Homework help that gives hints instead of answers

What the research found

Two large studies, run in very different ways, point the same direction. When students used AI to get answers, their practice or homework scores went up and their exam scores went down. When the AI was set up to give hints, or when students kept doing the work themselves, most of that loss went away.

A field experiment with about 1,000 high school students

Bastani and colleagues ran a randomized trial at a high school in Turkey in fall 2023 and published it in PNAS in 2025. Nearly 1,000 students in grades 9 to 11 took four 90-minute math review sessions. Each class was assigned to one of three groups for the practice part of every session:

• GPT Base: a chat tutor built on GPT-4 that worked like ChatGPT.
• GPT Tutor: the same chat with a prompt written with teachers. It was told to give hints without giving the answer, and it had the correct solution and a list of common mistakes for each problem.
• Control: the textbook and class notes only.

After practice, everyone took a closed-book exam without AI. Each exam problem was very close to one of the practice problems.

The researchers looked at what students typed. In the GPT Base group, students most often asked for the answer and copied it. In the GPT Tutor group, they more often asked for help or tried an answer first. Students in the GPT Base group didn't think they had learned less, even though their exam scores showed it.

Two more details matter if you use a general chatbot for math. When the researchers asked GPT Base “What is the answer?” ten times for each of the 57 practice problems, it gave a correct answer only 51% of the time. And the hint-only tutor removed the harm without raising exam scores above the control group. Knowledge at Wharton has a plain-language summary of the same results.

The authors name the study's limits themselves: one school, one subject, short-term outcomes, and AI models from 2023.

30 months of school records for 26,811 students

A June 2026 CEPR discussion paper by David Strömberg, Victor Lei and Yanhui Wu followed 26,811 students in grades 7 to 12 at nine schools in one county in central China. It combined 30 months of school records (weekly homework scores and completion times, monthly closed-book exams, and the high school and college entrance exams) with a June 2025 survey that asked each student when they started using AI. It then compared students before and after they started, against students who hadn't started yet. Because students chose when to start, this is an observational study, so it can't rule out every other explanation the way a randomized trial can. The authors run several checks on their design.

What they found:

• Six months after starting, homework scores were 18% higher and homework took about 30% less time, falling from 64 to 45 minutes per assignment.
• Over the same six months, scores on monthly closed-book exams fell by 20%.
• The entrance exams fell more slowly. After about two years, scores were 18% lower on the Gaokao, the college entrance exam, and 24% lower on the Zhongkao, the high school entrance exam.
• The losses were concentrated among about 80% of AI users whose homework looked outsourced: done much faster than any non-user, with very high scores. AI users who kept spending as much time on homework as non-users had only small losses.

Most students in the study used general chat apps that give quick answers. Fortune covered the paper on July 21, 2026.

What tutoring research says about help at each step

Older research on tutoring points the same way. In a 2011 review in Educational Psychologist, Kurt VanLehn compared human tutors, computer tutors and no tutoring across many experiments. Compared with no tutoring, human tutors had an average effect size of 0.79, and computer tutors that give hints and feedback on each step of a problem averaged 0.76. Tutors that gave feedback only on the final answer averaged 0.31 (VanLehn's summary table). Help on each step did much better than help on the answer alone.

Research on productive failure suggests that trying first helps too. A 2021 meta-analysis by Tanmay Sinha and Manu Kapur pooled 53 studies that compared students who tried problems before being taught with students who were taught first. Trying first had a moderate advantage (Hedges' g = 0.36). The authors note one exception: for younger children, in grades 2 to 5, being taught first worked better.

How to use AI for homework without it doing the work

The common thread in these studies is effort. Students who kept doing the thinking did about as well as students without AI. Students who let the AI do the thinking did worse, and in the PNAS study they didn't notice. These habits keep the thinking with you.

1. Try the problem first. Write every step by hand before you ask for anything. A wrong step is useful, because it shows exactly where the idea broke.
2. Ask about one step. “Is my third line right?” or “What should I check in this step?” gets you help where you need it. “What's the answer?” skips the part that teaches you.
3. Ask for one hint at a time. Tell the tool not to show the solution. Make the next step yourself before you ask again.
4. Check the answer yourself. Substitute it back into the original problem. The guide to checking handwritten math step by step shows how.
5. Watch your time. If AI cuts your homework time sharply while your scores jump, that's the pattern the CEPR paper linked to lower exam scores.
6. Redo a similar problem with no help. Both studies measured learning with closed-book exams, and a fresh problem with nothing open is the quickest way to find out whether you've got it.
7. Follow your class's rules. Some teachers allow AI for hints and some don't allow it at all. Ask when you're not sure.

What to look for in a homework help tool

• It works from your steps. A tool that sees your work can point at the line that went wrong. A tool that only sees the question can explain or solve it, but it can't tell you where your own work slipped.
• It gives hints and holds back the answer. In the PNAS study, the tutor told to give hints without the answer was the version that didn't hurt exam scores.
• It knows the right answer. GPT Tutor had each problem's solution in its prompt. GPT Base, asked for the answer, was right 51% of the time, and a hint built on a wrong solution can send you the wrong way.
• It leaves the next step to you. A question like “What should the 2 multiply?” points at the slip and still makes you do the fixing.
• It's clear about your data and age rules. Read the privacy policy before you upload schoolwork.

ChatGPT has a study mode for this. OpenAI's help page says it can guide you with questions instead of giving the answer right away, and that it can make mistakes and sometimes gives a direct answer. The comparison with ChatGPT study mode goes through the differences, and what to look for in an AI math tutor app for iPad has a fuller checklist.

How alai approaches it

alai is an iPad and web notes app with Ally, an adaptive AI tutor that teaches alongside you. In Homework help, you write each step yourself, and Ally doesn't interrupt while your steps are right. When a step goes wrong, Ally points at the first wrong line and gives one short hint about it, usually a question, in a single sentence. Ally is set up not to say the corrected step or the final answer while you work. If you stay stuck, the next hint gets smaller and more concrete.

For example, under 3(x + 4) = 2x + 19, someone writes 3x + 4 = 2x + 19. Ally underlines the +4 and asks, “Did the 3 reach every term inside the brackets?” Then Ally waits while you fix it.

When you bring a worksheet, alai reads the page, finds the problems and works out an answer key for each one, which Ally uses when checking your steps. Ally also remembers which mistakes you've made before and which hints helped you fix them. Ally can still misread messy handwriting or get something wrong, so it's worth checking your final answer yourself.

alai is for students 13 and older in the US (a parent sets it up for a child under 13), and it has a free plan. How alai works walks through a session from the first line to Ally's question.