AI in Teaching: What Helps

The market for educational AI is large and mostly aimed at administrators. From a teacher's position the useful applications are narrower and more mundane than the marketing suggests.

Where it genuinely saves time

Generating variations. Ten versions of a problem set, differentiated worksheets, alternative examples for a concept students did not get. This is real work that takes real hours.

Drafting administrative text. Parent communications, permission letters, report comment banks, meeting notes.

Rubric drafting. A first version to edit is faster than starting cold.

Adapting reading level. The same material at three levels, for mixed classes.

Producing practice questions across difficulty levels.

Translating communications for families who do not speak the school's language. Genuinely valuable, and it needs review before it goes out.

First-pass feedback on drafts, which the teacher then checks. Not replacing feedback — increasing how often students can get some.

Where it does not help

Knowing your students. Which one is struggling and why, who needs pushing, what happened at home this week. None of this is in the system.

Marking that matters. Automated marking of anything beyond the mechanical is unreliable and it removes the teacher's own diagnostic signal — you learn what the class misunderstood by reading their work.

Curriculum design without heavy expert editing.

Anything requiring knowledge of your specific syllabus, exam board or department norms.

The assessment problem

This is the real change and it is not solved by detection.

Any assessment that can be completed by a model at home will be. That is now true of most essay-style homework in most subjects.

Detection is not a solution. Detectors are unreliable, they accuse innocent students disproportionately, and they do not survive scrutiny in a disciplinary process.

Assessment redesign is the only durable response. The options that work:

In-class writing, supervised, handwritten or on locked devices.

Assessment on process — outline, draft, revision, reflection — where the sequence is visible.

Work that builds on class-specific material the model has no access to.

Oral components. A five-minute conversation about a submitted piece establishes authorship better than any tool.

Tasks requiring primary data the student collected.

Iterative work where each stage responds to feedback on the last.

None of this is new pedagogy. It is the pedagogy that was always better and was harder to justify the cost of.

Teaching students to use it

The more useful position than prohibition, and the one most institutions are moving toward.

Show them the failure modes. Have the class ask a model something they know well and find the errors. This does more than any warning.

Have them check citations. A lesson where students verify a model's references and discover several do not exist is memorable.

Teach the difference between using it for explanation and using it for production.

Be explicit about what is permitted in each assignment rather than issuing a blanket rule. "You may use it for X and not Y" is enforceable and teachable; "do not use AI" is neither.

Data protection, which schools frequently miss

Student work is personal data, and pasting it into a consumer AI tool is a disclosure to a third party.

Check whether the tool is approved by your institution before putting student names, work or assessment data into it.

Free consumer tiers generally permit training on inputs. Student work should not go into them.

On-device and institutional tools are the safer route where they exist.

This is the part that generates actual liability, and it is the part least discussed in the enthusiasm about classroom uses.

A defensible classroom policy

Blanket prohibition is unenforceable and blanket permission abandons the assessment. A per-assignment statement works better and takes a sentence.

State what is permitted for each task. "You may use AI to check grammar and to test your understanding. You may not use it to generate text you submit." Specific, teachable, enforceable.

Require a short declaration. What was used and how. Normalising disclosure is more effective than policing concealment.

Explain the reason, once. Students who understand that the assessment tests their thinking rather than their output comply better than students who are simply told no.

Assess something the tools cannot do where the stakes are high — supervised writing, oral discussion, work built on class-specific material.

Do not rely on detection. Build the policy on assessment design and honest declaration rather than on a tool that accuses non-native speakers disproportionately and cannot be defended in an appeal.