Why Chalkchalk exists
The problem
Homework is where the thinking happens.
The blank line under the problem. The wrong first try. The second one that works.
Now a chatbot fills it in.
Paste the question, get the answer. The working disappears, and so does the learning it was for.
How common it is
54% of U.S. teens have used AI chatbots for help with schoolwork.
One in ten say they do all or most of their schoolwork with a chatbot’s help.
Source: Pew Research Center, “How Teens Use and View AI,” February 2026 (opens in a new tab)1,458 U.S. teens ages 13–17, surveyed September–October 2025.
59% think using AI to cheat is a regular occurrence at their school.
Source: Pew Research Center, “How Teens Use and View AI,” February 2026 (opens in a new tab)1,458 U.S. teens ages 13–17, surveyed September–October 2025.
94% of UK undergraduates use AI to help with assessed work.
12% paste AI-generated text straight in, up from 3% two years earlier.
Source: HEPI and Kortext, “Student Generative AI Survey 2026” (opens in a new tab)1,054 UK undergraduates, surveyed December 2025.
What it’s used for
47% of students’ conversations with one AI assistant aimed to get the answer as fast as possible.
And the AI was doing the higher-order thinking: creating (39.8%) and analyzing (30.2%) were its most common tasks.
Source: Anthropic, “Anthropic Education Report: How University Students Use Claude,” April 2025 (opens in a new tab)574,740 student conversations. Anthropic notes that many direct conversations serve legitimate learning purposes.
1 in 4 U.S. teachers say AI tools do more harm than good in K-12 education.
Among high-school teachers, it’s 35%.
Source: Pew Research Center, “A quarter of U.S. teachers say AI tools do more harm than good in K-12 education,” May 2024 (opens in a new tab)2,531 U.S. public K-12 teachers, surveyed fall 2023.
What it costs
It feels like learning.
Researchers gave nearly 1,000 high-school math students three ways to practice: no AI, GPT-4 with no guardrails, and a GPT-4 tutor with guardrails.
Source: Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS, 2025 (opens in a new tab)Nearly 1,000 high-school math students in Turkey. The guardrailed tutor matched, but did not beat, the no-AI group on the exam.
+48% practice grades with unrestricted AI.
With the guardrailed tutor, +127%. So far, AI looks like a win.
Source: Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS, 2025 (opens in a new tab)Nearly 1,000 high-school math students in Turkey. The guardrailed tutor matched, but did not beat, the no-AI group on the exam.
−17% on the exam, without AI.
Students who had practiced with unrestricted AI scored 17% worse than students who never had it. The guardrailed tutor’s students scored about the same as them.
Source: Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS, 2025 (opens in a new tab)Nearly 1,000 high-school math students in Turkey. The guardrailed tutor matched, but did not beat, the no-AI group on the exam.
The difference was the guardrails: hints instead of answers.
And the students who leaned on AI thought they’d learned more than they had.
Source: Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS, 2025 (opens in a new tab)Nearly 1,000 high-school math students in Turkey. The guardrailed tutor matched, but did not beat, the no-AI group on the exam.
The fix
Not the answer.
What we built
Your next step.
Chalkchalk asks what you’ve tried, points to the page in your notes, and offers one step at a time. You do the climbing.
An answer machine is an elevator. The exam takes the stairs.
Does guidance work?
2× the learning gains, in less time.
In a Harvard physics course, students’ median learning gain with an AI tutor built on teaching research was more than double an active-learning class’s, in a median of 49 minutes against a 60-minute class.
Source: Kestin et al., “AI tutoring outperforms in-class active learning,” Scientific Reports, 2025 (opens in a new tab)194 students in a Harvard physics course, two lessons. A purpose-built research tutor, not Chalkchalk.
That’s the design we built on.
Sources
- Pew Research Center, “How Teens Use and View AI,” February 2026 (opens in a new tab)1,458 U.S. teens ages 13–17, surveyed September–October 2025.
- HEPI and Kortext, “Student Generative AI Survey 2026” (opens in a new tab)1,054 UK undergraduates, surveyed December 2025.
- Anthropic, “Anthropic Education Report: How University Students Use Claude,” April 2025 (opens in a new tab)574,740 student conversations. Anthropic notes that many direct conversations serve legitimate learning purposes.
- Pew Research Center, “A quarter of U.S. teachers say AI tools do more harm than good in K-12 education,” May 2024 (opens in a new tab)2,531 U.S. public K-12 teachers, surveyed fall 2023.
- Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS, 2025 (opens in a new tab)Nearly 1,000 high-school math students in Turkey. The guardrailed tutor matched, but did not beat, the no-AI group on the exam.
- Kestin et al., “AI tutoring outperforms in-class active learning,” Scientific Reports, 2025 (opens in a new tab)194 students in a Harvard physics course, two lessons. A purpose-built research tutor, not Chalkchalk.