AI Watermarks Are Coming. What Will Universities Do With Them?

From 2 August 2026, Article 50 of the EU AI Act introduces new transparency obligations for AI-generated content. Providers of generative AI systems are expected to make synthetic output machine-readable and detectable where technically feasible. The regulation covers more than text, extending to images, audio and video, and is part of a broader European attempt to make AI-generated material easier to identify.

The regulatory direction is clear. The technical implementation is not.

Major AI companies (Antropic, OpenAI) are moving toward different provenance and watermarking mechanisms, but there is still no single standard for text. It is also unclear how these signals will be checked, how reliable they will be after editing, and, most importantly for students, how universities will interpret them.

That last question may prove more consequential than the watermark itself.

A probabilistic signal is not proof

According to Nick Norman, whom I interviewed about the new requirements, the crucial qualification in the regulation is simple: content should be marked “where technically possible.”

That qualification matters because watermarking ordinary text is fundamentally more difficult than attaching provenance information to an image or video file.

With text, Norman explains, the signal is likely to be probabilistic. It may become stronger when a model generates a long continuous passage, but it can weaken when the text is edited, shortened, translated, divided into fragments or passed through another model.

The opposite problem is equally important.

“Human-written text can potentially trigger the marker as well,” Norman says.

For universities, this creates an obvious problem. Even if a detector identifies some statistical evidence of a watermark, what does that evidence actually mean?

Norman gives a simple example: “Maybe a professor sees that a detector gives a 17 percent probability of a watermark being present and starts penalizing students.”

Is 17 percent meaningful? What about 30 percent? Or 60 percent? At what point does a probability become sufficient evidence for an accusation of academic misconduct?

There is currently no common answer.

Students may change how they write

The danger is not limited to false accusations. Detection systems can also change student behaviour before any accusation happens.

Norman recalls a university group assignment in which one student was so concerned about AI detection that every paragraph was repeatedly checked and rewritten. Some members of the group had used GPT in parts of their work, while others had written independently. But the detector itself became the dominant constraint on the final text.

“Eventually we weren’t even writing in our own natural style anymore,” he says.

This is already a serious educational problem.

If students begin optimizing their writing for detectors rather than for clarity, argumentation or academic quality, the assessment system starts measuring the wrong thing. Instead of asking whether an argument is strong, a source is used correctly or a student understands the subject, students start asking whether a paragraph looks sufficiently human to an algorithm.

The new watermarking systems could intensify that behaviour.

The vocabulary paradox

There is another, more counterintuitive possibility.

According to Norman, some text-watermarking approaches depend not only on length but also on linguistic variability. A richer vocabulary gives the model more token choices, potentially making it easier to encode a detectable statistical pattern.

That creates a strange incentive.

A student who writes with a broader vocabulary and more sophisticated sentence structures could theoretically produce text that generates a stronger signal than a student who deliberately simplifies everything.

Norman formulates the problem directly: “What happens if texts become extremely primitive and simplistic simply because people are trying to comply with, or avoid being flagged under, this regulation?”

The result would be perverse. A transparency mechanism intended to improve accountability could indirectly encourage weaker academic writing.

And once academic or professional consequences are attached to watermark signals, another market will almost certainly appear around them: tools designed specifically to weaken, obscure or remove those signals.

Rewriting, translation, simplification, paraphrasing and passing text through another model may already interfere with detection. Dedicated “AI watermark removers” would simply formalize that process.

At that point, the system risks becoming similar to another familiar European regulatory phenomenon: cookie banners. The mechanism becomes nearly universal, but users learn to bypass, ignore or mechanically interact with it, reducing its informational value.

The deeper educational problem

This connects directly to a distinction I discussed in an earlier Canvas Assistant article about how students use AI.

The real divide is not simply between students who use AI and students who do not.

Some students already have a developed academic method. They know how to read sources, evaluate information, structure an argument and formulate their own conclusions. For them, AI can remove routine friction from the process. It can help organize material, clarify ideas or improve efficiency without replacing the underlying intellectual work.

Other students use AI differently. Without an established academic method, they can outsource the cognitive process itself: reading, synthesis, reasoning and sometimes even the formulation of the final argument.

In the second case, AI does not merely assist learning. It can replace it.

The problem with a rigid detection regime is that it may treat these two behaviours as equivalent.

If universities respond to AI misuse by restricting AI-assisted work for everyone, they risk designing the educational system around its weakest use case. Students who know how to use AI productively may lose tools that genuinely increase their efficiency because institutions are trying to control students who use the same technology to avoid thinking.

That would not solve the underlying educational problem. It would simply lower the ceiling for responsible users.

What changes for Canvas Assistant users?

For now, these developments do not directly affect the core functionality of Canvas Assistant: downloading educational video, creating transcripts, structuring lecture content and generating summaries.

Any practical effect will depend on two things that are still unresolved: the technical standards eventually adopted by AI providers and the policies universities create around those standards.

And that is the central issue.

The EU now has a transparency requirement. What it does not yet have is a universal technical standard, a perfectly reliable method for identifying AI-generated text, or a shared academic framework for interpreting detection signals.

For students, the important question is therefore no longer simply whether AI-generated content can be marked.

It is what universities will decide those marks actually mean.