What the AI features do, and what they must not be trusted to do.

Two features generate text with a model: the tutor, and feedback on a practice essay. Both are Pro features, both run only when you start them, and both run through the ChatGPT account you connect yourself — there is no background generation and nothing is generated about you while you are away.

Everything else on the site is ordinary software. Answer checking compares your response to the stored answer key with fixed rules. Progress figures and review scheduling are computed from your own recorded attempts. Published study pages are written records with cited sources, not text produced by a model on request.

The server prompt asks the tutor to teach rather than answer: a guiding question or a hint first, the full solution once you are stuck or ask for it, reasoning shown, and honesty about uncertainty instead of invented exam facts.

It is also instructed to treat your messages as material to help with rather than as instructions that change those rules, to refuse requests to cheat on a live exam, to decline unrelated requests, and never to claim that a conversation is saved or that quota is unlimited.

These are instructions to a model, not guarantees. A model can ignore an instruction, and a determined prompt can push it off course. Treat the tutor as a study partner that is often useful and sometimes wrong.

It cannot score you. Essay feedback is deliberately structured by rubric dimension — task response, development, organisation, control of language — and is instructed never to give a numeric score, a score range, a percentile, or an admissions prediction, because generated grading is not calibrated to official raters.

It cannot decide anything about your account. No generated output affects grading, progress, billing, entitlement, or any other decision the application makes about you. There is no automated decision-making based on model output.

It has not been benchmarked. We have not published an evaluation baseline, quality metrics, or a safety threshold for these features, so we make no accuracy claim. Check anything that matters against official exam materials.

It is not human expert review. A generated explanation is not a subject-matter expert's sign-off, and a page or answer is never marked as expert-reviewed on the strength of model output.

Generation runs through your own connected ChatGPT account. Your prompts, attached context, and the generated response are processed by OpenAI under the account terms you agreed with them. We have not verified OpenAI's retention, training use, processing locations, or abuse monitoring, and we do not represent them here — read their terms before sending anything sensitive.

Practically: do not paste passwords, payment details, identity documents, health information, or another person's personal data into the tutor, an essay, or a support message.

You can disconnect ChatGPT from the tutor page at any time. Disconnecting stops future requests; it does not by itself delete what a provider already holds, and exporting or deleting a saved chat here does not disconnect the account.

If the tutor gives harmful advice, states an exam fact that is wrong, reveals something it should not, or behaves in a way this page does not describe, report it through the contact form or support@keiko.study. Tell us the exam and roughly what you asked; please do not paste credentials or another person's data into the report.

We keep an internal record of AI issues. We will not tell you an issue is closed on the strength of a code change alone.

This page is reviewed on the schedule above, and again whenever the model catalogue, provider, help modes, prompts, retained fields, or intended audience change. If the behaviour you see does not match this description, that is a correction we want to receive.