Newsletter AI Learning, Keeping It Human Call 647-641-0643
Christopher Ross
,

Why didn’t the students ask for help?

Christopher Ross

10 min read
A coil of thick, natural thread unravels slightly, suggesting fragmented connections and unspoken questions.
In this post

In 18 middle schools in Tennessee, students were randomly assigned access to Khan Academy’s Artificial intelligence (AI) tutor, Khanmigo, for their maths practice, and researchers followed what happened over two years.

Nearly all of them tried it: 96 percent at least once. After that, the median student messaged the tutor on about a third of the days they practised, and in only 17 percent of the practice sessions where they’d made a mistake. When they did write to it, the messages were mostly bare answers or a click on a suggested prompt (Oreopoulos and Low, EdWorkingPapers, 2026).

The tutor still helped. Maths achievement rose by about 1.3 national percentile ranks per term, which Philip Oreopoulos and Nina Low put at roughly 0.06 to 0.08 standard deviations over a school year, a small effect in research terms. Their conclusion is what I’d keep: “The binding constraint appears to be engagement.” Access was never the problem. Getting students to reach for it when they were stuck was.

I think that’s the most hopeful research result of the summer, because it moves the problem to a place where teachers have real power.

Key takeaways

  • In a two-year trial in 18 Tennessee middle schools, Khanmigo raised maths achievement by about 1.3 percentile ranks per term, a small effect.
  • Nearly every student tried it, but the median student reached for it in only 17 percent of practice sessions with a mistake. The researchers name engagement as the binding constraint.
  • The best-evidenced AI so far makes the adult in the room better at the job, as Tutor CoPilot did for human tutors.

What the evidence actually separates

Last month Stanford’s SCALE Initiative published a brief whose title belongs on the wall of every procurement meeting: AI tutoring is not a monolith (Turano et al., Stanford SCALE, 2026). Susanna Loeb’s framing question is the one to borrow. Are we talking about homework help, or a high-impact tutor?

The brief sorts tutoring models by how much human involvement they keep. Live human tutoring still has the strongest evidence behind it. Fully automated tutors have the weakest record on simply getting used: in two randomised trials the brief cites, between 40 and 47 percent of students never used the tool at all, and active users logged two to five minutes a week, about the length of one song.

The model in between is where it gets exciting. Rose Wang, Dora Demszky, Susanna Loeb and colleagues gave 900 tutors working with about 1,800 K-12 students from historically under-served communities a tool called Tutor CoPilot, which suggests expert-style moves as the session happens. Students whose tutors had it were 4 percentage points more likely to master a topic, and 9 points more likely if their tutor was one of the lower-rated ones. It cost about $20 per tutor per year, and the tutors using it asked more guiding questions and handed over fewer answers (Wang et al., working paper, 2024).

AI tutoring models compared by human involvement
ModelWhat the evidence showsUse
Live human tutoringStrongest evidence of the threeThe comparison baseline
Human plus AI (Tutor CoPilot)About 900 tutors and 1,800 K-12 students: 4 points more likely to master a topic, 9 points more if the tutor was lower-rated. About $20 per tutor per yearTutors asked more guiding questions and handed over fewer answers
Fully automatedWeakest record on being used at all (two randomised trials cited in the Stanford SCALE brief)40 to 47 percent never used it; active users logged two to five minutes a week

In K-12, the best-evidenced AI so far is the kind that makes the adult in the room better at their job.

How good can it get?

When the design is right, a lot better than modest.

Gregory Kestin and colleagues built an AI tutor for a university physics course around the same teaching principles as the course’s own active-learning lessons, then randomised students between the two. Students working with the AI tutor learned significantly more in less time, and said they felt more engaged and more motivated (Kestin et al., Scientific Reports, 2025).

The Organisation for Economic Co-operation and Development (OECD) reaches the same place from the policy side in its 2026 Digital Education Outlook. Generative AI can support learning when it’s guided by clear teaching principles, and used without them it mostly lifts task performance with no real learning gain (OECD, 2026).

Picture that in a Grade 7 maths room. A student gets a problem wrong, opens a tutor set up to ask questions instead of giving answers, types what she tried, and gets back a question that points straight at her mistake. Meanwhile the teacher is working with the three kids who are properly stuck. Every piece of that already exists. It takes a well-chosen tool and a routine the class has practised, and both of those are a teacher’s calls.

One honest limit first. Stanford’s wider 2026 review went through more than 800 papers and found only 20 that produce strong causal evidence, and as of late 2025 none of them were high-quality causal studies of US K-12 students (Fesler et al., Stanford SCALE, 2026). The Tennessee trial belongs to the first real wave, and the next couple of years of trials will tell us a great deal more than this one.

Start with yourself

The clearest evidence for teachers so far is about teachers. Gallup and the Walton Family Foundation found six in ten US K-12 teachers used AI in their work in 2024-25, weekly users estimated saving 5.9 hours a week, roughly six weeks of a school year, and most said it improved the quality of that work at least somewhat (Gallup, 2025). Stanford’s review found teachers using AI for lesson preparation spent less time planning without any drop in lesson quality.

Most are doing it without a school playbook. Gallup’s follow-up this spring found 82 percent of teachers had no formal guidance on applying AI across their work tasks, and teachers in higher-needs schools were less likely to have any (K-12 Dive, May 2026).

What to measure instead of satisfaction

“Did they like it” tells you very little. Kestin’s team measured how much students learned and how students felt about it as two separate things, and that habit is worth stealing. In your own room, two numbers do most of the work.

  • The reach-for rate. When a student gets something wrong, how often do they go to the tool? In Tennessee it sat at 17 percent for the median student, and the researchers point to engagement as what held the results back.
  • Performance with the tool closed. Every so often, a short check without it. If that number doesn’t move, the tool is doing the work and the student isn’t.

Four things for Monday

  • Make the tutor the expected move when stuck. Engagement is the constraint, so build the reach into the routine: wrong answer, open the tutor, then show me what you asked.
  • Teach the question. The Tennessee messages were mostly bare answers. Put one of your own prompts on the board this week, bad first attempt included, and fix it together.
  • Pick the tutor that won’t just answer. Stanford’s review found tools with pedagogical guardrails, the kind that walk through reasoning step by step, show more promise than general-purpose chatbots.
  • Point it at your own prep first. That’s where the evidence for teachers is strongest, and you’ll meet its failure modes before your students do.

If you want a structured start, Google’s AI Educator Series offers free, self-paced modules for K-12 teachers with badges aligned to the International Society for Technology in Education (ISTE) standards, plus a free ten-hour virtual Badge-a-thon this Saturday, September 19 (Google, September 2026). It’s a vendor’s training built around its own tools, so treat it that way, but it’s free and it’s this week.

If you were hoping the software would do the teaching, a gain of 1.3 percentile ranks is a letdown. It also shows the tool helps when students use it, and getting students to use something well is ordinary classroom work.

One question I’d like answered, because the Tennessee study can’t tell us. If you’ve got students who actually reach for an AI tutor when they’re stuck, what did you change to get them there?

Next edition: teaching the question. If students mostly send an AI tutor bare answers, the missing skill is asking, and asking can be taught.

These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.

How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.

Common questions

Does Khanmigo work? A little. In a two-year randomised trial across 18 Tennessee middle schools, Khanmigo raised maths achievement by about 1.3 national percentile ranks per term, roughly 0.06 to 0.08 standard deviations over a school year. It helped when students used it, and most didn’t use it much.

Why don’t students ask an AI tutor for help? In the Tennessee trial, the median student messaged the tutor in only 17 percent of practice sessions where they’d made a mistake, and mostly sent bare answers. The researchers call engagement the binding constraint, which makes it a classroom routine teachers can work on.

How do you know whether an AI tutor is working? Watch two numbers. The reach-for rate: when a student gets something wrong, how often do they go to the tool? And performance with the tool closed, from a short check without it every so often. Whether students liked it tells you very little.

References

Working through something on your own site? Get in touch →