Douglas Hofstadter (2026), Philosophical Transactions of the Royal Society A 384(2320), 20240527, theme issue “World models in natural and artificial intelligence” — DOI-10.1098/rsta.2024.0527 (received 21 January 2025, accepted 19 June 2025, published 14 May 2026; CC BY 4.0).
Standing in the record
Filed on 2026-10-01 at the owner’s request for Douglas Hofstadter’s work in this record, and named by NOTE-136 as a candidate follow-up to the theme issue’s editorial (LIT-195). It is the issue’s lead contribution: the editorial opens its §2(a)(i) with it and takes its epigraph from it.
Unavailable because no lawful full text could be reached (see the history
note), so no NOTE is filed. What the record can say without the essay, all
of it second-hand through NOTE-136’s reading of the editorial:
- The editors quote Hofstadter that a system which passes a Turing test probing for consistent understanding should be credited with thinking, and that conscious AI may arrive “perhaps even by 2033” (NOTE-136, claim C8).
- NOTE-136 says, under its limitations, that Hofstadter “argues nearly the opposite” of the editors’ verdict that the issue converges on LLMs “probably” lacking coherent causal world models, consciousness and a likely path to superintelligence.
Is “nearly the opposite” right? It cannot be confirmed from what this record has read, and on that evidence it looks too strong. The editors’ verdict is about today’s LLMs. The quotation NOTE-136 reports is a criterion (understanding shown in a demanding Turing test suffices for thinking) and a forecast (conscious AI perhaps by 2033). A forecast that consciousness may arrive within a decade does not say present systems have it, and is compatible with the editors’ “probably not yet”. What the reported quotation does oppose is the editors’ framing: a behavioural criterion for thinking, where the editorial ties world modelling and consciousness to embodiment, agency and life. The defensible wording, until the essay is read, is that Hofstadter dissents from the editors’ framing and is more open than they are to machine thought and consciousness; whether he says current LLMs think or have an “I” is the question a reading must answer. That correction belongs in NOTE-136 once the essay is read, not before.
Where it would sit in the record:
- It is a party to the dispute THEORY-023 describes. A Turing-test criterion takes humanlike behaviour as evidence of thought, which is the evidence THEORY-023 says mimicry undercuts in a system trained on human output. Seth’s biological naturalism (LIT-135) and Roberts’s account of chatbot self-ascriptions as fiction (LIT-207) are the record’s readings on the other side; Freeborn’s conditions for understanding in deep networks (LIT-131) are the nearest attempt to make “understanding” testable.
- The title’s “I” is Hofstadter’s long-standing subject: the self as a self-referential pattern, the thesis of Gödel, Escher, Bach and I Am a Strange Loop, which are being filed beside this work. Whether the essay applies that account to language models is, again, for a reading to say.
- The Fluid Analogies Research Group’s papers filed with it (the Copycat paper, LIT-435, and the critique of hand-coded representations, LIT-414) are where his earlier criterion for understanding was worked out: perception and concepts built by the system itself. A reading should check whether the essay holds LLMs to that criterion.
What a reading should check: the exact form of the Turing-test criterion and of the 2033 forecast; whether the essay says present LLMs think or are conscious; and whether it engages Krakauer, Krakauer & Mitchell’s “less is more” contribution to the same issue, whose reading of OthelloGPT as a “bag of heuristics” the editorial reports (NOTE-136). Melanie Mitchell, Copycat’s co-author, is also among the editorial’s authors.