When the Magic Wears Off: Why Teachers Don't Trust AI Tools (And What We Can Do About It)

‍

Teaching in the Age of AI | Weekly Digest

Picture this: a teacher types "create a rubric for a 7th-grade argumentative essay" into ChatGPT. Thirty seconds later, a polished, five-category rubric appears. It looks right. It even sounds like something she might have written herself, on a good day, with a cup of coffee and two free periods. She uses it. It works fine.

And yet — something feels off.

That quiet unease isn't just anecdotal. A 2025 survey of 554 Michigan educators found that AI adoption is outpacing educator trust at a striking rate: teachers are using these tools faster than they feel confident in them. The researchers flagged this gap as one of the most important findings in the data. And when you look at how teachers are currently using AI tools in the classroom, the trust gap starts to make a lot more sense.

What Educators Are Actually Doing With AI

The Michigan Virtual report included something rare and valuable: a taxonomy of how educators are actually using AI, built from their own words. Seven categories emerged — instructional planning and resource creation, student learning support and personalization, assessment and feedback, student use of AI for learning, creative and visual applications, communication and administrative support, and AI literacy and ethics. And they tell a consistent story.

Teachers are using AI to plan lessons, generate rubrics, create quiz questions, draft feedback, and write emails. Students are using it to get writing help, research topics, and explore ideas. In the AI literacy category — perhaps the most telling of all — teachers described having students use AI "as a tool as well as bringing awareness to how it can lead them in the wrong direction."

What almost all of these use cases share is the same basic interaction pattern: type a request, receive an output. Prompt in, product out. No trail. No reasoning. No evidence that the rubric reflects what actually matters in a 7th-grade classroom, or that the quiz question targets the right cognitive level, or that the feedback is grounded in anything beyond pattern-matching on billions of prior texts.

The output might be perfectly good. But you have no way of knowing why it's good — or whether it actually is.

The Problem Isn't AI. It's the Black Box.

There's a well-documented concept in medicine called "black-box AI" — when an AI diagnostic tool produces an accurate result, but physicians can't see its reasoning. A 2021 study published in the Journal of Gastroenterology and Hepatology found that clinical adoption of AI tools has been slow because of a lack of trust — physicians are reluctant to trust and adopt something they do not fully understand, even when the tool is demonstrably capable. The problem isn't accuracy. It's the absence of an audit trail.

The same dynamic is playing out in classrooms, just without anyone naming it that way.

When a teacher receives a rubric from an AI tool for teachers, she has no idea whether it was generated by drawing on decades of writing research, or by stitching together the most statistically common rubric-shaped text on the internet. The output might pass a smell test. But in a profession where knowing your why is foundational — where teachers are expected to justify every pedagogical choice to students, parents, administrators, and themselves — a tool that can't show its work is going to feel untrustworthy. Not because it's wrong. Because it feels like magic.

Magic is impressive. Magic is also, by design, something you can't examine.

What an Evidence-First Approach to AI in Education Would Look Like

Here's the shift worth considering: what if AI in education were oriented around evidence first, rather than output first?

In an evidence-first model, the starting point isn't a typed request — it's real data from real classrooms. Student work, assessment results, observation notes, learning progression records. AI's role is to help teachers interpret and act on that evidence, not to generate plausible-looking materials from scratch.

Imagine a personalized learning tool that doesn't just produce comments on a student essay, but surfaces patterns across a class's writing — and shows you which students share similar gaps, traced back to the actual sentences where those gaps appear. Or an AI lesson planning assistant that starts with your last unit's assessment data and suggests adjustments grounded in what your students actually struggled with, not a generic template for a grade level it's never met.

The outputs would still be AI-assisted. But the reasoning would be visible. The evidence would be there. A teacher could look at a suggested intervention and follow the thread back to the classroom reality that generated it.

That's not magic. That's a tool you can trust — because you can interrogate it.

5 Questions Every Teacher Should Ask Before Trusting an AI Tool

Before adopting any AI tool for classroom use, run it through this checklist:

1. Can it show its work? When the tool gives you a recommendation, a rubric, or feedback, can it point to why? Is there an explanation, a source, or a data trail behind the output — or did it just appear?

2. Is the output grounded in your students, or someone else's? A tool trained on generic data across millions of classrooms isn't necessarily wrong — but it isn't looking at your kids. Does the tool use any actual data from your class, your school, or your students' prior work?

3. What would you tell a parent or administrator if they asked how this was made? If you can't explain how the AI arrived at an output, you can't stand behind it professionally. That's a red flag.

4. Does it help you make better decisions, or does it make decisions for you? The best AI tools for teachers support human-centered AI — they surface information, flag patterns, and present options. Tools that simply replace your judgment without explaining themselves are worth approaching with caution.

5. Has anyone studied whether this tool actually works — in classrooms like yours? AI literacy for teachers includes asking for evidence of effectiveness. Look for peer-reviewed research, case studies, or transparent impact data. If none exists, you're in pilot territory whether the company says so or not.

The Trust Gap Is Telling Us Something

The Michigan educators who said AI use is outpacing their trust aren't being irrational. They're being appropriately professional. A teacher who won't fully trust a tool she can't interrogate is doing exactly what good teachers do: asking for evidence before committing.

The answer isn't to tell teachers to trust AI more. It's to build — and demand — responsible AI in education that deserves that trust, by grounding outputs in the evidence of actual teaching and learning. The categories in the Michigan report — planning, assessment, feedback, personalization — are all areas where evidence-based AI could be transformative. But only if we stop treating "prompt in, output out" as the ceiling of what's possible.

Until then, the magic act will keep working. And teachers will keep feeling uneasy about it — for very good reason.

Teaching in the Age of AI is a weekly digest of research and ideas for educators navigating artificial intelligence in the classroom. Each week we follow the evidence — because that's what teachers do.

‍