Back to Blog

    Should you use ChatGPT for SAT prep?

    By Kim Strauch··Updated ·9 min read
    Should you use ChatGPT for SAT prep?

    Key takeaways

    • High school students with unrestricted GPT-4 access did 17% worse on a later exam than students who never had it.
    • A version that gave hints instead of answers largely removed that harm, though it didn't raise exam scores.
    • In the study's own check, GPT-4 answered its practice problems correctly only about half the time.
    • College Board's testing rules bar phones, smart glasses and any device that gives access to AI during the test and breaks.

    Yes, if you use it after you've tried a question, to get a hint or to understand a mistake. Used to get answers, it can leave you worse off than if you'd never opened it. The clearest evidence comes from a randomized experiment with nearly 1,000 high school math students, and whether it hurt students came down to how the AI was set up to respond.

    What happened when students practiced with GPT-4?

    In fall 2023, researchers from the University of Pennsylvania ran an experiment at a large high school in Turkey, published in 2025 in the Proceedings of the National Academy of Sciences. Across four 90-minute sessions, a teacher reviewed a topic, students worked practice problems, and then they took a closed-book, closed-laptop exam on problems that closely matched the practice.

    During practice, students were randomly assigned to one of three conditions. Some had a ChatGPT-style interface running GPT-4 ("GPT Base"). Some had a version prompted to give hints and step-by-step guidance instead of answers ("GPT Tutor"). The rest had only their textbook and notes.

    With AI available, practice scores rose: 48% for GPT Base and 127% for GPT Tutor compared with students who had no AI. On the exam, without AI, the GPT Base students scored 17% lower than students who had never had access. The authors found that students often used GPT Base as a crutch; the most common message students sent it was a request for the answer. The tutor version, according to the paper, essentially eliminated the harm, though its students didn't outscore the no-AI group on the exam either.

    Two other findings from the same study bear on how you use AI. First, the GPT Base students did not think they had performed worse or learned less. Second, when the researchers asked GPT Base for the answer to each of the 57 practice problems ten times, it gave a correct answer only 51% of the time, with logical errors in 42% of responses and arithmetic errors in 8%. That was GPT-4, the model students used in 2023, and the study didn't test newer models. A wrong explanation that sounds confident is still a risk with any chatbot, on math and on Reading and Writing questions alike, so treat its answers as claims to check.

    The GPT Tutor prompts were built with teachers' help. For each practice problem, teachers supplied correct solutions and the common mistakes students make, along with how to give feedback on them, which the authors say kept GPT-4 from giving incorrect feedback.

    Do students use AI tutors when they have them?

    Often, not much. In a two-year randomized trial in 18 Tennessee middle schools, released as an NBER working paper in August 2026, Philip Oreopoulos and Nina Low gave students in remedial math periods access to Khan Academy with Khanmigo, its AI tutor, set up to coach instead of giving answers. Math achievement rose by about 1.3 national percentile ranks per term, gains similar to those from Khan Academy practice without AI.

    The tutor itself saw little use. The authors measured usage in the trial's second year, among 563 students at 16 schools who had access to Khanmigo, activated their accounts and could be matched to practice records. Ninety-six percent of them messaged Khanmigo at least once, but the median student messaged it on 33% of the days they practiced, and roughly one message in seven contained a math question or a step of reasoning. The authors read their results "primarily as the effect of structured, individually targeted practice."

    A research summary from Stanford's SCALE Initiative and the National Student Support Accelerator (August 2026) reaches a similar conclusion. It reports that in one study, 40 to 47% of students given an AI tutor never logged in, and average use was two to five minutes a week. Those figures match a June 2026 Stanford working paper on elementary students given an AI literacy platform and scheduled time to use it, in which nearly half of the students working on their own never used it. The strongest results in the summary came from AI built for human tutors. In the Tutor CoPilot trial, with 900 tutors and 1,800 K-12 students, students whose tutors had an AI suggesting what to say during sessions were 4 percentage points more likely to master topics, and students of lower-rated tutors gained 9 percentage points.

    Studies of AI tutors and what they found
    StudyWhoWhat it testedWhat it found
    Bastani et al., PNAS (2025)Nearly 1,000 students at one high school in TurkeyGPT-4 with no limits vs. a hint-giving tutor version vs. no AI, during math practiceUnlimited GPT-4: 48% better in practice, 17% worse on the exam. Tutor version: 127% better in practice, no harm on the exam
    Oreopoulos and Low, NBER (2026)Students in remedial math in 18 Tennessee middle schools, over two yearsKhan Academy with the Khanmigo AI tutor vs. usual remedial workAbout 1.3 percentile ranks per term; students rarely used the tutor; gains attributed mainly to structured practice
    Robinson et al., EdWorkingPaper (2026)Elementary students in two districts, more than 80% economically disadvantagedAn AI literacy tutor, alone or with an in-person tutor encouraging useNearly half of students working alone never used it; human support raised use but not reading scores
    Wang et al., Tutor CoPilot (2024)900 tutors and 1,800 K-12 studentsAI suggestions for human tutors during live sessionsStudents 4 percentage points more likely to master topics; 9 points with lower-rated tutors

    Sources: linked in the table. Checked October 30, 2026.

    None of these studies looked at the SAT, and the students differ from a high schooler prepping on their own. Together they point the same way: AI helped most when it made students do the thinking, and the Khanmigo authors credit their gains mainly to structured practice.

    How should you use ChatGPT for SAT prep?

    Treat it the way the GPT Tutor version was designed: as something that withholds answers until you've done the work.

    Try the question first. Work it fully and commit to an answer before you open a chat. If you're stuck, give yourself a few more minutes; that effort is part of the practice.

    Ask for a hint, not the answer. Paste the question and tell it what you want:

    Don't tell me the answer. Give me one hint about where to start.

    Ask why your wrong answer was tempting. Knowing which mistake led you to a wrong choice is the most useful thing to take from a missed question.

    I picked B. The correct answer is D. Explain what made B look right and what I missed.

    Redo the question without help. Close the chat and work the problem again from the start. Then find a similar question and do that one cold. That is the test the PNAS study gave, and it's how you find out whether you learned it.

    Check explanations against the official rationale. College Board's Educator Question Bank lets you view answer rationales for official questions without signing in, and after each Bluebook practice test, My Practice shows an answer explanation for every question. When the AI and the rationale disagree, trust the rationale.

    Don't use it to write anything you'll be tested on. Having it choose your answer to a Reading and Writing question, or solve a Math problem you then copy, is the answer-copying pattern the PNAS authors linked to lower math exam scores. Don't ask it to generate SAT practice questions either: nothing checks that they match the test's content or difficulty, while official questions come from the people who write the test. We've written about what AI does well in SAT prep and what a tutor provides.

    Track the mistakes that repeat. One missed question tells you little. The same error across five questions (misreading what a question asks for, or dropping a negative sign) tells you what to practice next. Keep a running list, and bring the patterns, not single questions, to the AI. Our free Bluebook score report breaks a practice test down by skill, which makes those patterns easier to spot.

    Sharp's AI tutor starts from the mistake: when you miss a question, it explains the misconception behind your answer and how to approach similar questions next time, using teaching guides written by experienced tutors.

    Where should your practice questions come from?

    Official College Board material remains the reference for what the test looks like. We've written about which SAT prep resources to use for each stage. If you haven't started yet, we've also written about how to begin with a diagnostic test and a plan.

    Use full-length Bluebook practice tests as checkpoints and the question banks for practice by skill, and bring AI in to understand the questions you miss there. The findings above fit with the broader research on SAT prep, where practice that makes you retrieve and apply what you know has the strongest support. AI earns a place in that routine when it sends you back to the question instead of past it.

    Can you use AI on the SAT?

    No. College Board's SAT testing rules bar you from using phones, smartwatches, smart glasses or other eyewear that gives access to the internet or to any AI feature or service at any point during testing, including breaks. The same goes for other electronic devices beyond your approved testing device and calculator. The rules say any attempt to gain an unfair advantage, including with unauthorized smart devices, will lead to consequences that include score cancellation. On test day you'll have Bluebook's calculator and reference sheet but no AI, so practice that way.

    The point of practice is getting the next question right on your own, and Sharp's practice and AI tutor are set up to build that habit.

    Kim's take

    The part of this research I find most useful is the finding that students who leaned on ChatGPT didn't think they had learned less. Feeling fluent while you practice and being able to do the problem alone are different experiences, and you can't tell them apart from the inside. The only reliable check is to close the chat and try a similar question cold. If you can't, the AI did the learning that session.

    Kim Strauch

    Kim Strauch, SAT Tutor & Founder

    Kim scored a perfect 1600 on the SAT, graduated summa cum laude from Dartmouth, and founded Sharp after years of tutoring students one-on-one.

    How we write and fact-check our posts

    Frequently asked questions

    Is it cheating to use ChatGPT for SAT prep?

    No. Using AI to study for the SAT is allowed, and it can help if you use it for hints and explanations after you've tried a question. Using AI or any device that accesses it during the test or on breaks is against College Board's rules and can get your scores canceled.

    How accurate is ChatGPT on SAT math?

    It can be wrong while sounding certain. In a study run in 2023, GPT-4 gave the correct answer to the researchers' high school math problems only about half the time. Newer models weren't tested, so check every AI explanation against the official answer and rationale.

    Should I ask ChatGPT to make SAT practice questions?

    No. Nothing checks that AI-generated questions match the SAT's content, difficulty or wording. Use College Board's Bluebook tests and question banks for practice questions, and use AI to understand the ones you miss.

    Does using AI while studying make you worse at the SAT?

    It can. In a randomized study of nearly 1,000 high school students, those who had unrestricted GPT-4 during practice scored 17% lower on a later exam than students who never had it. Students using a version that gave hints instead of answers didn't show that drop.

    What's the best way to use ChatGPT after missing a question?

    Ask it to explain why your answer looked right and what you missed, without solving the problem again for you. Then close the chat and redo the question on your own, followed by a similar one. If you can't do the second question alone, you haven't learned it yet.

    Sources

    Sign up for our newsletter

    SAT strategy and updates from Sharp, by email.