Using AI for Homework Without Cheating

The question people ask — is AI cheating? — has been the wrong shape for about forty years. Institutions have already absorbed calculators, answer keys, solutions manuals, paid tutors, study groups, and the entire internet. Each arrived with the same panic and each was resolved with the same principle: what you submit has to be evidence of what you can do.

That gives you a test that works on any tool. Ask what the assignment was set up to measure, then ask whether you or the tool did that specific thing. A problem set on factoring quadratics measures whether you can factor quadratics; if the solver factored them, the measurement is of the solver.

What is new is the range. A calculator could only ever do arithmetic, so the boundary was self-enforcing. A solver can do the whole task, which means the boundary now lives inside your own decisions rather than inside the technology.

Read the policy text rather than summaries of it. The summaries are consistently more alarmist than the documents.

Stanford University.The Board on Conduct Affairs adopted guidance on 16 February 2023 stating that, absent a clear statement from a course instructor, "use of or consultation with generative AI shall be treated analogously to assistance from another person. In particular, using generative AI tools to substantially complete an assignment or exam (e.g. by entering exam or assignment questions) is not permitted. Students should acknowledge the use of generative AI (other than incidental use) and default to disclosing such assistance when in doubt."

The analogy to another person gives you a usable heuristic: if a classmate doing exactly this for you would be fine, the AI doing it is fine. Substantially complete is the actual prohibition, and the example given — pasting in the question — is precisely the act most students think of as ordinary use.

The Russell Group. In July 2023 the twenty-four research-intensive UK universities published five shared principles, none of which is a ban. They commit to AI literacy, to adapting teaching and assessment, and to upholding academic rigour and integrity — and they explicitly encourage departments to apply the institution-wide policy within their own context. So the university-level page is not your rule. Your rule is in the module handbook, and it can differ between two modules in the same term.

The JCQ.For UK school qualifications, the exam boards' joint guidance is the strictest and most specific. Where AI has been used it must be acknowledged with the name of the tool and the date the content was generated, and the student must retain the prompt and output in a non-editable form — a screenshot — with a short explanation of how it was used. Then the clause that matters most: content reproduced directly from an AI response will not allow a student to demonstrate that they independently met the marking criteria, and therefore will not be rewarded. Ofqual's 2026 guide for schools restates the same requirement in public: students must acknowledge any use of such tools or it is malpractice.

Disclosure, in other words, is not a loophole. Declaring that you copied does not convert copied work into credited work; it converts undeclared malpractice into a zero on that section. It protects you from the sanction, not the mark.

Two students hand in the same page. Both have solved $2x^2-7x+3=0$ and both have written $(2x-1)(x-3)=0$ followed by $x=\frac{1}{2}$ and $x=3$. The first found the factor pair herself, then typed the equation into a solver to confirm. The second typed it in first and copied what came back.

The artefact is identical. The acts are not, and no amount of forensic examination of the page will separate them. That is why the regime is written about your conduct rather than about the document: you are the only person with access to the evidence that decides this case. Uses of a solver fall on a gradient, and it is worth naming where you are on it:

• Verifying a finished answer. Fine almost everywhere, because it replaces none of the thinking.
• Unblocking one step. You attempted it, stalled, and asked what move comes next. Usually fine, and often the fastest route to learning the move.
• Generating a method for a problem you have not attempted. Reasonable on a practice set. Not on an assessed problem set where method selection is the graded skill.
• Producing the submitted artefact. Not acceptable under any policy in this article, disclosed or otherwise.

Before you submit anything you used a solver on, run these four. They take under a minute and they map closely onto what a panel would ask.

1. Could I reproduce this tomorrow, closed book, on blank paper? The strongest of the four, because it tests the thing the assignment was proxying for. If the answer is no, you are also in trouble on the exam.

2. Would I be comfortable if my instructor read the exact prompt I typed? The prompt is the honest record of intent. "How do I decide between substitution and elimination here?" and "solve question 4" are different sentences written by different people doing different things.

3. Did I attempt it first? If you have already fought with the problem, the explanation lands on a hook — you know which step you could not find. If you have not, it lands on nothing and evaporates within a day.

4. Could I explain every line if asked right now? Not paraphrase it — explain why that step is legal and why it was chosen. A marker who suspects copied work asks this in the tutorial rather than filing a report. Failing questions 1 and 4 together means the work is not yours in any sense that matters, whatever the paperwork says.

Most students who get into trouble for AI use were not trying to cheat. They used a tool in a way their instructor would have waved through, then wrote nothing about it, and the absence of a note became the story. The JCQ format is a good default even outside UK schools, because it captures everything an assessor could want: the tool, the date, the prompt, the raw output, and what you did with it. A complete example, attached to a problem set:

AI use: Gauth (gauth.io), 3 September 2026. Question 4 only. I attempted it first and got $x=2$ instead of $x=\frac{1}{2}$. I entered the equation to locate the discrepancy and found I had dropped a negative when subtracting the bracket on line 3. I then re-solved the question by hand. Screenshot of prompt and output attached.

That is specific enough to be verifiable, it shows the tool was used after the attempt rather than instead of it, and it names the actual error. Compare the vague version — "I used AI to help with some questions" — which invites exactly the follow-up you do not want.

For short maths homework, one line under the question is enough: checked Q4 and Q7 with a solver after completing them; no changes made. We keep a fuller account of how students should use this product on our academic integrity page, and an honest account of where the solver itself gets things wrong on the accuracy page.

Set the ethics aside and treat this as a selfish calculation, because it loses on those terms too. In Roediger and Karpicke's 2006 experiment, students who re-read a passage four times scored 83% five minutes later against 71% for students who instead recalled it from memory — then 40% against 61% a week later, while remaining the more confident group. Fluency, the smooth feeling of reading something you recognise, is a poor signal of whether you can produce it later, and it is exactly the feeling a clean worked solution gives you. Copying is weaker still, because it is transcription rather than reading to learn.

The second cost is specific to maths. Exams do not reproduce the problem you copied; they produce a cousin of it with different numbers and one structural change. Recognising the cousin is the actual skill, and it is built by attempting problems and failing at them. It is also why marks go to method rather than answers, which we cover in how markers award method marks.

We sell a solver, and here is our position. We would rather you used it three times well in a week than thirty times badly. The version of this product that makes us the most money in the short term is the one you use to avoid thinking, and it is the version that leaves you unable to pass the exam.

None of this requires you to work harder. It requires you to attempt the problem before you ask, to ask for the next move rather than the solution, and to close the answer and reproduce it from a blank page before you believe you have learned it. The reasoning behind that order, and a seven-day plan built on the same evidence, is in our guide to studying for a maths test. For the algebra most of these gaps turn out to be, start with the four factoring patterns and the algebra solver.

Four situations come up repeatedly where the substitution test does not settle things cleanly.

Language, not content. A student writing in their second language asks an AI to fix the grammar of an argument they constructed themselves. If the assignment assesses mathematics, this is not a substitution of the assessed skill. If it assesses written communication, it is. Most institutions land on the first reading for STEM coursework, but the syllabus decides and you should ask.

Accommodations and group work. Where a tool is part of an approved disability accommodation, the accommodation governs; get it in writing through the disability office. And if three of you split six questions and one used a solver, all three of you submitted it. Agree the rule at the start.

Silence.A course that says nothing about AI has not permitted it; it has failed to consider it. Stanford's guidance exists to fill that gap, and its default is that AI is treated like another person. Assume the restrictive reading, ask by email, and keep the reply.