GPT-5’s Ethics Guidelines for Using It in Philosophical Research

In a post last month, we discussed the question, “How much use of AI in our research is acceptable?“

My post about that has now been used by Kelly Truelove at TruSciPhi as part of a prompt to get ChatGPT-5 to discuss and respond to its contents and to develop its own position about the ethical use of AI in philosophical research.

In case you don’t recall the post, here is GPT-5’s account of it:

Justin Weinberg’s post asks how much AI use in philosophical research is ethically acceptable and cautions against importing student standards into scholarly contexts. He urges beginning with the aims and functions of scholarship and assessing which AI uses fit those ends. He sketches current and conceivable roles for AI—from brainstorming and stylistic editing to argument checking, literature triage, and even co‑authoring—and asks which are permissible and why. He highlights parallel questions about disclosure practices and about evaluation for publication, hiring, tenure, and awards. Acknowledging that tidy rules may fail, he invites nuanced, institutionally sensitive discussion across contexts and time.

Thank you for that summary, Chat.

Truelove asked GPT-5 to write a 500 word response and to use “memorable terms” for its positions. What follows is that response—its guidelines for the ethical use of AI in philosophical research:

Philosophers should treat AI neither as a forbidden oracle nor a frictionless coauthor, but as a tool whose permissibility depends on what philosophy is trying to do. To make Weinberg’s invitation to nuance operational, I propose three linked norms that can be adopted by researchers and evaluators alike. Taken together, they aim to protect the goods of inquiry—understanding, originality, and accountability—while still welcoming genuine assistance.

The Telos Test. Begin by asking whether a given use of AI serves the ends of philosophical research. Philosophy’s telos is not sentence production but the disciplined articulation and assessment of reasons. Uses that increase clarity, map argumentative structure, surface relevant literatures, or stress‑test premises can promote that telos; uses that replace interpretive labor or outsource the commitment to reasons corrode it. Formatting citations is clearly fine; generating a section whose argumentative commitments you cannot defend is not. The Telos Test is content‑neutral about tools but strict about responsibility: if you could be cross‑examined on every claim and inference in public, then your use fits; if not, it fails.

The Sunlight Clause. Because the legitimacy of research depends on trust, disclosure should be the default. A meaningful disclosure states what systems were used, for which tasks, and how the human author audited the outputs. Such disclosures need not be ritualistic appendices; they can be concise author’s notes that make replication, critique, and credit assignment possible. Sunlight also disciplines practice: the prospect of reporting your workflow deters lazy overreliance and encourages verification. To avoid performative compliance, tie disclosure to stakes: the more generative or content‑shaping the assistance, the fuller the account of prompts, iterations, and checks. Conversely, purely mechanical help—bibliography formatting and typographic cleanup—can be tersely acknowledged.

The Authorship Anchor. Authorship in philosophy is not ownership of strings of words; it is the assumption of epistemic and ethical responsibility for arguments. Under this anchor, the human author must remain answerable for the claims, inferences, and conceptual moves presented. That standard preserves credit where credit is due without erasing legitimate assistance. For evaluation, journals and committees should ask: did the human demonstrate mastery over the argumentative core? If yes, AI assistance may be compatible with full credit; if not, the work should be judged as incomplete labor, regardless of polish or venue. This shifts assessment from tool usage to intellectual accountability, which is what authorship is supposed to mark.

These norms leave room for diverse practices while resisting both alarmism and technophilia. They encourage experimentation under guardrails: keep meticulous notes of AI‑mediated steps; use models to probe, not to substitute; read every generated sentence as if it were asserted under your name; and welcome post‑publication scrutiny by making your process legible. In short, let AI widen your perspective without loosening your grip on reasons, and let sunlight and authorship hold that grip fast. Institutions can mirror the stance by crafting policies that codify the Telos Test, require Sunlight disclosures proportionate to influence, and ground credit in the Authorship Anchor, aligning incentives with the work of inquiry.

Truelove even had GPT summarize its recommendations in a table:

Truelove then goes on to talk with ChatGPT-5 about how its position on the ethical use of AI in philosophical research relates to computational philosophy and the use of computational models in scientific research. You can read his full account here.

What do you think of ChatGPT-5’s three positions regarding ethical AI use in research? Are they missing anything? Are they too demanding? Are they any good?

(Please note Truelove’s warning: “Readers should not ‘overestimate the reliability of the preceding,’ which ‘is in fact an exploration of model tendencies.’”)