Disclosing AI assistance
Disclosing AI assistance is not a confession, it is a methods section, and people who can write a methods section read as competent. The people who get burned are the ones who said nothing, got asked late, and had no records to point at. By the time a committee is asking, the window for looking careful has closed.
Three places it comes up
- Coursework
- Policies vary by instructor and by assignment, and a department-level policy does not override a syllabus. Read the syllabus first; if it is silent, ask before you submit, not after. Assume a written record of the exchange is worth having.
- Thesis or dissertation
- The methods section is the safer default — it describes how the work was produced, which makes it methodologically relevant. Some programs instead require a standalone statement or an acknowledgement; check your own before you submit. Whichever location, name the tools, the model versions, and which stages they touched.
- Conference or journal submission
- Most major venues now carry an explicit policy. The consistent through-line: generative tools cannot be listed as authors, because authorship requires accountability a tool cannot hold, and the human authors remain fully responsible for everything submitted — including anything the tool produced. Check the specific venue’s policy for the submission year, not the one you remember.
The line between assistance and authorship
The test that survives contact with an actual committee: can you defend it? If someone asks why a method was chosen, why a parameter has that value, or what a paragraph means, and your honest answer is that the model produced it and you did not check, that is not assistance you can sign your name to.
- Assistance — drafting text you then rewrote, generating code you read and tested, summarising sources you also read, suggesting analyses you then justified.
- Not assistance — results you did not verify, citations you did not open, code whose behaviour you cannot explain, or claims you could not defend if challenged.
The curriculum’s framing for the same problem is the verification habit in Part 6, and the AFD test in §1.4 — the category of work where being wrong hurts someone who is not you.
A statement you can adapt
Adapt it; do not paste it unchanged. A disclosure that does not match what you actually did is worse than none.
AI assistance disclosure
Tools used: [tool and interface, e.g. gemini-cli 0.x, Gemini 3.1 Pro
via the browser, Google Colab Data Science Agent]
Model versions: [exact versions and the dates they were used]
Stages involved: [e.g. literature search, exploratory analysis, code
scaffolding, prose drafting, copy-editing]
Not involved: [stages done without assistance — state these explicitly]
Verification: All generated code was read and tested by the author. All
cited sources were retrieved and read in full. All reported results were
reproduced from the logged model version, prompt, and random seed recorded
in [location].
The author is responsible for the entire content of this submission,
including any portion produced with the assistance described above.Make it reproducible or do not claim it
A disclosure is only worth as much as the record behind it. Log the model version and the exact prompt alongside every result, pin your dependencies, seed every random process, and commit the context file with the code — a model updated underneath you is an unlogged change to your method. Unit 06 covers this in full.