fib(50) never finishes. It asks you why, then draws every call and rings the ones it repeats.In a 2025 field experiment in a high school, students who practiced math with a plain AI chat did better during practice, then 17% worse on the exam once the AI was taken away, compared with students who never had it. A version built to give hints instead of answers mostly avoided the harm. DuckyHelper is built the second way: it guides you to the answer.
“unfettered access to GPT-4 can harm educational outcomes”

It waits for you.
Why does the river bend? It asks where the water is fastest, lets you answer, then draws what you worked out.
What the research found
Hamsa Bastani and colleagues ran a field experiment with high school math students. Some practiced with "GPT Base", a plain chat interface; some with "GPT Tutor", the same model with prompts designed to protect learning; and some with no AI. During practice, both AI groups did better: grades rose 48% with GPT Base and 127% with GPT Tutor.
Then the AI was taken away for the exam. Students who had used GPT Base did worse than students who never had it, a 17% reduction in grades. In the authors' words, “unfettered access to GPT-4 can harm educational outcomes.”
The guarded version mostly avoided this. In the authors' words:
These negative learning effects are largely mitigated by the safeguards in GPT Tutor
(Bastani et al., 2025)
The authors' explanation is that without guardrails, students used the AI as a crutch during practice and then could not do the work alone. Other research points the same way. Language models solve math well but tend to give the solution away too early when they tutor:
While models like GPT-3 are good problem solvers, they fail at tutoring
(Macina et al., 2023a)
And long before chatbots, students who used a tutor's hints just to get to the answers ("gaming the system") learned much less:
students who frequently game the system score substantially lower on a post-test than students who never game the system
(Baker et al., 2004)
Feature by feature
Try first, then guide
On homework and anything graded, it does not just hand you the final answer. You take a first try, then it guides you one step at a time. Really stuck? It works through a similar problem with you, and then you do yours. In the app.
Hints, not answers
Stuck on a problem in the app? Ask for a hint. They come one at a time: a nudge, a bigger hint, the worked step, and the answer last. It holds the answer back until you ask for it a second time. In the app.
A nudge on a wrong answer
In a quiz, a wrong answer gets a one-sentence nudge, and a hint request gets a hint without the answer. In the app.
Steps you have reached, not the whole solution
On homework, worked steps stop where you are and add a hint for your next step, instead of the whole solution. On the blackboard beside your work, math is worked out step by step, not just announced. In the app.
Questions of fact still get answers
Guiding is for problems you are learning to solve. Ask a simple fact and you get a clear, short answer, plus one small thing to try next. In the app.
What this page does not claim These studies tested other people, other schools and other tools. DuckyHelper has not been tested in a study like these, so none of these results are DuckyHelper's results. They are the reasons behind how it teaches.
The papers on this page
Generative AI without guardrails can harm learning: Evidence from high school mathematics
Bastani et al., 2025. Proceedings of the National Academy of Sciences (PNAS). Unguarded AI help lifted practice grades, then cut exam grades 17%.Read the paperMathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems
Macina et al., 2023. Findings of EMNLP 2023. LLMs solve math well but reveal solutions too early when tutoring.Read the paperOff-Task Behavior in the Cognitive Tutor Classroom: When Students "Game the System"
Baker et al., 2004. CHI. Students who abused hints and feedback to get answers learned much less than students who did not.Read the paperPedagogical Alignment of Large Language Models
Sonkar et al., 2024. Findings of EMNLP. Trains LLMs with preference learning to guide students through problems rather than reply with the answer at once.Read the paperFrom Problem-Solving to Teaching Problem-Solving: Aligning LLMs with Pedagogy using Reinforcement Learning
Dinucu-Jianu et al., 2025. EMNLP. Uses reinforcement learning with simulated students to turn an LLM into a tutor that guides instead of giving away answers.Read the paperLearnLM: Improving Gemini for Learning
LearnLM Team, Google, 2024. arXiv (Google tech report). Training on teaching instructions made the tuned model the experts' pick for learning.Read the paperTutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise
Wang et al., 2024. arXiv preprint (v1). AI guidance for human tutors raised student topic mastery by 4 points.Read the paper

Try it on your own work
Press fn on a Mac (Ctrl+Shift+Space on Windows and Linux) and say it out loud. It sees your screen, so it works on your own problem with you, and shows you.
“Don't tell me the answer, just what to try next.”
DuckyHelper is an app for Mac (Apple silicon, macOS 14 or later), Windows and Linux.
Questions
Does using an AI chatbot for homework hurt learning?
It can. In Bastani et al. (2025), high school students who practiced with a plain AI chat did better during practice but 17% worse on the exam once it was taken away, compared with students who never had it. A version with learning safeguards largely avoided that harm.
Will DuckyHelper ever just tell me the answer?
For a plain question of fact, yes, briefly. For homework and practice problems it guides you first: it asks what you think, gives hints one at a time and shows similar worked examples. In the app the answer comes last.
Is a tutor that withholds answers slower?
Sometimes, on purpose. The research above suggests that the effort of working it out is part of how the learning happens. DuckyHelper keeps that effort small with quick hints and steps instead of long waits.
Sources
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. (accessed 2026-10-09)
- Macina, J., Daheim, N., Chowdhury, S., Sinha, T., Kapur, M., Gurevych, I., & Sachan, M. (2023a). MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems. Findings of EMNLP 2023. (accessed 2026-10-09)
- Baker, R. S., Corbett, A. T., Koedinger, K. R., & Wagner, A. Z. (2004). Off-Task Behavior in the Cognitive Tutor Classroom: When Students "Game the System". CHI. (accessed 2026-10-09)
- Sonkar, S., Ni, K., Chaudhary, S., & Baraniuk, R. (2024). Pedagogical Alignment of Large Language Models. Findings of EMNLP. (accessed 2026-10-09)
- Dinucu-Jianu, D., Macina, J., Daheim, N., Hakimi, I., Gurevych, I., & Sachan, M. (2025). From Problem-Solving to Teaching Problem-Solving: Aligning LLMs with Pedagogy using Reinforcement Learning. EMNLP. (accessed 2026-10-09)
- LearnLM Team, Google (2024). LearnLM: Improving Gemini for Learning. arXiv (Google tech report). (accessed 2026-10-09)
- Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. (2024b). Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. arXiv preprint (v1). (accessed 2026-10-09)