You Are the Part That Makes It Work

Read the research side by side and the teacher looks like the reason AI works when it works. Shape the tool, keep the thinking with the student, and let your judgment reach every desk.

If you teach, AI handed you two jobs this year and no extra time for either. Keep students thinking for themselves while a tool that will do the thinking for them sits in every pocket. And work out what to do with that same tool in your own work, usually with no training and no policy to lean on.

You are not alone in either one. In a June 2026 NPR and Ipsos poll of 545 K-12 teachers, 54 percent said AI makes it harder for students to learn to think for themselves. Training is patchy: RAND put the share of teachers who had received training on adapting their teaching to AI at about half by spring 2025, and in June 2026 about four in ten told NPR and Ipsos their school offers any at all. The rules are further behind. In a Walton Family Foundation and Gallup survey of 2,069 teachers in February and March 2026, 18 percent said they receive any formal guidance from administrators on how AI should be used at work. A year earlier, 19 percent worked at a school with an AI policy. On that, nothing moved.

You are right to worry about the thinking. This piece is about why answering it does not have to cost you more hours, and about a pattern in the research I have not seen named as plainly as it deserves.

Two things before the research. First, who is writing this. I am not a teacher, and I do not have a degree in education. I spent two years as a para-educator at a California high school. I also run an AI governance advisory firm, and Still in Charge is its student initiative, so I have an interest in schools taking this seriously. What follows comes from reading the research closely, not from running a classroom of my own.

Second, nothing here says AI belongs in every classroom, or that any teacher should be made to use it in their own work. AI is a tool, like a hammer. Swing it well and you drive the nail. Swing it carelessly and you hit your thumb. Whether to pick it up at all is your call. If you choose to, here is what the research suggests about swinging it well.

What I noticed reading the studies side by side

At the University of Pennsylvania, researchers gave nearly 1,000 high school math students in Turkey practice with a plain chatbot, with a tutor version, or with no AI at all. Once the tool was taken away, the plain chatbot group scored 17 percent worse on the exam than the students who never had it. The tutor version, which "includes teacher input and is designed to guide students with hints rather than directly giving answers," showed no such drop.

At Harvard, the instructors of an introductory physics course built their own AI tutor. They wrote their step-by-step solutions into it and told it to give away one step at a time, never the whole answer. Learning gains with the tutor were more than double those from the same topics taught in an active-learning class.

In Nigeria, a World Bank team ran a six-week after-school English program. Every session opened with a prompt built to make the AI tutor rather than answer, which the teacher delivered, and teachers kept students on task and led a reflection at the end. The authors estimate the gain at about a year and a half of typical schooling there.

At Stanford, AI was given to the tutors instead of the students. About 1,800 students worked with 900 human math tutors, half of whom had an AI assistant suggesting what to ask next. Their students were 4 percentage points more likely to master a topic, and 9 points more likely when the human tutor was one of the lower-rated ones.

And in five UK secondary schools, Google DeepMind and the learning platform Eedi had an AI tutor work with students while expert tutors approved every message before it went out. Students did as well as or better than they did with a human tutor alone.

None of these studies set out to measure the teacher. Each was testing a tool. But read them together and the same thing is sitting in every one where students came out ahead: a teacher or tutor shaped the tool, supervised it, or stood beside it. The one version that clearly hurt learning was the bot on its own, with no teacher's hand in it.

The teacher's part looked different in each one. A teacher opened each session with a tutoring prompt. Instructors wrote the solutions. Tutors approved every message. What stayed the same was a professional who could tell whether the tool was right, deciding what the tool was allowed to do.

That is my observation, not any study's finding. I think it is the most useful thing in this research for a teacher to know, because it suggests the part doing the work may not be the technology. What the better results had in common was hints instead of answers, one step at a time, a teacher's own solution underneath, a reflection at the end. That is teaching, written down where the tool can follow it.

You are not the part AI replaces. In this research, you look like the part that makes it work.

That also means the skill travels. In the Stanford study, the biggest gains came from the lower-rated tutors, the ones the AI was helping think more like an expert. A newer teacher, or a tutor, can borrow some of that judgment the same way.

The time is real. The reach is bigger.

England's Education Endowment Foundation tested the workload side directly. In a randomized trial, 259 science teachers in 68 secondary schools were split in two, and half were given ChatGPT and a guide on using it well. The ChatGPT teachers spent 31 percent less time planning lessons, about 56 minutes a week instead of about 82. An independent panel of teachers reviewed the materials and found no noticeable difference in quality.

That is a use of AI almost nobody argues with: a professional using a tool on their own work, and checking what it makes.

But minutes saved is the smaller win. A high school teacher may have 50 minutes with a class of 40 and see more than 200 students a day. Nobody has time for a two-minute conversation with each of them, and every teacher knows it. What AI can change is how far one teacher's judgment reaches inside those 50 minutes: your instructions, at every desk at once.

Where the two jobs meet

Write the instructions once. Before students use a tool in your class, give it the instructions you would give a student teacher. Give hints, not answers. One step at a time. Ask what the student tried first. Here is the correct solution, check against it. Then add one more: explain it in the student's own terms. Not a different language, their language: the words they use, the game they play, the example that already makes sense to them. The content stays yours. The tool carries it to each student in a form they recognize, which no one can do for 40 students at once. It takes some time up front, and you will adjust it after you watch students use it. But you write it once, wherever your school's approved tools let you, and it goes with every student who opens the tool. Instructions shape the tool. They do not lock it, and a determined student can talk past them, which is why the check below happens in the room.

Let it write the quiz, not the answers. A major 2013 review of learning research rated practice testing one of the two most effective study techniques it examined. Writing practice questions is also one of the slowest parts of the job. Have the tool draft ten questions on this week's material, fix what needs fixing, and hand them out. The tool did your prep. The students did the remembering. That is Think Better, our first principle, working the way it should.

Check what it makes for you. Verify Outputs, our third principle, applies to teachers first. The Harvard team wrote their own solutions into the tutor because AI can be confidently wrong. One wrong answer key teaches the whole class the wrong thing.

Trade the detector for the question. Issue 15 argued that detection asks the wrong question. It is also slow: running papers through a checker, then arguing over the result. The better question is the walkthrough: explain one choice in your work, or tell me what would change if one fact were different. You cannot ask it of 200 students a day, one at a time. You can ask all of them at once. Have the tool draft a few walkthrough questions for the assignment, then give students the first five minutes of class to answer one on paper, screens away. That answer is one a tool did not write. It will not tell you who wrote the homework, but it tells you what the student can explain without help, which is the thing you wanted to know. Read the answers that do not sound like the student, and save your conversations for the few who need one.

Move the weight off the homework. In the NPR and Ipsos poll, about four in ten teachers said they had already moved more work into the classroom or back onto paper, so this is less a proposal than a description. If the worry is that students will hand their homework to a tool, look at how much of the grade is riding on it. The pull toward the shortcut is strongest where the stakes and the stress are. Issue 14 quoted MIT's own report: students said the temptation was greatest when they feared missing a deadline. Count homework as practice, worth enough that students do it and reflect on it, and put most of the grade on work where you can see the thinking happen: in class, out loud, on paper. Weighting is often set above your desk, and moving it touches accommodations and the students carrying a job or a family at home, so treat this as a question for a department or a school rather than a rule for every class. It also leaves you with fewer papers to police.

All of this, with a starting set of instructions you can copy, is on one page for teachers: stillincharge.org/teacher-guide

For schools

Every result above that went well had a teacher or tutor who knew how to shape the tool, working inside a design someone had set on purpose. About half of teachers have had some training. Eighteen percent have formal guidance. If your school is going to use these tools, decide what they are allowed to do before you buy one, give the teachers who choose to use them the time to write the instructions, and name who owns the answer key. In this research, that is the part that was present wherever the products worked.

The short version

The worry is that AI takes over the teacher's job. Read the research side by side and the teacher looks like the reason it works when it works. Shape the tool, keep the thinking with the student, and let your judgment reach further than one room and 50 minutes ever allowed. Those were never two different goals.

What the checking turned up

Six things about the evidence behind this piece, because you should be able to go look.

The teacher surveys at the top are three different instruments, and on training they do not agree. NPR and Ipsos polled 545 teachers, Gallup polled 2,069, and RAND drew on its own national panels. Asked slightly differently about the same spring, Gallup put the share of teachers with no AI training at 68 percent and RAND put it at about half. Treat the training picture as a range, not a number. Gallup's 2025 and 2026 questions also differ: the first asked whether a school had an AI policy, the second whether a teacher receives formal guidance.

The pattern this piece is built on is my reading across five studies. None of them was designed to compare the same tool with and without a teacher, so none of them proves the teacher was the reason. Each is linked below so you can read them side by side yourself.

The studies vary in strength. The Penn and Harvard studies are peer reviewed. The Education Endowment Foundation trial is an independent evaluation. The Nigeria and Stanford studies are working papers, and the DeepMind and Eedi study is described by its authors as exploratory, with 165 students.

The Harvard result cuts both ways. The AI tutor beat the same instructors' own active-learning class, which can be read as the tool outperforming the teacher. Those instructors had written their solutions into it, and the authors say it should not replace in-person teaching.

Several measure learning over a short span. The Harvard study covers two lessons with a test right after each. None of them shows how long the gains last.

The lesson-planning trial measured planning time, not a teacher's whole workload, and its quality review used a small sample of materials.

NPR and Ipsos, Most K-12 teachers say AI's impact on education will eclipse the internet or computers, 2026: www.npr.org/2026/06/05/nx-s1-5779757/school-ai-education-students-teachers-poll-critical-thinking

Walton Family Foundation and Gallup, Teaching for Tomorrow: Closing the Expectations Gap, 2026: news.gallup.com/poll/710534/teachers-receive-no-formal-guidance.aspx

Doss et al., AI Use in Schools Is Quickly Increasing but Guidance Lags Behind, RAND, 2025: www.rand.org/pubs/research_reports/RRA4180-1.html

Gallup and Walton Family Foundation, Teaching for Tomorrow, 2025: nextgeninsights.waltonfamilyfoundation.org/wp-content/uploads/2025/06/Teachers_AI_Report.pdf

Bastani et al., Generative AI without guardrails can harm learning, PNAS, 2025: www.pnas.org/doi/10.1073/pnas.2422633122

Wharton's plain-language summary of the same study: knowledge.wharton.upenn.edu/article/without-guardrails-generative-ai-can-harm-education

Kestin et al., AI tutoring outperforms in-class active learning, Scientific Reports, 2025: doi.org/10.1038/s41598-025-97652-6

De Simone et al., From Chalkboards to Chatbots, World Bank, 2025: openknowledge.worldbank.org/bitstreams/cd9cca5b-2ffd-4be1-8313-0f4d2ed6ef4f/download

Wang et al., Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise, Stanford, 2024: arxiv.org/abs/2410.03017

Eedi and Google DeepMind, AI tutoring can safely and effectively support students, 2025: arxiv.org/abs/2512.23633

Education Endowment Foundation and NFER, ChatGPT in lesson preparation, 2024: educationendowmentfoundation.org.uk/projects-and-evaluation/projects/choices-in-edtech-using-generative-ai-chatgpt-for-ks3-science-lesson-preparation-2024-teacher-choices-trial

Dunlosky et al., Improving Students' Learning With Effective Learning Techniques, 2013: www.psychologicalscience.org/news/releases/which-study-strategies-make-the-grade.html

When we get something wrong in a published issue, it will be listed here too.