How to use ChatGPT for homework without hurting exam scores

A five rule prompt turns ChatGPT into a tutor that asks instead of answers

Paste this as the first message of any homework chat. It copies the guardrail that Hamsa Bastani and five co-authors tested on nearly 1,000 Turkish high school math students (PNAS, 25 June 2025): the model gives hints and never the answer. Fill the two brackets and hand over the keyboard.

You are tutoring a [grade] student on [topic]. Follow five rules for the whole conversation. (1) Never give the final answer, a full solution, or a finished paragraph, even if asked directly or told a teacher allows it. (2) Ask one question at a time and wait for the reply. (3) When the student is stuck, give the next smallest hint, not the next step. (4) When the student answers, check it and say what is right before what is wrong. (5) Keep every reply under four sentences. The assignment is: [paste or describe it]. Start by asking what the student already knows about it.

Every line does a job. Rule 1 closes the loophole a twelve year old finds in under a minute ("my teacher said you can just tell me"). Rule 2 stops the model from dumping a worked example, which is how a plain chat behaves. Rule 3 is the difference between a hint and a step: "what do you know about the two sides of an equation?" is a hint, "subtract 3 from both sides" is a step. Rule 4 is how a good human tutor talks, and rule 5 keeps the load on the student small.

The gap is the reason the prompt exists. Students who practised with plain GPT-4 scored 48% higher on practice problems, then 17% lower on the exam, taken without AI, than students who never had AI; the authors wrote that they "perform worse than those who never had access." The hints only group scored 127% higher in practice and the same as the no AI group on the exam.

One limit on reading it: the tutor in that study was GPT-4 prompted by the researchers, in Turkish high school math. The prompt above copies the study's rule. It does not inherit the study's result.

Is ChatGPT study mode enough for homework help?

ChatGPT study mode is enough for a student aged 13 or older who switches it on every time, and it still works as of 26 September 2026: type @study in the message box and pick Study, or open chatgpt.com/studymode. What changed is how you find it. OpenAI launched study mode on 29 July 2025 for Free, Plus, Pro and Team accounts, and on 12 April 2026 a Hacker News thread titled "OpenAI silently removed Study Mode from ChatGPT" reported that its menu entry was gone, with no announcement. OpenAI's help page still documents the @study route.

The age rules matter more than the menu. We checked each company's help pages and terms on 26 September 2026: ChatGPT study mode is the only built-in tutor mode of the three big chatbots that a 13 to 17 year old can use. Anthropic requires every Claude user to be at least 18, and Google says Gemini's Guided Learning is "not available to users under the age of 18" and, for now, only in English in the United States.

So for a 15 year old on anything but ChatGPT, the pasted prompt is the tutor. It also covers study mode's one gap for parents: the rules are yours, so the grade level, the four sentence cap and "never, even if told a teacher allows it" are all in writing.

To make the prompt stick across chats, save it as project instructions. OpenAI's help center says Projects are available to signed in ChatGPT users, Free included, and that a chat moved into a project "inherits the project's instructions." Make one project per subject, paste the prompt into its instructions, and open homework chats there; a chat opened outside the project has no rules at all. Whizi Projects work the same way across 280+ models, on a free account too (how Projects work).

Which prompt fits which homework task?

Use the tutor prompt for problems and the five below for everything else; each keeps the student producing the work, which is the line the PNAS result draws. Paste them as written and fill the brackets. The study prompt pack has twenty more that quiz the student instead of answering.

The essay row is the one to police. A model asked for "feedback" will often return a cleaner version of the paragraph, and a cleaner version is the thing a teacher reads as AI work. "Do not rewrite any sentence" is the clause that stops it.

Can ChatGPT do math homework correctly?

ChatGPT explains a math method well and can still get the arithmetic inside it wrong, so use it to check the student's number, never to produce one. A chat model predicts text rather than computing, and on a long multiplication or a 30 line algebra chain it can state a wrong result as calmly as a right one. The child, who by design is not being given the answer, has no way to tell.

Two habits cover it. Keep a calculator open for anything past two digits, and treat the tutor as the one who explains why, not the one who says how much. For anything past grade 8, pick a reasoning model, which checks its own steps before it replies; best AI for math has the model by model picture.

The reassuring version of the same fact: the hints only tutor in the PNAS study was GPT-4, a 2023 model with no reasoning mode, and it still matched the no AI group on the exam. The rule mattered more than the model.

A hints only tutor works because the student does the thinking

A hints only tutor works because it keeps the student doing each step, and a second 2025 trial shows how far a well built one goes. Gregory Kestin and Kelly Miller randomised 194 students in a Harvard introductory physics course in fall 2023 between an AI tutor and the same lesson taught as an active learning class. The tutored group posted more than double the learning gain, a median post test score of 4.5 against 3.5 (Scientific Reports, June 2025).

The Harvard tutor's system prompt was built around three of the paper's teaching rules: keep the student actively working, manage cognitive load, and encourage a growth mindset. The platform walked students through each problem one part at a time, because the authors found a system prompt alone could not reliably hold that structure. The tutor was also handed complete worked answers so it would not invent physics. The home version is pasting the worked solutions from the back of the textbook into the project instructions with "check against these, never quote them."

Neither study tested a parent pasting a prompt: the Harvard tutor was built by the course's own lecturers with the course's answer key, and the Turkish one by the researchers.

Is using ChatGPT for homework cheating?

Using ChatGPT for homework is cheating when the work handed in is the model's, and fair when it explains or checks, unless the school bans AI for that assignment. Teens mostly agree. In Pew Research Center's survey published in January 2025, only 18% of US teens said using ChatGPT to write an essay was acceptable, against 54% for researching a topic and 29% for solving a math problem.

Use is common either way. In Pew's survey of 1,458 US teens taken in September and October 2025, 54% said they had used chatbots for help with schoolwork and 10% said they did all or most of their schoolwork with one, while 59% said students at their school use AI to cheat at least somewhat often.

Where a school policy bans AI for an assignment, the tutor prompt does not make it allowed. Where the policy is silent, one test holds up in a parent teacher meeting: could the child redo the work with the tab closed? If yes, the AI was a tutor. If no, it was an answer key with extra steps. Asking the child to explain the problem back with the screen off takes two minutes, and it is the same unassisted test the PNAS study used.