The question German staffrooms are actually arguing about is not whether AI belongs in school. It is narrower and more awkward than that: what should a student produce individually, what may be supported, and how can performance still be graded fairly?
I do not think there is a general answer. But there is a workable local one, and it is what I have built into the timed writing assessments I use with my own secondary classes — with the same infrastructure now being extended to vocational and university writing tasks. It is offered as a position to argue with rather than a finished solution.
Two pressures, not one
The discussion usually gets framed as an integrity problem. It is really two problems that happen to have arrived together.
Marking load. The Philologenverband put it starkly this spring: teachers report correcting themselves to death on Klassenarbeiten and Klausuren, with little time left for preparation or individual support. Anything that adds verification work on top of existing correction load will not be adopted, however principled.
Unverifiable evidence. Written work produced at home no longer reliably demonstrates what a student can do unaided. This is a measurement problem before it is a moral one. A task that cannot distinguish a student’s English from generated English has stopped functioning as assessment, regardless of how well designed it is.
Solutions that address only one of these tend to make the other worse. Requiring handwritten drafts increases marking. Abandoning written assessment for oral examination is defensible but expensive in contact time — and it is where much of the current discussion is drifting by default rather than by decision.
The distinction that does the work
The single most useful move I have made is to stop treating all tasks as the same kind of thing.
Practice is open. Vocabulary work, comprehension, drafting, revision — students may use dictionaries, translation tools, language models, each other. The learning happens in the attempt. Restricting tools here achieves nothing except teaching students to conceal what they are doing. For adult vocational learners, pretending otherwise destroys credibility immediately; using a model to check a draft ticket is a legitimate professional workflow.
Evidence is invigilated. Where a mark contributes to a report, a certificate, or a decision, the task runs under conditions that make unaided production plausible.
Students are told explicitly which category any given task belongs to. In my experience the ambiguity is what generates resentment, not the invigilation. A student who knows the rules of a task rarely objects to them; a student who suspects the rules are unenforceable learns something worse than cheating — that the mark was never real.
This distinction also solves the marking-load problem, because it turns out most tasks do not need securing. Only the ones producing a number someone will act on.
What the invigilated pages actually do
The timed writing assessment pages are static HTML with client-side integrity measures. Grouped by what they are for rather than listed flat:
Preventing pasted answers
- The paste event is blocked
- Ctrl/Cmd+V is blocked
Preventing the prompt being copied out
- Scenario and vocabulary boxes are set non-selectable
- Right-click is disabled during the exam
Detecting attention leaving the task
- Tab and window focus changes are detected via
visibilitychangeandblur - An on-screen warning appears when the student returns
Detecting inspection or tampering
- F12, Ctrl+Shift+I, Ctrl+Shift+J and Ctrl+U are blocked
- A DevTools size heuristic runs every four seconds
Detecting non-human input patterns
- Typing-speed anomaly detection flags bursts above sixty characters in under half a second
Enforcing task compliance
- A time gate disables submission for the first ten minutes of a sixty-minute task
- Word count is enforced — submission is disabled outside the accepted range
- Required vocabulary is checked at submission, so target language must actually appear
Making copying from a neighbour pointless
- Four scenario variants, assigned by a hash of the student’s name
Creating a record
- Progress snapshots every three minutes, submitted with the final text
- An integrity payload accompanies the work: tab switches, paste attempts, anomaly flags, DevTools detection, user agent, session ID
What this does not do
A checklist presented as a solution would be dishonest, so:
It is not prevention. A phone beside the laptop defeats every item above. Nothing implemented in a browser can prevent a student reading from a second screen. What the measures remove is the frictionless path — the tab-switch-and-paste that takes four seconds and requires no planning.
It is not proof. A typing anomaly means a burst of fast input occurred. That might be a paste through an unblocked route; it might be a student who touch-types and had a sentence ready. The flag is grounds for a conversation, never a verdict. Automated integrity decisions would be both unjust and, in a school context, indefensible.
The human stays in the loop. This is the phrase circulating in the German debate and it is the right one. The data informs a teacher’s judgement. It does not replace it, and a system that tried would fail on its first false positive.
The quiet measure is the strong one. Of everything on that list, scenario randomisation by name hash is the most effective and the least visible. Adjacent students write different prompts. Unlike the blocking measures, it cannot be circumvented by a second device, and unlike the detection measures it produces no false positives. Prompt design does more integrity work than surveillance does — a prompt concrete and specific enough that a generic answer is obviously generic is worth more than any keyboard block.
The Datenschutz question, since it will be asked first
Any German colleague reading that feature list will reach the same question immediately, and they are right to.
The integrity payload is behavioural data about a named minor. Tab switches, session identifiers and user agent strings are personal data. Collecting them requires a legal basis, a retention decision, and transparency toward students and parents — not merely a page that works.
Concretely, this has meant:
- Telling students in advance, in specific terms, what is recorded. This is both a legal requirement and a pedagogical one: a fifteen-year-old who discovers mid-task that their tab switching is logged will reasonably feel ambushed, and will be right.
- Collecting only what informs a judgement. Anomaly counts and flags, not keystroke logs.
- Migrating the backend. Submissions originally posted to Google Apps Script with an email fallback. Following a data-protection restriction on Google services at school, that is moving to Excel Online and Power Automate. Given how much of the current debate concerns Datenschutz, building for that constraint rather than around it has proved the right decision, if a laborious one.
Anyone adopting measures like these should treat the data question as part of the design rather than a compliance step afterwards. It changes what you build.
What I would say to a colleague considering this
Start with prompt design, not blocking. A specific, situated prompt with a named context and a required vocabulary set does most of the work, costs nothing, and produces no privacy questions.
Then secure only what needs securing. If you find yourself invigilating everything, the category distinction has collapsed and the marking load is about to return.
Tell students exactly what is happening and why. The argument — that a mark only means something if the conditions were clear — is one most students accept when it is made honestly.
And be careful about the claim you make for it. The measures buy the ability to keep written production as a meaningful assessed component instead of retreating entirely to oral examination. That is a real gain. It is not the same as having solved anything.
The position, stated plainly
The current framing treats integrity measures and trust as opposites. I do not think they are. Clear conditions are what make a mark meaningful, and a meaningful mark is what makes feedback worth reading.
The question was never whether students will use AI. It is which tasks are supposed to prove something — and whether those tasks are actually built to do it.
This is the first of five posts. The remaining four cover secondary curriculum exercises, vocational IT English, Abitur writing preparation, and MSA exam practice — each of which refers back to the framework set out here.
Improve your English skills 