A history teacher at a secondary school pastes twenty test questions into a free tool on a Sunday evening and gets them back at four levels. On Monday a colleague in the staff room shares an extension that summarises lesson videos. Neither reported anything, and neither broke a rule, because no agreement covers this.
That is shadow AI in education: AI use that staff started themselves, without a decision from the school. It comes from three directions, all of them out of ICT's sight: free tools on a personal account, AI features that appear inside learning tools already in use, and browser extensions. What the school needs is not a ban, but a view of which tools are opened, a decision per tool, and that decision at the moment someone opens the tool. No teacher has to be tracked for that.
Where does shadow AI in a school come from?
The general definition is in what is shadow AI: self-serve, no decision, actually in use. In schools it takes three recognisable forms.
| Where it comes from | Example | Why it stays out of sight |
|---|---|---|
| Free tool on a personal account | The free version of ChatGPT, MagicSchool, Curipod, a translation service | No procurement, no school sign-in, often at home on a personal device |
| AI feature in a learning tool already in use | A pupil tracking system or a method that adds a summarising or marking feature | Same domain, same data processing agreement, new processing nobody reassessed |
| Browser extension | An extension that summarises videos, rewrites emails or checks text | Installed in two seconds, with read access to every page, with no admin involved |
The first form is the best known and the easiest to see. The second is the quietest: the tool was already in the register as allowed, and the new feature arrived with an update. Why that is a trigger for reassessment in its own right is explained in why tool catalogues age. The third form is skipped almost every time, even though an extension with read access is an AI tool too, and one with a wider reach than the chatbot itself.
Why does it grow in schools in particular?
For the same reason as everywhere: the work gets lighter. Feedback on thirty essays, a test at three levels, a letter to parents that reads without jargon. The allowed tool usually exists, but on a Sunday evening it does not feel adequate, or the teacher simply knows it less well than the tool that was already open at home.
Three things make a school different from an office. Staff work a lot from home and on personal devices, so the school network sees little. The data passing through belongs to children, with a care note or a development plan as the heaviest category. And the school works with data processing agreements through the sector's privacy covenant, while a free account runs under consumer terms without such an agreement. In education, that is the first question for every tool that surfaces, before the question of what it can do. The Dutch Data Protection Authority reported several data breach notifications in which employees had put personal data into an AI chatbot, and advises organisations to make clear agreements on whether it is allowed and which data may go in.
A ban does not solve this. It does not remove the tool, only the tab you could still see, and the use moves to the phone and the personal account. What remains is exactly the account type you did not want.
Why does a quarterly survey fall behind?
A board with twelve schools asks its ICT coordinators in October: which AI tools are used at your school? The list comes to seventeen tools, six of them without a decision. The board assesses them, records the statuses and circulates the overview in December.
By January there are three new tools, one of which turned up at four schools at once after a webinar. A learning tool that sat in the register as allowed has gained a feature in the same period that comments on essays. And the extensions were not on the October list in the first place, because nobody recognised them as AI tools.
That is not a failed survey. It shows that a list people fill in from memory is a snapshot, while the use keeps moving. Which methods exist and what each of them misses is set out in how to discover shadow AI.
How do you keep track without tracking teachers?
By limiting the measurement to the question the school actually has: which AI tools are opened, and which of them have no decision yet? For that question the domain of the service is enough. That an AI tool was opened at the board's schools this month, which one, and how often. Who did it adds nothing, and as soon as you record it anyway the measurement changes from a count of services into a record of people. That difference, and what it means for the GDPR and staff representation, is in AI monitoring versus employee monitoring.
BeeSensible therefore records only the registrable domain, at most once per day per device, with no user id, no path and no page content. A new tool lands in an inbox with the status "no decision yet"; the board's administrator assesses it and records the decision. This is what that register looks like:
