Where does a teacher's week actually go? Not into lesson planning. It goes into grading, into feedback, into differentiating one worksheet six ways, into emails to parents, into administrative forms nobody reads. Yet almost every AI-for-teachers listicle opens with lesson planning — because lesson planning demos beautifully. Type a topic, watch a plan appear, feel the future arrive.
So this ranking is by hours returned, not by demo quality. Feedback and grading first. Differentiation second. Lesson planning third, because a generated lesson plan still needs to be checked, aligned to your standards, adapted to your class and rewritten in your voice — which is most of the work you were trying to avoid.
And a warning up front, because it is the question every teacher eventually asks: AI detectors do not work reliably enough to accuse a student. That section is at the bottom, and it is the most important part of this piece.
Tier one: feedback and grading (the real hours)
This is the pile that eats evenings. It is also, awkwardly, where AI is most useful and most legally fraught.
Rubric-driven feedback drafts
The workflow that works: give a model your rubric, one or two exemplar responses with the feedback you would have written, and then the student's work. Ask for feedback comments aligned to each rubric criterion, in your register, with specific reference to what the student actually wrote.
What you get back is a first draft of comments — often decent on structure and evidence, reliably generic on the things that make a piece of writing that student's. You edit. You cut the flattery. You add the one observation that proves you read it.
The rule that must not bend: the model does not assign the grade. It can describe. It can suggest where a piece sits against a criterion. But the judgment, and the accountability for it, is yours, and it needs to remain yours in fact and not merely on paper.
The privacy problem you cannot skip. Student work is student data. In the US, FERPA applies; in the UK and EU, UK GDPR and GDPR do; most districts and schools have their own approved-vendor list and a signed data-privacy agreement with each one. Pasting an identifiable student's essay into a consumer chatbot account is very likely a breach of one of those, whatever your personal view of the risk. The practical minimum: use tools your district has approved, strip names, and ask your data protection lead before you build a habit.
The tools worth knowing
Brisk Teaching runs as a Chrome extension over Google Docs — it gives inline feedback, and it can show a document's writing process through revision history, which is a far more honest signal than a detector score. MagicSchool packages dozens of teacher-specific generators behind an education-focused product with school agreements, which is precisely why teachers use it instead of a raw chatbot. Neither has a page on ToolsPantry, and both are worth your attention anyway.
For the underlying model, ChatGPT and Claude both handle long student texts and rubrics well. Claude in particular tends to be more literal about following a rubric and less inclined to be nice about a weak argument, which for feedback is a feature.
Who this tier suits: any teacher with a large marking load and standardised rubrics — English, humanities, anything essay-shaped. Who it does not: teachers of very small cohorts, or of subjects where feedback is inherently mathematical or diagrammatic. Setting up the workflow costs more than it returns at that scale.
Tier two: differentiation (the underrated one)
Differentiation is where the technology is genuinely, quietly transformational, and almost nobody leads with it — because it is invisible work. Nobody sees the three reading levels. They only see the one lesson.
Things that used to cost an evening and now cost minutes:
- Reading-level rewrites. One source text, three versions, same content. This is a task language models are close to ideal for, and the failure mode — an over-simplified version that loses a nuance — is one you can spot in ten seconds of reading.
- Home-language family letters. DeepL produces translations that are consistently more natural than the free general-purpose options, and communicating with families in their own language is one of the highest-value things in the building. Have a fluent human check anything consequential.
- Tiered tasks and scaffolds. Sentence starters, vocabulary supports, a version with the worked example still visible.
- Accessible variants. Reformatting a task for a student who needs shorter chunks, larger structure, or fewer simultaneous instructions.
Diffit and Curipod are built specifically for this and are worth a look. But a general model with a well-written prompt does most of it, and you already have one.
Tier three: lesson planning and materials (fun, overrated)
Generating a lesson plan is the demo everyone shows and the task that saves the least, for a boring reason: your plan has to align to your standards, your scheme of work, your class, and the sequencing you did three weeks ago. A model that knows none of those produces something plausible and generic, and editing it into usability takes about as long as writing it.
Where it does earn its place: as a divergence machine. Ask for eight ways into a topic, six discussion questions that would produce disagreement, four analogies for a concept students always get wrong, three ways this could go badly with a Year 9 class after lunch. That is idea generation, where being generic costs you nothing, because you are selecting rather than accepting.
Materials are a better use than plans. Gamma turns an outline into a presentable slide deck faster than any tool in this category and is genuinely a time-saver when you need a deck by Thursday — see AI presentation tools for the alternatives. Canva's education tier remains the pragmatic default for worksheets and displays.
Tier four: student-facing tools
These do not save you hours directly. They change what happens when you are not in the room.
Khanmigo is Khan Academy's tutor, notable mainly for a design choice the rest of the field mostly did not make: it is built to be Socratic and to withhold the answer, which is the difference between a tutor and a cheating machine. It also has teacher-facing tools. Who it suits: schools already inside the Khan ecosystem.
Socratic is a homework-help app that leans towards explaining rather than solving. Watch out for: the obvious. Any tool that will explain a problem will, with enough persistence, also do it.
Quizlet AI generates practice sets and study material from your content. Watch out for: generated distractors in multiple-choice questions are frequently poor — implausible, or accidentally correct. Read every one before it reaches a student.
Ranked by what it actually returns
| Task | Hours it genuinely returns | Where to do it | Watch out for |
|---|---|---|---|
| Rubric-aligned feedback drafts | High — the biggest single win | ChatGPT or Claude with your rubric; Brisk in Docs | Student data privacy; never let it assign the grade |
| Reading-level and tiered versions | High, and consistently underrated | Any capable general model; Diffit | Simplification quietly dropping a key nuance |
| Family communication in home languages | High, low risk | DeepL | Have a fluent human check anything consequential |
| Quiz and practice generation | Medium | Quizlet AI, Kahoot's generator | Weak or accidentally-correct distractors |
| Slide decks and materials | Medium | Gamma, Canva | Design over substance; check every fact |
| Lesson plans from scratch | Low — the demo lies | Any model, as an idea generator | Generic output; alignment is still all yours |
| Detecting AI-written student work | Negative. It costs you | Nothing reliable exists | False positives; bias against non-native writers |
Swipe the table sideways to see every column →
The detection question, answered honestly
This is the section that matters, so here it is plainly: AI text detectors have a documented false-positive problem, and you should not accuse a student on the basis of one.
The public record is not ambiguous.
- OpenAI withdrew its own AI Text Classifier in 2023, citing its low rate of accuracy. The organisation with the most direct knowledge of how these models write concluded it could not reliably detect their output.
- Detectors are biased against non-native English writers. A Stanford study by Liang and colleagues, published in Patterns in 2023, ran a set of TOEFL essays written by non-native English speakers through several widely used GPT detectors. A majority of those human-written essays were flagged as AI-generated, while essays written by native-English-speaking US students were classified far more accurately. The mechanism is intuitive once you see it: detectors partly key on low lexical variety and predictable phrasing, which is also what writing in a second language often looks like.
- Institutions have switched detectors off. Vanderbilt University disabled Turnitin's AI detection feature in 2023, citing accuracy concerns and the harm of false accusations. Others followed.
Now sit with the consequences. A false positive is not a rounding error in a metric. It is a teenager in a meeting, being told they cheated, with no way to prove a negative — and the students most likely to be in that meeting are disproportionately the ones writing in their second language. That is not a tolerable error rate. It is not close.
What to do instead
Design the question out of existence.
- Make the process the artefact. Google Docs version history, drafts, annotated outlines, in-class planning. A document that grew over two weeks is evidence in a way no detector score is.
- Move high-stakes writing into the room. Some of it, at least. In-class writing and oral defence of a submitted piece answer the question definitively and without accusation.
- Ask for things models are bad at. Something that happened in the classroom on Tuesday. A local reference. A connection to a text your class actually read. Personal reasoning grounded in specifics.
- Write an explicit AI policy per assignment. Not a school-wide slogan — a line at the top of each task saying whether AI is permitted, for what, and how to disclose it. Most academic-integrity conflicts are policy failures wearing a technology costume.
- If you have a genuine concern, talk to the student. Ask them about their argument. A student who wrote it can discuss it. This is old, unfashionable, and works.
A first term with AI, if you are starting now
- Pick one task. Make it grading feedback, because that is where the hours are.
- Check the approved-tools list before you touch student work. If you do not know who to ask, that is itself the finding.
- Build one strong prompt with your rubric and two exemplars. Reuse it. The setup cost is real and it is paid once.
- Never send unread output to a student or a parent. The moment you do, you have outsourced the relationship rather than the admin.
- Tell students what you use it for. The double standard — teachers use it silently, students get accused — corrodes trust faster than any single incident.
FAQ
What is the best AI tool for teachers?
There is no single one, and the honest answer depends on where your hours go. For grading and feedback — the biggest drain for most teachers — a general model like ChatGPT or Claude, driven by your own rubric, does more than any purpose-built product. For classroom-safe packaging with school data agreements, the education-specific tools like MagicSchool and Brisk exist for good reason. Start from the task, not the tool.
Can AI grade student essays?
It can draft feedback against a rubric, and it is genuinely useful at that. It should not assign the grade. Models are inconsistent across runs, are influenced by surface fluency in ways that disadvantage substantive-but-clumsy writing, and cannot be held accountable for a mark. Use it to write the comments faster; keep the judgment.
Are AI detectors accurate enough to catch cheating?
No. OpenAI withdrew its own detector in 2023 over accuracy, published research has found that detectors disproportionately misclassify writing by non-native English speakers as AI-generated, and universities including Vanderbilt have disabled AI detection tools over exactly these concerns. A detector score is not evidence. Use process artefacts, in-class work and a conversation with the student instead.
Is it safe to put student work into ChatGPT?
Not by default. Student work is protected data under FERPA in the US and GDPR in the UK and EU, and most schools maintain an approved-vendor list backed by a signed data-privacy agreement. A personal consumer chatbot account is almost certainly not on it. Remove identifying details, use school-approved tools, and ask your data protection lead before it becomes a habit.
Does AI actually save teachers time on lesson planning?
Less than the demos suggest. A generated plan still has to be aligned to your standards, fitted to your sequence, adapted to your class and rewritten in your voice — and that is most of the job. It is genuinely good at idea generation: analogies, discussion prompts, alternative approaches to a concept students keep missing. Treat it as a colleague throwing out ideas, not as a planner.
What free AI tools can teachers use?
The free tiers of the major models cover most of the ground in this piece, and the free tools filter surfaces the wider set. Khan Academy's materials remain free, DeepL has a usable free tier for family communication, and Canva's education tier is free for eligible teachers. Check your school's approved list before adopting any of them for work involving student data.
Where to go next
The full shelf is in AI education tools, the teachers hub collects what is relevant to classroom practice, and the students hub is worth reading precisely because it shows you what your class is already using. Found something that returns real hours rather than demo-hours? Submit it — this is a category where the loudest marketing and the most useful tools are rarely the same product.