Ask a teacher what they’d do with an extra hour on a Tuesday night and the answers are rarely glamorous. Mark three more essays. Build a differentiated version of Thursday’s worksheet. Phone a parent about a detention. The promise of AI tools for teachers isn’t a robot that stands at the front of the room. It’s clawing back that hour.
That’s the honest framing. Most of what’s on the market right now is a planning assistant, a drafting assistant or a marking assistant. Knowing which job each one does well saves you from the trap of signing up for fourteen platforms and finishing the term using two.
Where AI Actually Saves Time in a School Week
Talk to teachers who’ve stuck with these tools past the novelty phase and the same three jobs keep coming up.
- Planning and differentiation. Turning one worksheet into three versions at different reading levels used to be a 40-minute job. It’s now a five-minute editing job.
- First-pass feedback. Not the grade itself, but the scanning: spotting the twelve students who never answered the second part of the question.
- Admin. Parent emails, trip letters, report comments, the department newsletter nobody has time to write.
Notice what isn’t on that list: teaching. Every credible trial and every slightly exasperated head of department says the same thing. The value sits in preparation and paperwork, not in delivery.
Differentiation is the strongest use case
If you only try one thing, try this. Paste a text into a chatbot and ask for two rewrites: one at a lower reading age with shorter sentences and glossed vocabulary, one with the inference left implicit for your high attainers. Add a line like “keep the same facts and structure, do not simplify the science.” The output won’t be perfect, but it gets you to 80% in under a minute, and the last 20% is the part you’re actually qualified to do.
The same trick works for scaffolding. Ask for a writing frame, a sentence-starter bank, or a set of six comprehension questions ranked by difficulty. What used to be a Sunday afternoon of fiddling becomes a quick review pass.
The Tools Worth Knowing About
General chatbots
ChatGPT, Claude and Gemini do the heavy lifting, and most teachers never get past the free tier. The step change comes when you upload a document rather than typing a vague prompt. Drop in your scheme of work, a past paper and the mark scheme, then ask for a lesson sequence that targets the two weakest question types. The answer improves dramatically because the model finally has something concrete to work with.
If you teach computing, subject-specific assistants are worth a look too. This breakdown of what GitHub Copilot does well and where it still fails is a reasonable starting point before you let students near it.
Purpose-built classroom tools
MagicSchool, Diffit, Brisk and Eduaide wrap the same underlying models in teacher-shaped buttons: text leveller, rubric generator, IEP-language helper, exit ticket maker. The advantage is speed and a cleaner interface. The trade-off is less control, and free tiers that cap how much you can generate per month. Most offer a school licence, which is where the data protection conversation starts.
Quizzing and retrieval practice
Quizizz, Kahoot and formative-assessment add-ons will generate a 20-question retrieval quiz from a chunk of text in about 30 seconds. Treat the output as a first draft. Roughly one question in five needs rewriting, usually because it’s ambiguous or has two defensible answers, and a bad multiple-choice question causes more confusion than no question at all.
Videos and Visuals Without a Media Budget
Explainer clips used to mean a weekend with a screen recorder and a failing microphone. Text-to-video generators have collapsed that. A tool like the one reviewed in this InVideo AI walkthrough will take a paragraph of script and return a narrated, captioned clip suitable for a flipped-classroom task or a cover lesson. Quality varies, and you’ll want to check every fact and every stock clip before it reaches a class, but for a two-minute recap of the water cycle it’s genuinely good enough.
Image generation is the other quick win. Diagrams of a cell, a timeline visual, a stylised map of a trading route, a cover image for a display board. This guide to generating images inside ChatGPT covers the prompting basics, including the important habit of specifying “no text in the image” so you don’t end up with nonsense labels on a diagram of the digestive system.
The Marking Problem Nobody Wants to Talk About
Here’s where enthusiasm meets reality. Teachers are rightly worried about students submitting AI-written homework, and plenty of schools have responded by running essays through a detector. The trouble is that detectors are statistical guesses, not proof. Several studies have found false-positive rates high enough to be alarming, and they fall hardest on students writing in a second language, whose phrasing is more predictable.
Accusing a Year 11 student on the strength of a percentage score is a serious thing to do. A sensible approach to AI writing detectors treats the result as one weak signal among several: a conversation about the work, a handwritten in-class paragraph, knowledge of how that student usually writes. That sequence protects both the student and you.
There’s also a practical route that sidesteps the whole argument. Design tasks that are hard to outsource. In-class drafting, verbal defence of a piece of work, process logs, questions tied to a specific lesson discussion. AI is weakest exactly where real learning is strongest.
Before Your School Signs Up: The Questions That Matter
Free tools are free because your data, or your students’ data, has a value. Before anything goes on a whole-school licence, get straight answers on these.
- Where is the data stored, and is it used to train future models? Under UK GDPR and similar regimes, you need a data processing agreement, not just a terms-of-service page.
- What is the minimum age, and does the student-facing version differ from the teacher version? Many consumer tools set 13 or 18; education tiers often sit separately.
- Who reviews outputs? A human in the loop is not optional for anything that reaches a student or a parent.
- What happens when it’s wrong? A hallucinated date, a misattributed quote, a maths step that doesn’t work. Someone has to own the check.
And a personal rule worth adopting regardless of school policy: never paste a student’s name, SEN record or safeguarding note into a public chatbot. Strip it to the content and re-add the identity yourself.
What It Looks Like in Practice: One Unit, Three Hours Back
A Year 8 geography teacher preparing a three-week unit on rivers might work like this. Ten minutes uploading the scheme of work and asking for a lesson sequence with two retrieval starters per lesson, then editing out the two suggestions that don’t fit the department’s textbook. Fifteen minutes generating a differentiated reading at three levels with a glossary. Five minutes on a labelled diagram prompt, redrawn by hand on the board because the AI version had a wonky tributary. Twenty minutes turning the end-of-unit assessment into a quiz, with half the questions rewritten.
Total: under an hour, against an evening. Nothing in that list required the teacher to hand over judgement. The AI drafted; she decided.
Teaching Students to Use These Tools Honestly
Banning is a losing position, and students know it. The more useful move is teaching the difference between using AI to think and using it to avoid thinking. Show them a chatbot confidently inventing a source in front of the class, then show them how to check it. Ask them to paste a paragraph and critique it, line by line, as an exercise in editing rather than writing.
Set explicit rules for each task type: no AI for the timed essay, AI allowed for brainstorming in the project, AI encouraged for feedback on a draft you’ve already written. When students know which mode they’re in, the anxious guessing disappears and so does most of the cheating.
The teachers getting the most out of this technology are not the ones with the longest list of subscriptions. They’re the ones who picked two or three tasks they dread, automated the boring part of each, and kept every decision about what a student actually needs in their own hands.