How we use AI
Every place AI is used on Preptics, which model, who runs it, exactly what is sent, what is kept, and the rules it works under. If any of this stops being true, this page changes before the product does.
At a glanceLink to the section: At a glance
Four uses. Two of them involve your results; two never touch anything about you.
Your practice-test debrief
- Where
- The results page, after you submit a practice test
- Model and who runs it
- DeepSeek V4 Flash, run by zero-data-retention hosts through OpenRouter
- What is sent
- A summary our software has already calculated: score ranges, accuracy by difficulty, the skills that cost you points, pacing, kinds of mistake, reviews due. No identifier, no questions, no answers.
- What is kept
- The summary and the written template, 90 days
- Limits
- 6 a day, 20 a week per account
Highlight to explain
- Where
- Lessons, drills and the review after a practice test — never during one
- Model and who runs it
- DeepSeek V4 Flash, run by zero-data-retention hosts through OpenRouter
- What is sent
- The phrase you highlighted, up to 400 characters
- What is kept
- Nothing — not the phrase, not the answer
- Limits
- 20 a day, 60 a week per account
Writing and checking our questions and lessons
- Where
- Offline, before anything is published
- Model and who runs it
- Anthropic's Claude models
- What is sent
- Draft questions, lessons and our instructions — never anything about a student
- What is kept
- What passes the checks becomes our content
- Limits
- Every question passes seven checks, the last one a person
Search index of our own content
- Where
- Offline; not yet used by any feature you can see
- Model and who runs it
- OpenAI's text-embedding-3-small, through OpenRouter under zero data retention
- What is sent
- Our published explanations, lesson cards, lesson questions and help articles
- What is kept
- The resulting search vectors, in our database
- Limits
- Nothing about you is involved
The hosts eligible for our requests, and what each company receives, are listed on third-party services.
The rulesLink to the section: The rules
These hold for every use above, and for anything we add.
AI never calculates your results
Scores, ranges, skill estimates, review timing and your study plan are all calculated by our own software. In the debrief, the model writes wording with placeholders, and your figures are filled in from our calculations. A draft that states a number of its own is rejected.
Never inside a timed test
The exam runner does not contain the explain feature at all, and the debrief exists only after a practice test is submitted. That is structural, not a setting.
It does not decide anything about you
No AI is involved in your score, your leaderboard eligibility, integrity checks, account decisions, which questions you are served or how your plan is ordered. Appeals are read by a person.
Only providers that keep nothing
Every request carries a zero-data-retention requirement: the host may not store what we send or train on it. If no host meeting that requirement is available, the request is not sent anywhere else — you get our standard wording.
Who you are never leaves
Requests carry no name, email address, account number or other identifier. The debrief summary contains numbers and skill names; the explain request contains the phrase you chose.
What it may not say
The debrief is rejected if it talks about passing or failing (the SAT has no pass mark), compares you with other students, predicts a test-day score, or promises an outcome.
Limits are checked before sending
The per-account limit is counted before a request leaves. If the count cannot be checked, the request is refused rather than let through.
Labelled where you see it
A debrief written by the model says "drafted by AI". One written from our standard wording says "Written from your own results".
How a debrief is writtenLink to the section: How a debrief is written
Most debriefs never reach a model at all. Here is the whole path.
Our software builds a summary of your results — numbers and skill names, nothing that identifies you.
If exactly the same summary has been written up before, the stored wording is reused and no request is sent.
If another student's results were in the same situation — the same section held back by an easier second module (or neither), the same most common kind of mistake, and reviews due or not — and very close in overall shape, that wording is reused with your own figures filled in. A piece of wording can be reused at most 200 times, and it never contains anyone's figures, only placeholders.
Otherwise your limit is counted, and the summary is sent to the model with a strict format to answer in.
The answer is checked: every placeholder must match your own results, and no number or forbidden kind of statement may appear outside them. A failing draft gets one retry that names the problem; a second failure falls back to our standard wording.
The accepted wording is stored for 90 days, so reopening the page does not send another request.
Reusing wording across students is deliberate: it keeps the feature fast and affordable without a limit so low it stops being useful. It is safe because the stored wording holds placeholders, never figures — what one student reads cannot contain another student’s results. When you delete your account, the summary written from your results is deleted; wording that was reused stays, because nothing in it is about you.
Why this model, run this wayLink to the section: Why this model, run this way
A debrief should arrive while you are still on the results page, and an explanation while you are still looking at the phrase. DeepSeek V4 Flash is fast and inexpensive enough to do both within the limits above.
We do not send requests to DeepSeek’s own service: every request excludes it by name. Requests go through OpenRouter, which routes each one only to hosts that run the model under a zero-data-retention commitment. A model is only as private as the place it runs, so we chose the place first.
Questions written with AILink to the section: Questions written with AI
Our questions and lessons are written with the help of Anthropic’s Claude models, and we would rather say so than have you guess. A question is published only after it passes seven checks: structure, a blind solve by three models that never see the answer key, arithmetic run as code, an adversarial critique, duplicate and fairness checks — and a person.
One limit we want you to know about: the three blind solvers are all Claude models today, so they can share a blind spot. Human review and the report button exist partly for that reason. How questions are made has the details.
Something wrong?Link to the section: Something wrong?
If a debrief or an explanation is wrong, misleading or just unhelpful, tell us at support@preptics.com or through feedback. A person reads it.
A study assistant that can answer questions about your own practice is planned. It is not built, and when it is, it will appear on this page — with what it sends and keeps — before it appears in the product.
SAT® is a trademark registered by the College Board, which is not affiliated with, and does not endorse, this site. All practice questions and reading passages are original works created by Preptics and are not actual SAT® questions.