Using AI Tools Ethically
AI assistants are now part of the research toolkit, and pretending otherwise helps no one. Used well, they save real time. Used carelessly, they invent citations, launder unverified statistics into your work, and creep into thinking you're supposed to be accountable for. The skill is neither "AI good" nor "AI bad." It's knowing, task by task, when to go ahead, when to use it but verify, and when it's simply off-limits. This lesson gives you that decision sense, then lets you practice it on eight situations you'll actually face.
Where AI genuinely helps
Plenty of legitimate uses have nothing to do with faking results. AI can explain a cryptic error message, suggest why your code won't run, act as a tireless devil's advocate against your argument, help you brainstorm explanations for a surprising finding, or tidy the grammar and flow of prose you wrote yourself. In each case you stay in charge: you decide whether an idea is any good, you run and check the code, you keep your own voice. Think of it as an eager, fast, occasionally-wrong intern — useful, but never the person whose name is on the work.
Where it fails: fluency is not truth
A language model generates plausible text. Left to itself it does not know facts, look anything up, or calculate, and that single property explains its three most dangerous failure modes. Many assistants can now search the web or run code when they choose to, which genuinely helps, but it changes the size of the problem rather than its shape: you usually cannot tell from the answer whether it searched or simply sounded confident, a retrieved source can still be summarized wrongly, and code still runs on whatever inputs you handed it. The safe assumption remains that nothing has been checked until you have checked it.
- Fabricated citations. Ask for references and it will happily produce authoritative-looking ones (real journal, plausible authors, convincing title) for papers that do not exist. This isn't rare; it's how the tool works.
- Unverified statistics. Ask it to "compute" a p-value or a mean and it will output a confident number that may be wrong: unless it genuinely ran the calculation, what you are reading is a prediction of what the answer would look like, and it arrives formatted identically either way.
- Hallucinated facts. It will state historical "facts," study findings, or definitions with total fluency and no flag when it's simply making them up.
The unifying lesson: fluency is not evidence. Confident, well-written output is exactly what the model is built to produce, whether or not the content is true.
The rule that never bends: you own every word and number
Whatever help you take, verification is non-delegable. If a sentence, statistic, or citation appears in your work, you are vouching for it as if you'd produced it yourself, and "the AI told me" is not a defense to a supervisor, an examiner, or a journal. Every number must be reproducible from your data in real software; every citation must point to a source you have actually read; every factual claim must be checkable. You can use AI to get to a draft faster, but you cannot outsource responsibility for what that draft says.
The data-protection trap. Never paste participant data (transcripts, survey responses, anything identifiable) into an external AI tool. Doing so discloses your participants' data to a third party they never consented to, breaking your data-protection duties and probably your ethics approval. Even "I removed the names" isn't safe, because stories and quasi-identifiers still expose people (see Privacy & Confidentiality and Data Privacy Basics). Keep participant data inside approved, secure systems only.
Disclosure, and why AI can't be an author
Whether you must declare AI use depends on your institution and your target journal, and those policies genuinely differ and are still changing, so check yours and disclose honestly when asked. One thing is settled everywhere, though: an AI tool cannot be an author. Authorship requires accountability (the ability to stand behind the work and answer for it), and software cannot take responsibility (the same accountability criterion from Plagiarism, Authorship & Citation). If AI helped substantially, the honest move is a short note in your methods or acknowledgements describing how, not a byline.
Run the decisions
Eight everyday tasks. For each, decide whether it's a go (fine as described), a verify / disclose (usable, but only with a specific safeguard), or off-limits, then reveal the reasoning. What matters is the judgment, not a perfect score.
🤖 Can I use AI for this?
Route each task through the go / verify / don't gates.
Same tool, three verdicts. The gate deciding each case is never the tool itself; it's what's at stake. Touch your thinking or writing in a way you still own and verify: usually fine. Touch facts, numbers, or citations you'd pass on unchecked: verify first, because fluency isn't truth. Touch participant data, reproducibility, or authorship: off-limits, because those are duties you can't hand to a machine.
Why it matters: AI won't get you in trouble for existing. It gets you in trouble when it lets you skip the parts of research that are yours to own: reading your sources, reproducing your statistics, protecting your participants, and standing behind your words. Treat it as a fast assistant whose every output you check, disclose your use honestly per your policy, and you get the time savings without trading away your integrity.
Common questions
Is it unfair to use AI to improve my English if I am not a native speaker?
Using a tool to say your own idea more clearly is not the thing academic integrity rules are aimed at, and it is worth saying so plainly, because this question causes a lot of unnecessary anxiety. The line falls in the same place as it does for a proofreader or a writing center: the ideas, the analysis and the responsibility for every claim stay yours, and what the tool touches is the expression. Polishing grammar in a paragraph you wrote is on the safe side of that line; asking for a paragraph about a topic and adapting it is not, however much you edit afterwards. Two practical cautions. Check that the polished version still says what you meant, since these tools smooth meaning as readily as grammar and are confident either way. And follow whatever your institution asks you to declare, which for language editing is often nothing, but is a question to answer from the policy rather than from hope.
Why does AI invent references that don't exist?
A large language model generates text by predicting plausible next words rather than by consulting a database of real papers. A reference is just a pattern (author, year, journal, title), so the model can produce one that looks completely convincing yet corresponds to no real paper: right-sounding authors, a real journal, a plausible title, all assembled from nothing. This is often called hallucination, and it is a structural feature of how the tool works, not an occasional bug. The practical rule is to treat every AI-suggested reference as unverified: find the actual paper, confirm it exists and says what you're citing it for, and cite that — never paste an AI reference list into your work unchecked.
Do I have to disclose that I used AI in my research?
It depends on your institution's and your target journal's policy, so the honest answer is: find out and follow it. Norms are still settling and they genuinely differ — some ask you to declare any AI assistance including editing, others only substantive use, and many now require a statement describing how AI tools were used. What is universal is that AI cannot be listed as an author (it can't take responsibility for the work) and that you remain fully accountable for everything in the document. When in doubt, disclose: a short, honest note on how you used AI costs nothing and protects you, whereas hiding substantive use you were required to declare is the integrity problem.