Photo by Vitaly Gariev on Unsplash
Lists of AI tools for teachers tend to assume the hard part is finding the tools. In a school the hard part is the two decisions that come before any tool: what you are allowed to type into it, and what you do when you suspect a student used one.
Both have clearer answers than the discourse suggests, and getting them wrong carries more consequence than picking the second-best lesson planner.
Rule One: Student Information Does Not Go In
A consumer AI account is a third-party service. Putting personally identifiable student information into one means disclosing education records to a vendor your district has not vetted, and that is squarely FERPA territory rather than a matter of personal caution.
The Department of Education’s Student Privacy Policy Office publishes guidance on schools using online educational services, including the conditions under which a vendor can be treated as a school official acting on the district’s behalf. A free account you signed up for yourself does not meet those conditions.
The Rewrite That Solves Most of It
Nearly every genuinely useful teacher prompt works without a single identifying detail. Strip and generalise before you paste.
| Instead of | Write |
|---|---|
| “Marcus wrote this paragraph, is it any good” | “Here is a Year 7 paragraph on the water cycle. Give feedback against these criteria.” |
| “Write a report comment for a struggling reader named …” | “Draft three report comment templates for a reader working below expectation, warm but honest.” |
| “How do I handle this incident with two of my students” | “How do I structure a restorative conversation after a classroom conflict between two pupils?” |
The output is the same quality. The disclosure risk goes to nearly zero. This one habit resolves the overwhelming majority of the problem.
Photo by Thomas Park on Unsplash
Nearly every genuinely useful teacher prompt works without a single identifying detail.
Rule Two: Do Not Accuse on Detector Output
This is the more serious of the two, because the harm lands on a student rather than on a policy.
AI writing detectors are not reliable enough to support an accusation. Stanford researchers published in the journal Patterns after running TOEFL essays written entirely by humans through seven detectors: the tools flagged non-native English writing as AI-generated at an average rate above sixty percent. Essays by US students were mostly classified correctly.
The detectors were not measuring authorship. They were measuring how predictable the prose was, and careful writing in a second language looks predictable.
A tool that disproportionately flags two groups is worse than no tool, because it launders a bias into something that looks like evidence.
What to Do Instead When You Suspect It
Ask about process, not authorship. It is a conversation rather than a verdict, and it works whether or not AI was involved.
- “Talk me through how you got to this argument.”
- “Which part took you the longest?”
- “What did you cut from your first draft?”
- “Where did you check this fact?”
A student who did the work can answer these easily. A student who did not will usually say so at about the second question, and the conversation becomes about how to use the tool honestly rather than about catching them.
The design fix works better than the detection fix: assessment that requires drafts, in-class writing, or an oral defence removes the incentive rather than policing it. More on why detection itself is unreliable in How to Tell If Text Was Written by AI.
Photo by Vitaly Gariev on Unsplash
Photo by National Cancer Institute on Unsplash
Where AI Actually Saves a Teacher Time
Sorted by how much time it returns for how little risk it carries.
| Task | Time saved | Risk |
|---|---|---|
| Differentiating one worksheet to three levels | High | None, no student data involved |
| Generating practice questions from a topic | High | None |
| Rewriting a text to a lower reading age | High | None |
| Drafting parent communication templates | Medium | Low, if kept generic |
| Report comment scaffolds | Medium | Moderate, easy to over-share |
| Marking actual student work | Low | High, and the judgement is the job |
The top three are where the hours are, and they involve no student information at all. That is a convenient alignment: the highest-value uses are also the safest ones.
FAQ: Frequently Asked Questions
Can I put student work into ChatGPT for feedback?
Not with identifying information attached. Remove names and identifying detail and the same request is usually fine. If your district has an approved tool with an agreement in place, use that instead.
Are AI detectors accurate enough to catch cheating?
No. Published research found average false-positive rates above sixty percent on human-written essays by non-native English speakers. Use them as a prompt to have a conversation, never as evidence.
What is the safest way to start?
Lesson materials. Differentiating a worksheet or generating practice questions involves no student data, saves real time, and lets you judge output quality on something low-stakes.
Should students be allowed to use AI at all?
That is a curriculum decision rather than a technology one, and blanket bans have proved hard to enforce. Assessment designed around process, drafts and discussion holds up regardless of which way the policy goes.
General information for educators, not legal advice. FERPA obligations depend on your role and your institution’s agreements; follow your district’s policy and consult your data protection lead before using any tool with student information.