How our questions are made
Our questions are written by AI. We think telling you that plainly, and then showing you exactly what happens before a question reaches you, is worth more than pretending otherwise.
A practice exam is only worth anything if you can trust the answer key. Here is what a question has to survive. Where else AI is used — and where it never is — is on how we use AI.
How are Preptics SAT questions checked before publication?Link to the section: How are Preptics SAT questions checked before publication?
Seven gates, and failing any one discards the question — there is no publish-anyway route. Three AI models solve it blind with the key removed, arithmetic is executed as code rather than trusted, and a person reviews the question and its key last. During the beta that human review is still catching up, so some live questions have passed only the six automated gates.
Blueprint
Nothing is written freehand. Generation follows the College Board's published description of the Digital SAT: its content domains and skills, how much each is weighted, and the question formats that actually appear. The source material is public — never a real or recalled test question, and never another company's question bank.
Structure and style
Deterministic code, no AI. It rejects malformed questions, wrong option counts, missing rationales, stems that are too long or too short, catch-all options like "all of the above", and — importantly — any question where the correct answer is conspicuously longer than the alternatives, because that is a question a test-wise candidate can answer without knowing the material.
Blind solve
The question is given to three AI models with the answer key removed, and they have to solve it cold. At least half must reproduce the intended answer. If none can, the question is ambiguous and it is discarded; if they all agree on a different answer, the key is probably wrong and it is discarded. One limit we want you to know: the three solvers are currently all Anthropic's Claude models, so they can share a blind spot — one reason the last check is a person.
Arithmetic executed, not trusted
A Math question whose answer comes from a calculation carries a machine-checkable expression that is run as code. Language models are unreliable at arithmetic, so we do not ask one to check another one's sums. If the computed value disagrees with the question, it is discarded outright.
Adversarial critique
A separate model is instructed to attack the question: is any wrong answer actually defensible? Is the key actually wrong? Could it be answered from surface cues alone, without knowing the skill? Has anything it depends on gone out of date? Any hit sends the question back — and a critique that cannot be read counts as a failure, never as a pass.
Duplicates and fairness
Each question is fingerprinted and compared against the whole existing bank, so it does not slowly fill with the same question reworded. A separate pass rejects region-locked assumptions — units such as miles, pounds or Fahrenheit, and US-only institutions — because students sit the SAT worldwide.
Human review
A person reads each question with its answer key, every option's rationale and any graph or table it depends on, and accepts it, sends it back or rejects it — the founder, or an SAT tutor reviewing with them. A rejected question leaves circulation the moment the decision is recorded. During the beta this review is still under way: some questions students meet have passed only the six automated checks above, which is why every question has a report button, and three open reports take a question out of circulation automatically.
A high rejection rate is the point, not a problem. We would rather generate five questions and publish two than pad a bank with questions that teach you the wrong thing.
What happens when a practice question turns out to be wrong?Link to the section: What happens when a practice question turns out to be wrong?
Every question carries a report button, and three open reports take a question out of circulation automatically — no one has to be awake for it. Affected exams are then rescored with the question excluded, which can never lower your score. The gates catch most errors but not all of them, and claiming otherwise would be the dishonest part.
Every question has a report button
One tap during review. You do not need to explain yourself, though it helps if you do.
Reported questions are pulled automatically
Once a question crosses a report threshold it leaves circulation immediately — no human has to be awake for that to happen.
Scores are recalculated, never lowered
If we remove a question, affected exams are rescored with it excluded. Removing a bad question can never reduce your score.
Statistics find what people miss
When strong candidates get a question wrong more often than weak ones, that is the classic signature of a wrong answer key. We flag those automatically.
Are these real SAT questions?Link to the section: Are these real SAT questions?
No. Every question on Preptics is an original work written for this site, and we never reproduce real SAT questions or passages. Sites publishing 'real' or 'recalled' questions are copying the College Board's copyrighted work. It is also bad practice: a bank of memorised questions trains recall, and the form you sit will not contain any of them.
Sites that publish "real" or "recalled" SAT questions are reproducing the College Board's copyrighted work, usually gathered by students who agreed at the test centre not to disclose it.
That is their legal problem. Yours is different and more practical: a bank of memorised questions teaches you those questions, and the test you sit will not contain them. Every form is new. Practising on leaked material trains recall when the SAT is testing reasoning, which is the most expensive way to waste a month.
Every question here is an original work written from public documentation. That is also why we can afford to be this open about our process — we have nothing to hide about where the content came from.
What makes a practice exam honest rather than just a quiz?Link to the section: What makes a practice exam honest rather than just a quiz?
The details that a quiz skips. Module clocks are held on our server, so closing the tab does not pause them and reopening does not reset them. The second module is chosen by how you did on the first, the break is a real ten minutes, and Math answers can be typed rather than picked from four options.
The answer key never reaches your browser
Questions are sent one at a time with the key stripped out. You could read every byte your browser receives and not find the answer.
The clock is on our server
Changing your device time, or opening the developer console, does not buy you a second. The exam is timed the way the real one is.
Option order is different for everyone
Two people sitting the same exam see the options in a different order, so a shared answer key is worthless.
Rankings are earned
Ranked attempts are limited and scored under consistent conditions. We would rather have a small honest leaderboard than a large meaningless one.
Spotted something we have got wrong? Tell us — including about this page.