The Metaphilosophy of Human–AI Research Collaboration: Discovery, Selection, and the Conditions Under Which an AI-Generated Object Counts as a Contribution

Existing literature on AI and scientific inquiry mostly asks whether AI can discover independently. Ganapini (2025) answers this directly: current systems do not meet the criteria for independent scientific discovery, because discovery is an epistemic achievement requiring a competence and self-reflectivity AI does not yet possess. Duede (2023) asks a related but distinct question, whether a breakthrough produced by an opaque process can be justified at all. Neither question is the one a human–AI research collaboration actually needs answered, because neither assumes AI as a standing collaborator embedded throughout generation, critique, and self-audit rather than as a candidate independent discoverer. This paper states and answers four sub-questions that follow once AI is treated as a collaborator rather than a candidate agent. First, authorship: reviving Popper's classical discovery/justification distinction — which already held that a hypothesis's origin is epistemically irrelevant to its scientific status — dissolves most anxiety about who authors an AI-generated candidate, but relocates the real question to selection, where authorship attaches to whoever supplies and defends the criterion that separated one candidate from the rest, not to whoever or whatever produced the candidate set. Second, genuine critique: extending Lakatos's distinction between novel facts and ad hoc accommodation gives an operational test for when an AI-generated objection constitutes a real risk to a thesis rather than a risk the researcher's own prior framing had already prepared an absorption move for — directly addressing this corpus's own documented finding that every Objections section it has produced shows complete thesis survival. Third, the reasoning/generation distinction: because fluent text and truth-tracking text are not distinguishable by inspection, the paper grounds the distinction in a sensitivity criterion — whether a claim would have come out differently had the underlying fact been different — rather than in any property of the output's surface coherence. Fourth, closure: building on this corpus's own prior finding that AI self-review underperforms independent review, a genuine Popperian loop with AI is available at the generation stage but not at closure, because determining whether a given piece of evidence actually falsifies a claim requires exactly the accountability a model structurally lacks.