY, cs.LG, q-bio.NC) and the full v3 PDF (88 pp.). I read the Executive Summary, Table 1 (indicator properties) and the table of contents, and skimmed §3.2 (the Transformer/LLM case study under GWT). A footnote on p. 1 records that an earlier version of a key sentence was softened (“shows” became “suggests”). I found no journal version.