Derivative Epistemic Agency: When AI Helps Write the Science, Who Can Still Answer for It?
21 October 2026 · 5 min read
Most discussion of AI in research is about plagiarism, made-up references, disclosure statements and who counts as an author. These are real problems. But my article "Derivative Epistemic Agency: AI-Mediated Scientific Cognition and the Limits of Human Answerability", accepted in JMIR Metascience and Research Integrity, argues that they are symptoms of a deeper shift in how scientific claims are produced.
The problem beneath the familiar worries
Generative AI can help with scientific work while leaving authors responsible for claims they may not adequately understand. Authorship rules and disclosure statements make sure a human name stands behind every paper. But stating that one supervised the work does not show that one has the understanding needed to answer for it.
The article draws on social epistemology, theories of distributed and extended cognition, automation research, work on responsibility gaps, and metascience to describe this shift more precisely.
Not a yes-or-no question
Using AI is not simply either acceptable or unacceptable. The article offers a graded account of how researchers relate to the systems they rely on.
At one end is integrated epistemic extension: the researcher delegates work to a tool but keeps comprehension, the ability to evaluate the result, control over it, and the ability to answer for it. At the other end is derivative epistemic substitution: these capacities fall below what genuine answerability requires — even though the researcher is still listed as the author.
Responsibility asymmetry
The gap between these two is what the article calls responsibility asymmetry: a mismatch between the responsibility formally assigned to an author and the actual competence that author retains over what their paper claims.
The article sets out when this asymmetry arises, how it differs from ordinary reliance on tools, and why publication pressure in today's research system makes it more likely.
From diagnosis to practical rules
The article tests existing guidance — the ICMJE authorship criteria, COPE guidance, and contributorship frameworks such as CRediT and MeRIT — against this account, and finds that they establish who is accountable without checking whether accountability can actually be exercised.
It then proposes a governance model scaled to risk: a profile of how each author took part in the reasoning, tiered reporting of where key thinking came from, an enforceable standard for whether the reasoning can be audited, and fair treatment of AI use, for instance for researchers writing in a second language. The aim is rules that journals and institutions can actually apply, not just aspirations.
Why it is interesting: the article moves the debate from "Was AI used?" to "Can the author still answer for what the paper claims?" That question matters to journals, research institutions and readers of science alike, especially in health research, where published claims shape practice.
Elden, Å. Derivative Epistemic Agency: AI-Mediated Scientific Cognition and the Limits of Human Answerability. Accepted in JMIR Metascience and Research Integrity. Preprint: https://preprints.jmir.org/preprint/105907
Share this note