Is Using ChatGPT Cheating? Where the Line Actually Falls

The short answer

Using ChatGPT is not cheating in itself. It becomes cheating at the point where it stops assisting your work and starts producing it. Nearly every academic integrity code draws its line in the same place — you may get help with language, ideas and organisation, but the text and the argument you submit must be yours — and using AI to generate the words or the reasoning you are being examined on crosses that line whatever tool does it. The complication is that the exact placement of the line is set by your own institution, and the rules genuinely differ between universities, faculties and individual supervisors.

So there are two questions, not one. The general question — where does academia draw the line, and why there — has a fairly stable answer, and this post gives it. The specific question — what may you do, in your programme, this semester — has only one reliable source, which is your own examination regulations and the wording of the affidavit you will sign on submission.

What does the spectrum of AI use look like?

“Using AI” describes acts so different from each other that the phrase is close to useless in an argument. Laid out roughly from least to most contested:

1. Spellcheck and autocorrect. Machine language technology that nobody has objected to for thirty years, and which is built into the word processor your university licenses for you.
2. Grammar and style correction. Fixing agreement, punctuation, article use, clumsy phrasing — in text you wrote.
3. Translation. Reading a source in another language, or rendering your own writing into the submission language.
4. Brainstorming and discussion. Talking through possible angles, testing whether your research question makes sense, asking what the standard objections to a position are.
5. Outlining and structuring. Asking how a methods chapter is normally organised, or whether your chapter order is logical.
6. Summarising sources. Having a paper condensed rather than reading it.
7. Drafting text. Generating sentences, paragraphs or sections that then appear in your document, edited or otherwise.
8. Generating the substance. Having the argument, the analysis or the interpretation produced for you.

The useful thing about the list is that it makes the disagreement smaller than it feels. Very few people are arguing about items 1 to 4. Almost nobody defends items 7 and 8 in an assessed submission. The live argument is a narrow band in the middle — and it is much easier to ask your supervisor a precise question about that band than to ask whether “using AI” is allowed.

Which uses are generally permitted?

The lower end of the list is permitted in most institutions, and often not even mentioned in policy because it was never considered controversial. What these uses have in common is that they operate on work you already did.

• Correcting your own text. Spelling, grammar, punctuation, and the light touch of style correction a proofreader would apply. Universities have permitted human proofreading for decades; a tool doing the same job is not a new category of act.
• Understanding sources. Translating a paper so you can read it, or asking for a concept to be explained to you. You are still the one who reads, judges and cites.
• Thinking out loud. Nothing produced in a brainstorming conversation reaches your document. The output is your own subsequent thinking.
• Mechanical and administrative work. Reformatting a reference list, converting citation styles, generating a table of contents, checking that your figure numbering is consistent.

There is a caveat that belongs here rather than as a footnote: these tools invent citations, confidently and in correct format. Anything a model tells you about the literature is a lead to verify, never a source to cite. A fabricated reference in your bibliography is a problem under the rules that existed long before AI did.

Which uses are generally not permitted?

The upper end is where institutions are consistent, and it is worth being blunt about it, because the euphemisms in this area do students real harm.

• Text that goes into your submission having been generated rather than written. Editing generated text afterwards does not change its origin. “I rewrote it in my own words” describes paraphrasing someone else’s work, which is the thing academic integrity rules were written about in the first place.
• The argument, the analysis or the interpretation. This is the thing you are actually being examined on. A thesis is an assessment of whether you can do it.
• Anything you would not be able to explain in a viva. A practical test that works better than most policy language: if you could not defend a paragraph, or say why you made the choice it describes, it is not yours.
• Generated data, results or sources. This is fabrication, and it is treated far more seriously than unauthorised assistance.

Why is the line drawn at authoring rather than at AI?

Because the line was never really about AI. Universities have always distinguished assistance from substitution, and the rules that govern ChatGPT are mostly the same rules that already governed proofreaders, writing centres, study groups, statistics consultants and translators.

A degree certifies that a particular person can do a particular thing. Help that improves your ability to demonstrate what you can do is compatible with that: a writing centre pointing out that your paragraphs lack topic sentences makes the assessment more accurate, not less, because it removes an obstacle between your competence and the examiner. Help that supplies the competence being assessed breaks the certification, because the document no longer evidences anything about you.

Two consequences follow from seeing it this way. The first is that the same tool can fall on either side depending on the job you gave it.There is no list of forbidden software, and asking “is ChatGPT allowed” is the wrong shape of question. The second is that what is being assessed matters. Translation is language support in a thesis assessed on its research contribution, and much closer to substitution in a language degree where writing in that language is the point.

What about the genuinely unclear middle?

Summarising, outlining and heavy editing are where honest people disagree, and where policies differ most. Three questions resolve most cases without needing a rule to point at:

• Did any generated wording end up in the document? If yes, you are in authoring territory regardless of how the task was described.
• Did it replace something you were supposed to do? Summarising a paper instead of reading it replaces the reading you are being credited for. Summarising a paper you have read, to check you understood it, does not.
• Would you be comfortable writing it in the declaration? This is the most reliable test, because you may soon have to. Reluctance to name a use is usually accurate information about that use.

What is the AI clause in the submission affidavit?

Every thesis is submitted with a signed declaration of independent work — the affidavit, or in German the Eigenständigkeitserklärung — in which you confirm that you wrote the work yourself and used no sources or aids other than those you have named. Many universities have now added a clause covering AI specifically.

Wording varies by institution, and yours is the only one that binds you, so read it rather than any example. In shape, though, the added clause usually does one of three things:

• Requires disclosure. You state which AI tools you used and for what purpose, sometimes in an appendix listing tool, task and section.
• Prohibits generation. You confirm that no part of the submitted text was generated, while permitted uses such as language correction remain allowed.
• Assigns responsibility.You confirm you are accountable for the entire content, including anything produced with such assistance — which forecloses the defence that a fabricated citation was the tool’s fault.

Two things about this clause matter more than its exact wording. First, it is a formal declaration, and a false one is a distinct and more serious offence than the underlying rule-breaking would have been — this is why concealing permitted use is a bad trade. Second, it means the safest position is not minimal AI use but accurately declared use.“I used DeepL Write for grammar correction throughout, and a language model to discuss possible research questions in the planning phase” is a sentence that ends a conversation. Silence is a sentence that starts one.

How do you find out what your own rules say?

Nobody can tell you your rules from outside, and any site that gives you a confident universal answer is guessing. The rules exist in four places, and they can contradict each other:

1. The examination regulations and any AI guidance the university has issued alongside them.
2. The faculty or programme handbook, which is frequently stricter than the university-wide document.
3. The affidavit itself, which is the one you sign and therefore the one that binds you.
4. Your supervisor, who may set narrower conditions and who will be the person any question is routed to anyway.

Ask specifically and ask in writing. “May I use AI?” invites a cautious no. “I intend to use a grammar tool on my own drafts and to have German sources translated while reading. Is that within the rules, and how should I declare it?” invites an answer you can keep. Save the reply.

Does honest use still get flagged?

Yes, and this is the part that catches people who did nothing wrong. A detector reads the finished text. It has no access to your permission, your declaration or your process, and heavy but entirely permitted editing pushes prose toward exactly the regular, predictable patterns these tools react to. So does careful academic style, and so does writing in a second language.

The protection is not writing worse on purpose. It is keeping the record: version history intact, drafts saved, notes dated, and a declaration that already names what you used. We set that out in how to prove you wrote it yourself, and what to do if it has already gone wrong in what to do when you are accused of using AI.

If you want to know what a detector sees in your own writing before your university runs one, our AI check is free for the first 1,500 words with no account. Be clear about what it does and does not do: it cannot tell you whether your use was permitted, that is a question for your regulations, and it cannot prove you wrote anything — no detector can establish authorship in either direction. What it can do is show you which passages of your own text read as statistically machine-like, next to our measured false-positive rate for your language and length, including the cases where we decline to give a verdict because no honest one exists.