Detect tower · floor

Is using ChatGPT cheating, or does it depend on the rule?

Asking whether is using ChatGPT cheating feels like a moral question and is mostly not one. It is a question about which rule applies to you, and there are two very different regimes writing those rules right now — one for students, one for everybody publishing work under their own name. They disagree about the default, and both have written their reasoning down.

Notice what that policy does and does not say. It does not say the tool is dishonest, or that using it makes you a worse student. It says permission is not assumed, and the person who grants it is your instructor. Which means the genuinely useful move — before any argument about ethics — is to read your own course’s statement, because it exists and almost nobody looks at it until something has gone wrong.

That convergence is worth more than either policy alone, because it tells you the safe assumption when you cannot find your own institution’s wording: absence of a rule is not permission. San Diego is explicit about the corollary too — if your instructor did authorise it, “don’t assume you can use it in ways not authorized or for other assignments or courses”. The permission is as narrow as it was given.

San Diego also makes an argument neither of the other two documents makes, and it is the sharpest thing said about this anywhere. The problem is not only permission — it is citability. “You can’t really ‘cite’ ChatGPT, because the professor can’t follow up on your citation, unless you submit your chat history to your professor.” A citation is a promise that somebody else can go and check. An output generated once, for you, in a conversation nobody else can open, cannot keep that promise. That is a structural point rather than a moral one, and it explains why so many policies land on disclosure rather than prohibition: telling somebody you used a tool is possible, whereas citing it in the ordinary sense is not.

And they offer the single best practical test on the subject, worth more than any threshold: can you “articulate where your words/ideas end and the machine’s begins”? If you can, you almost certainly have not crossed a line. If you cannot, you have — whatever any detector says.

The harder question is the one your instructor’s policy probably does not answer in detail: where inside “using it” the line actually falls. Nobody thinks running a spellchecker is cheating. Almost everybody thinks submitting a generated essay is. Between those two points sit the questions people genuinely search for — is it cheating to use ChatGPT to proofread, for a grammar check, for ideas — and the most careful public answer comes from an unexpected direction.

So, applied to the three questions people actually ask.

Proofreading and grammar. Green, on that framework, and comfortably so: it supports expression without introducing new intellectual content. The reasoning is sound for coursework too — with one caveat that matters more in a classroom than in a journal. If the thing being assessed is your command of the language, then handing that part to a machine removes the very skill being measured. In a language course, or a piece marked on written expression, a grammar tool sits closer to amber than to green, and the reason is not ethics but assessment.

Ideas and structure. This is where the framework gets genuinely interesting, because it splits the two. Suggesting the structure of a piece is green. Contributing to the reasoning is amber — permitted, but requiring “verification and oversight” and disclosure, because AI at that point “does introduce new intellectual content”. The distinction is a good one and worth carrying into your own work: arranging your thinking is not the same as doing your thinking.

Writing it. Red, and the framework names the act precisely: “generating hypotheses, analyses or conclusions and presenting them as human-derived”. The offence is not that a machine was involved. It is the presentation — the claim of authorship over reasoning you did not do. That is why disclosure moves so many uses out of the red band, and why concealment is what actually defines the violation.

The detector is a separate question, and a weaker one. Students often collapse “was this allowed” into “will I be caught”, which gets both wrong. The University of Georgia, which permits only one detector on campus, publishes three cautions about it that are worth knowing: Turnitin “reports a sentence-level false positive rate of 4%”, and studies show that rate “is higher for students who speak English as a second language”; unlike a similarity report, “it is not possible for any AI detection tool to offer such context or suspected source”, so there is nothing to check the flag against; and the detector “was trained on output from GPT-3.5”, so “those with the resources to access more sophisticated tools than GPT-3.5 are better able to avoid detection”.

Read that last one again. The tool is least effective against the people best resourced to evade it, and most likely to misfire on people writing in a second language. Whatever else it is, it is not an instrument of fairness — and it is a poor reason to decide what is acceptable. Decide that from the rule, not from the probability of being flagged.

The short version, if you want one. Find your course policy, because it outranks every general argument including this page. Assume permission is not automatic. Where you do use a tool, keep the reasoning yours and say so if asked — disclosure is what separates most amber from most red. And keep your drafts, which protects you from a false accusation far better than any argument about ethics ever will.

What this tower will not do

It will not give you a verdict that overrides your institution. Your course policy decides, this page does not, and any page that claims otherwise is selling you something.

It will not help anybody conceal anything. The whole framework above turns on disclosure, and the detector section is here because people deserve to know how weak that instrument is — including the staff relying on it.

And it will not carry an affiliate link to a writing tool or a detector, which on this particular question means turning down both sides of the argument at once. What holds instead is simple: the three documents quoted here are published by two universities and a major academic publisher, each quoted with the date it was read.