Once it's demonstrated that LLMs can produce referee-report-style philosophical artefacts with minimal prompting, the debate over whether LLMs can "do philosophy" must relocate. The crude skeptical question "can they even produce philosophy-flavoured text?" becomes untenable. The real question becomes: do LLM outputs instantiate the constraint structure that distinguishes good [[abductive reasoning]] from merely persuasive dialectic? ## Why this matters Any anti-LLM argument that relies on "they can't even do X" must now target deeper constraint properties—stability, non-ad hocness, cost discipline, commitment management—rather than mere ability to produce structured criticism. The "appearance vs reality" rhetoric ([[Floridi's critique of LLM abduction|Floridi's "abductive appearance"]]) has to be cashed out by identifying specific text-internal failures, not by gesturing at the production mechanism. ## The minimal-prompt phenomenon Consider cases where very thin genre-governing prompts ("is this paper correct?", "be philosophically robust", "focus on the arguments") are sufficient to elicit substantial philosophical structure: argument reconstruction, hinge identification, cost-accounting, alternative-theory comparison. If the only question were "can they produce philosophy-flavoured text?", these cases would settle it trivially. Since skeptics don't accept that, they must be demanding something more—and that "something more" is precisely the constraint structure. This connects to the [[Dialectical saturation thesis]]: the training distribution contains enough philosophical dialectic that LLMs have effectively learned the public norms of philosophical assessment. ## Relation to the Floridi-Williamson foil This responds to the pessimistic argument that combines [[Luciano Floridi]]'s claim (LLMs produce only "abductive appearance") with [[Timothy Williamson]]'s claim (abduction is central to serious philosophy). The response isn't to deny either premise, but to attack the inference by showing that philosophical abductive constraints are largely text-internal and publicly codified—and therefore learnable and instantiable by systems trained on philosophical corpora. --- Source: [[Web Clippings/ChatGPT 1|ChatGPT conversation on Floridi, Williamson, abductive competence]]