How to Use AI as a Teacher: Uses, Limits, Risks

A teacher's week splits into two different kinds of work filed under one job title: preparing material, and reading a specific room of students in real time. A general-purpose AI tool is genuinely good at the first — a first-pass worksheet, a bank of practice questions, a parent email — and has no access to the second, because noticing that one student is stuck on a concept for a different reason than the student before them comes from watching them work, not from anything a chat window can see. Knowing which half of the week you are handing to a tool, and which half stays yours, is most of what makes this useful rather than risky.

Will AI replace teachers covers that split at the level of the whole profession — what is genuinely shrinking and what is not. This is the practical version: what to actually hand over this week, and where a fluent-sounding draft still needs a check before it reaches a classroom. Profession endangered by AI covers the same task-by-task test for any job, not just this one, if the worry started somewhere more general.

Where AI genuinely helps

1. 01Drafting a first-pass worksheet or a bank of practice questions at a given difficulty from a topic and a grade level, so you are editing and levelling a draft rather than writing every question from a blank page.
2. 02Turning a dense source — a textbook passage, a research summary — into a plain-language explanation you then adapt in your own words for a specific class.
3. 03Drafting a first-pass comment on a routine piece of writing, which you edit into feedback that actually reflects what this student needs to hear next.
4. 04Writing the first draft of a standard parent newsletter or a routine announcement, so you are reviewing tone and detail rather than composing from nothing.
5. 05Generating a few variations on the same practice problem at slightly different difficulty levels, for the handful of students who need an easier or harder version of what the rest of the class is doing.

A worked example

Weak: "Make a worksheet about fractions."

Better: "Write ten practice problems on adding fractions with unlike denominators, for a Year 5 class. Include three word problems. Order them from easiest to hardest. Give the answer key separately at the end." The second version states the grade level, the specific sub-skill, the mix of formats and the exact output structure, rather than leaving the model to guess at difficulty and format — how to write a prompt that works on the first try covers that same discipline more generally.

Check every problem and every answer before it reaches a student. A worksheet that is factually wrong, or pitched a year above where your class actually is, does more harm than the time it saved — checking an AI answer when you are not the expert is the general version of that habit, and it matters here specifically because the person reading the mistake is a child who has no way to know it is one.

Where it stays your call

The techniques with the strongest evidence behind them — spaced practice and retrieval practice — work because a teacher notices which specific gap a specific student has and times the right question for them, not because a worksheet exists in the abstract. Deans for Impact's synthesis of the learning-science evidence makes the same point: the evidence describes what works when a person applies it to a particular class, and that application is not a step a general-purpose tool can do for you.

A model can generate the practice questions faster than you can write them. It cannot tell you, reliably, which one this specific child needs next.

The APA's guidance on practice for knowledge acquisition covers the same evidence from the psychology side: distributed, targeted practice beats generic drilling, and 'targeted' is the part that needs a person who knows the student, not a tool that only knows the topic.

Where it goes wrong

• —A fact, a date or a worked solution stated with total confidence and wrong. A widely cited survey of hallucination in large language models covers why fluent and accurate are not the same property in these systems — a mistake in a worksheet answer key reads no differently from a correct one until a student gets marked wrong for a right answer.
• —Student work or identifying details pasted into a tool with no agreement covering how that data is used — check what your school has already decided before putting any student-specific material into a general consumer tool.
• —Feedback copied straight from a tool without being read against the actual piece of writing, so a comment ends up generic or, worse, addresses a mistake the student did not actually make.
• —A generated worksheet or explanation treated as finished rather than as a draft pitched at an average student who does not exist in your actual class.

What to do this week

1. 01Find out what your school has already said about AI tools and student data — many now have a written policy, worth reading before assuming anything is fine to use.
2. 02Pick one purely preparatory task — a worksheet, a practice-question bank, a parent email — and try it for a week, checking every output against your own knowledge of the material and the class.
3. 03Keep the time actually saved in mind rather than a general impression that it helped, and spend that time on the part of the job a tool cannot do: watching how a specific student works.
4. 04Find the repetitive part is a short test for which other prep tasks are worth handing over once the first one is routine.
5. 05If your school is also evaluating a platform that bundles AI into its courseware, AI-powered LMS covers what that label actually promises versus what it does not — a separate decision from how you personally use a general tool for prep.

The short version

Use AI for first-draft worksheets, practice questions, explanations and routine correspondence. Check every fact, every answer key and every piece of feedback before a student sees it, and never put identifying student information into a tool your school has not cleared. The prep work gets faster. Knowing this particular student does not move to software, and it is still the part of the job worth protecting the most time for.

That split between faster prep and protected judgement is also what the wider labour market has started pricing in — PwC has measured a 62% wage premium for workers with demonstrable AI skills, up from 57% the year before. Coursium teaches that practical layer directly: writing a specific, levelled request and the habit of checking the result before it reaches anyone else. Stay ahead of AI by learning the tools on your phone.