Epistemic Dependence at Machine Speed: Extending Kitcher's Division of Cognitive Labor and Hardwig's Conditions for Warranted Trust to Human–AI Research Collaboration

The question a research programme using AI assistance owes its reader is not whether the researcher depended on a source they could not personally verify from first principles — no researcher, with or without AI, achieves that kind of independence, and Hardwig (1985) showed why demanding it produces worse beliefs, not more reliable ones: a researcher who trusted no testimony would hold only what they could personally re-derive, which is a narrower and cruder set of beliefs than the one any working scientist actually holds. The question is which conditions make a given act of epistemic dependence warranted, and how a research programme's division of labor should be organized so those conditions can be checked by a reader rather than taken on faith. This paper extends two established results in the philosophy of science to the specific case of human–AI research collaboration. Kitcher's (1990) division of cognitive labor shows that a research community accumulates more knowledge, faster, by allocating investigative effort unevenly across specialized contributors than by requiring every contributor to independently replicate every step — a result that reframes the human–AI question from whether to delegate to along which lines delegation is efficient. Hardwig's (1985, 1991) conditions for warranted epistemic trust — track-record competence in the specific domain of the claim, and an accountability structure that allows the dependent party to be asked how they know and to bear consequences for being wrong — supply the test for where that line should fall. Applied to current AI systems, competence is measurably domain-specific rather than global: recent benchmark evidence (Halder et al., 2026) finds leading models generate and recombine claims capably but fail to catch their own errors in the majority of self-review trials, performing markedly worse than when an independent process performs the check. The accountability condition fails more fundamentally: a model has no continuous, consequence-bearing self across sessions to answer for a claim, which is the specific and principled reason verification cannot be delegated to the same process that generated the claim, independent of any single documented error. The paper closes with a general, forward-looking protocol — not a response to any particular incident — for dividing research labor and disclosing that division in a way a reader can audit.