What AI can and cannot do for a teacher

Two claims about AI in teaching are both wrong. One is that it will replace the explaining, marking and knowing-your-students that make up the job. The other is that it is a toy with nothing to offer a working teacher. The useful position is narrower and more boring: AI is very good at producing a first draft of something you can check quickly, and unreliable wherever the output is a judgement nobody will verify.

The dividing line

Ask one question of any task: can I tell in seconds whether the output is wrong?

A generated set of twenty practice questions: yes. You read them, you spot the two that are broken, you delete them. A generated report comment on a student you teach: yes, because you know the student. A generated mark for a long answer, or a claim about which topic your class is weakest in, drawn from data you have not checked: no. Nothing in those outputs tells you it is wrong, so the error passes straight through into a decision.

Where it saves real time

• Question variants. You write one good question; you need eight versions for two exam papers and a homework sheet. This is genuine drudgery with a fast check.
• First-draft materials from your own documents. A summary sheet, a set of recall questions, a worksheet built from a chapter you supply. Note "your own": output grounded in your material is checkable against it.
• Rewriting for a different level. The same explanation for a weaker group, or in the other language.
• The administrative sentence. Parent messages, session summaries, the wording of an instruction. Small individually, and one of the largest blocks in a tutor's week in total.

Where it quietly fails

The dangerous cases are the plausible errors, not the obvious ones. A generated question with a subtly wrong answer key, marked confidently, teaches thirty students something false. A generated Arabic explanation that reads fluently and uses a term your curriculum does not use costs a lesson to undo. Both look fine at a glance, which is exactly why they need reading rather than skimming.

The rule that keeps it useful

Never put AI output in front of students unread. Not as a matter of principle but of arithmetic: checking twenty questions takes four minutes, and one wrong answer key costs a session plus the trust of a group. The time saved is real, and it comes from not having to write the draft, not from skipping the check.

Sources

• UNESCO — Education
• OECD — Education and Skills