Learning Center AI
Tutor Training Nov 9, 2026 · 5 min By

Better Work Isn't Learning: Training Tutors to Hand the Thinking Back

AI can make a student's homework better without making the student better at it. What new research on metacognitive offloading and an inverted Bloom's model mean for tutoring sessions, plus three prompts for tutor training.

A student comes into the math lab with a finished Calculus I homework set. Every problem is worked and every answer is boxed. The tutor points to problem 4, a related rates question about a ladder sliding down a wall, and asks, “Why did you set it up that way?” The student can't answer without looking back at the screen.

The homework is right. The student just skipped the part that builds the skill. That moment is happening in centers everywhere, and a few pieces I saved this week help put a name on it.

Better work, same student

Alexandre Rodrigues argues in Chief Learning Officer that AI can clearly improve how well people do a task without doing much for what they learn from it. His framing is short: “Performance is visible. Learning is harder to see.” That problem set shows exactly that.

Rodrigues leans on a long line of learning research. Doing well during practice doesn't reliably predict what someone keeps or can use later, so getting problem 4 right with AI's help says little about whether the student can do a similar one on next week's exam. People also remember more when they generate an answer than when they read one. And a 2025 study of knowledge workers found that the more people trusted AI, the less critical thinking they put in.

His practical advice: let AI remove the friction that doesn't matter and protect the friction that does, like retrieving, explaining, deciding, and catching errors. On problem 4, the friction that mattered was deciding which quantities are changing, drawing the triangle, and picking the equation that links them. That's exactly what got skipped.

Hand the checking back

The second piece is a new paper by Yizhou Fan and colleagues, which Philippa Hardman flagged on LinkedIn. They describe two directions. In metacognitive offloading, the learner hands monitoring and judging over to the AI. In metacognitive onloading, the support is designed so those jobs shift back to the learner over time.

That's the useful part for us. The question isn't whether the student used AI. It's who did the checking. If the AI solved problem 4, checked the answer, and explained its own steps, the student was a spectator. If the student has to predict, judge, and fix, the student is learning, even with AI in the room.

Good tutors already know this move. It's the same instinct as not picking up the pencil. What's new is that the pencil now talks back, and students arrive with problem sets that look finished.

Bloom's, upside down

Michelle Kassorla's Inverted Bloom's model is a helpful map here. Her observation is that students now tend to create first, with AI, and understand later, if at all. So she flips the taxonomy and gives each stage an agency level, from AI doing the creating at the bottom to the student recalling the principles without the tool at the top. She has been sharing this with faculty, including at AAC&U, as a way to rebuild assignments for AI.

For a learning center, that ladder is basically a session plan. Our student is on the bottom rung: the AI created and the student copied. One rung up, the student checks the AI's answer against a quick estimate. Then they fix a step without help, explain why it works, and finally do a new ladder problem cold. A tutor's job is to figure out which rung the student is on and help them climb one.

What this looks like at the table

Here's how a tutor might work problem 4 with that student, no new software needed:

  • Open with recall, not the worked solution. Before looking at the screen, ask the student what problem 4 is asking and what they would draw first. Now you know where they are.
  • Make the student the checker. Instead of asking AI to find the mistake, ask the student whether the answer makes sense. The top of a sliding ladder is moving down, so its rate should be negative. Does the boxed answer agree?
  • Ask why it worked. If the student fixes a sign or a setup, have them explain why the new version is right. That's where understanding starts to stick.
  • Close without the tool. End with a new ladder problem with different numbers, no AI. That answers Rodrigues's question: can the skill survive without the tool?
  • Fade support across visits. If the student comes back next week, start with less help, maybe just asking what to draw. That's onloading in practice.

Train for it on purpose

Most tutor training still treats AI as a policy topic: what's allowed, what to disclose, what to report. That matters. But tutors also need practice spotting offloaded thinking and handing it back. Role-plays work well here. Give a new tutor that same problem set, have another tutor play the student, and let them practice getting the student talking about problem 4.

The same lens works for exam prep. If the student asks AI to build a study plan for the Calc I exam, they haven't practiced planning. Have them draft it first, then use AI to poke holes in it.

The point

This isn't about telling students to stop using AI. Their instructors may require it, and they'll use it at work. Our job is narrower: when that student leaves the session, they should be able to set up problem 4 without the screen. A finished problem set is nice. It isn't the measure.

Try It

These prompts work in whatever AI tool your campus provides. Fill in the bracketed parts before you paste. None of them need student data, so keep student names and ID numbers out.

1. Write a tutor training role-play on AI-assisted work. For tutor supervisors and coaching supervisors planning a training session.

You are an experienced tutor trainer at a college STEM tutoring center. You design short, realistic role-plays that help new peer tutors practice specific session skills.

Goal: write a role-play that helps tutors practice handing the thinking back to a student who brings in AI-assisted work they can't fully explain.

Context:
- Center type: [MATH LAB / STEM TUTORING CENTER / SCIENCE LEARNING CENTER]
- Course: [e.g., College Algebra, Calculus I, General Chemistry, Intro Physics]
- Topic: [e.g., related rates, stoichiometry, Newton's second law]
- Tutors in training: [e.g., first-semester peer tutors, mostly sophomores]
- Our AI policy for tutoring sessions: [PASTE OR SUMMARIZE, or write none yet]
- Training time available: [e.g., 20 minutes]

Include:
1. A one-paragraph scenario and a short sample of the student's AI-assisted work (two or three fully worked problems or a lab calculation) that looks correct and complete but hides a gap in understanding, such as not knowing why a formula applies or what the units mean.
2. A role card for the student with what they understand, what they don't, and how they will respond if the tutor just explains the solution.
3. A role card for the tutor with the goal but not the answers.
4. Five tutor moves that put the student back in charge of checking and judging (for example: explain the setup without looking at the screen, estimate a reasonable answer first, find the step most likely to be wrong, check units, try a similar problem with new numbers and no AI).
5. Three debrief questions for the group.

Format the output as a handout with clear headings. Keep it under 600 words. Use plain language a 19-year-old tutor would understand.

2. Audit your session protocol for who does the thinking. For directors, coordinators, and tutor or coaching supervisors revising how sessions run.

You are a learning scientist who advises college STEM tutoring programs on AI.

Goal: review our session protocol, show where AI use could let students hand off the thinking, and suggest changes that hand the checking and judging back to the student.

Context:
- Program: [DROP-IN MATH LAB / APPOINTMENT-BASED STEM TUTORING / SUPPLEMENTAL INSTRUCTION] at [COLLEGE NAME]
- Courses we support most: [e.g., College Algebra, Calculus I, General Chemistry, Intro Biology]
- How students tend to use AI before or during sessions: [e.g., bring AI-solved homework, photograph problems for step-by-step solutions, ask AI to explain lab calculations]
- Typical session length: [e.g., 30 minutes]
- Our current session protocol:
[PASTE YOUR SESSION STEPS OR CHECKLIST]

Use this lens: a student moves from low ownership to high ownership as they go from letting AI solve the problem, to evaluating the solution, to fixing errors themselves, to explaining why each step works, to solving a similar problem without AI. Effort that involves setting up the problem, choosing a method, estimating, explaining steps, or catching errors should stay with the student. Effort like looking up a formula sheet or checking arithmetic on a long calculation can be handed to AI.

Give me:
1. A table with three columns: protocol step, where the student might hand off the thinking, and a suggested change.
2. A revised version of the protocol with the changes built in, keeping our original steps where they already work.
3. Two or three questions I should answer before rolling this out to staff.

Do not add steps that would make a session run longer than our typical length. Keep the language plain enough to share with student staff.

3. Build no-AI exit checks for a course. For tutor supervisors and coordinators prepping tutors for a specific STEM course or unit.

You are an experienced peer tutor and content specialist for [STEM COURSE, e.g., Calculus I, General Chemistry].

Goal: create short exit checks that tutors can use in the last five minutes of a session to see whether a student can now do the skill on their own, without AI.

Context:
- Course: [COURSE NUMBER AND TITLE]
- Topics or skills in this unit: [LIST, e.g., chain rule, limiting reactants, free-body diagrams]
- Student level: [e.g., first-year students, many in their first college math course]
- Materials students use: [e.g., OpenStax textbook chapter 3, instructor lecture slides]

For each topic, give me:
1. One quick task the student does without AI (a similar problem with different numbers, a one-sentence explanation of why a step works, or a prediction of the answer's size, sign, or units), set up slightly differently from typical homework so it tests whether the skill transfers.
2. The correct answer and what a solid explanation sounds like, in one or two sentences, for the tutor.
3. One follow-up question the tutor can ask if the student gets stuck that prompts thinking rather than giving the answer.

Format the output as a one-page reference table tutors can print, with one row per topic. Use only content that fits the topics I listed. If you are unsure whether something is covered in this course, flag it instead of including it.

Sources

AI use statement

Sources come from my own reading, plus articles flagged by Elvis, an AI agent I built in Muse (Meta, 2026). I save the ones worth keeping with a note on why they matter for learning centers. Claude (Anthropic, 2026) helps me review those notes, choose topics, and outline and draft the article and its prompts. I revise every draft, check claims against the sources, and decide what gets published.

I tested the prompts in ChatGPT (OpenAI, 2026), Claude (Anthropic, 2026), Copilot (Microsoft, 2026), Gemini (Google, 2026) and Grok (xAI, 2026).

References

About the author

[[shortcode.authors.mark-woolwine]]

← Back to The Feed