Why this guide, and why now
Most of the public argument about generative AI in schools has been about students using it. This guide is about you, the teacher, using it for lesson prep. That distinction matters because the evidence base is finally starting to include teachers as the subjects of study rather than students.
According to reporting by the Hechinger Report in July 2026, a working paper studied teachers using generative AI to prepare lessons and found a pattern worth taking seriously: teachers saved time, but their students may pay a price Source. The paper's title, as listed, is blunt: "Generative AI Can Harm Teaching" Source. One of the researchers associated with the work is Alp Sungu, whose faculty page lists the study Source.
We want to be direct about what we can and cannot stand behind here. Bright Learning did not conduct this research and did not independently verify the study's raw statistics. The primary working paper on SSRN could not be opened during our review (it returned an HTTP 403 error), so the exact title, full author list, and precise numbers were checked only against the Hechinger Report account and Sungu's faculty listing, not against the primary document itself Source. Treat the specific claims below as reported findings from a preprint, not settled fact.
What the study appears to show
The headline pattern reported is that using AI output as a finished product, rather than as a first draft, may lower student motivation, and for weaker teachers it may also lower achievement Source. The reporting frames the harm as landing hardest when AI-generated material is delivered to students essentially as-is.
There is a companion line of research on the student side. A peer-reviewed study titled "Generative AI Can Harm Learning" examined how AI assistance affected students and is a separate, published piece of work Source. The teaching-side paper echoes that framing in its title but focuses on teachers as the actors.
What we cannot claim
Before any practice advice, the honest caveats:
- The teaching study is a preprint and is not yet peer-reviewed Source.
- Control-group teachers could use other AI tools, which muddies the comparison between AI and no-AI conditions Source.
- Researchers did not observe classrooms or analyze the AI-generated materials directly, so the mechanism behind the harm is inferred, not proven Source.
This means the "why" in everything below is our reasoned interpretation of a reported result, not a demonstrated causal chain. We flag it so you can weigh it accordingly.
The practice: draft it, don't ship it
The single most defensible takeaway from the reporting is a workflow rule, not a ban. AI is useful for generating a rough draft. The harm signal appears when that draft becomes the finished product students receive Source.
1. Treat every AI output as a first draft to rewrite
Generate freely, then rewrite in your own voice. The rewriting step is where your subject judgment, your knowledge of this class, and your pedagogical choices re-enter the material. If you hand students unedited AI output, you are removing exactly the layer the study's framing suggests matters.
2. Protect the personality-driven parts of a lesson
Some parts of a lesson carry your personality: the analogy that only you use, the running joke, the story from your own experience, the way you frame a hard idea. These are plausible carriers of student motivation. Since the reported harm centers on motivation, be especially unwilling to let AI replace these elements. Write them yourself.
3. Be most cautious when tired or teaching outside your strength
The reported finding that weaker teachers saw achievement effects suggests the danger rises when your own subject expertise is thin Source. The practical translation: the moments you are most tempted to accept AI output wholesale (late at night, or in a subject you are covering outside your training) are the moments you can least afford to skip the rewrite. In those exact conditions, do less with AI, not more, or reserve extra time for editing.
4. Watch engagement as an early-warning signal
If motivation effects precede test-score effects, then student engagement is your leading indicator. You will feel a room going flat weeks before it shows up in an assessment. Treat a dip in participation, energy, or effort on AI-assisted lessons as a prompt to inspect and rewrite your materials, not as noise.
Safety, privacy, and suitability
- Privacy: do not paste student names, grades, IEP details, or any identifiable student information into public AI tools. Use anonymized or generic inputs.
- Age suitability: AI drafts can produce content pitched at the wrong reading level or including inappropriate examples. Your rewrite is also your safety check.
- Accuracy: AI can produce confident errors. In subjects where you are outside your strength, this compounds with the achievement risk above. Verify facts against a trusted source before use.
- Institutional policy: check your school or district rules before using any tool for prep.
Failure modes to recognize
- Ship-it drift: starting with good intentions to edit, then gradually pasting AI output straight through as the term gets busy. This is the exact behavior the reporting associates with harm.
- Voice flattening: your lessons start sounding generic and interchangeable. Students notice.
- Expertise masking: AI output looks polished enough that you stop noticing you do not actually understand the material well enough to teach it live.
- Overcorrection: banning AI entirely and losing the genuine time savings. The evidence points to how you use it, not to abstinence.
Bottom line
The reported result is not "AI is bad for teaching." It is closer to "AI output shipped as a finished product may cost you student motivation, and if your subject footing is shaky, achievement too" Source. The workflow that follows is simple and low-risk: draft with AI, rewrite in your voice, guard the human parts, be extra careful when tired or out of your depth, and treat engagement as your smoke alarm. Given that the underlying study is an unverified preprint with an inferred mechanism, this is guidance to adopt with your eyes open, not a mandate.