An interview is an exam with no AI in the room. If the way you prepare is to ask a chatbot for good answers and read them, you are practising the part the interview will not test.
ChatGPT has a setting that changes this. In Study mode it asks you questions and waits for your attempt before it explains. It is on every plan, the free one included. Study mode is made by OpenAI, not by webridgeAI.
The setup, in three steps
- Turn on Study mode. In an ordinary chat, select +, then Study. On the web you can also type
@study, or open chatgpt.com/studymode. It does not appear in Temporary Chats, GPTs or Projects. - Paste the job description. Add the role, the round you are preparing for (screening, technical, behavioural, final) and how many years of experience you have. OpenAI's own advice is to give your level, the exact topic and what you are preparing for.
- Make it quiz you, one question at a time, with no answer until you have tried. This is the step most people skip, because reading a model answer feels like progress.
The prompts
Paste the job description first. Then use these as they are.
Start the session
I have an interview for the role above. This is the [screening / technical / behavioural / final] round.
Quiz me one question at a time, the way this interviewer would.
Do not show a model answer until I have given mine.
After each answer, tell me what was strong, what was missing, and ask one follow-up.A behavioural round
Ask me behavioural questions this employer is likely to ask, one at a time.
After I answer, check it against situation, task, action and result.
Tell me which of the four was thin, and ask me to say that part again.A technical round
Ask me technical questions at the level this job description implies, one at a time.
If I am stuck, give one hint, not the solution.
Give a second hint only if I ask. Show the full answer only after my final attempt.Grade my answer
Here is my answer to "[question]":
[your answer]
Score it out of 10 as this interviewer would. Give the two changes that would raise the score most.
Do not rewrite it for me. I will rewrite it and send it back.What a weak answer looks like
For the question "[question]", describe what a weak answer and a strong answer have in common and where they differ.
Do not write either answer out in full. Then ask me the question.The night before
From everything in this chat, list the three questions I answered worst.
Ask me those three again, one at a time, with no hints.Why hints and not answers
In 2025 a team led from the University of Pennsylvania published a field experiment in the Proceedings of the National Academy of Sciences. Nearly a thousand high school students in about fifty classes practised maths in one of three ways: with a chat interface like ordinary ChatGPT, with a tutor version told to give hints and not answers, or with only their notes and textbook.
| Group | Practice, with the tool | Exam, without it |
|---|---|---|
| Ordinary chat | 48% higher than the group with no AI | 17% lower than the group with no AI |
| Hints only | 127% higher than the group with no AI | About the same as the group with no AI |
The most common message students sent the ordinary chat was "What is the answer?" The authors describe this as using the tool as a crutch. Students in that group did not think they had learned less.
What this does not show
- It is one study. One school, one subject, one country, during the 2023 to 2024 school year.
- It studied teenagers learning maths, not adults preparing for interviews. Applying it to an interview is our reading, not the authors' finding.
- Study mode was not what the researchers tested. They built their own tutor. Study mode follows the same idea, questions before answers, and OpenAI has not published a comparable trial of it.
- The hints group did not beat the group with no AI on the exam. It avoided the damage. The gain is in the practice itself, so the practice still has to be yours.
- Study mode can be wrong. OpenAI says so. Check any fact you plan to repeat in an interview.
Sources
- Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman, "Generative AI without guardrails can harm learning: Evidence from high school mathematics", PNAS 122(26), 2025. Full text read on PubMed Central on 11 October 2026.
- OpenAI Help Center, "Using study mode in ChatGPT", read on 11 October 2026.