Does ‘AI‑watermarking’ mean the party is over for cheating students?

Concerns about within schools and universities.

Some lecturers have even students have been “lobotomised by AI” amid reports are patchy and far from fool-proof. Last week, assignments for Year 12 to try to stop AI use in assessments.

At the same time, a major generative AI system has developed a way to “watermark” whether or not a text is AI. Developer Anthropic, has the text output of Claude (one of the five big AI tools) will include a hidden label or watermark indicating the text is AI-generated.

Files generated by Claude, including Word and PowerPoint, will also contain an additional watermark hidden in the file itself.

What does this mean for efforts to stamp out cheating?

Read more:

Why a watermark?

Anthropic took this step after signing up to the European Union’s . The act requires providers of AI systems that create synthetic audio, video, images or text to identify the outputs as AI-generated.

Other tech companies, (which runs ChatGPT), have developed a text watermarking tool, but not rolled it out publicly.

The watermarking is not an image superimposed over the text. It is . Readers will need a “detection mechanism” (yet to be released by Anthropic) to pick up the watermark.

If this sounds too good to be true, that’s because it is.

There are still loopholes

As with AI detection, watermarking using other AI tools. The absence of a marker also doesn’t guarantee that AI wasn’t used.

Even Anthropic their watermarking cannot tell “whether the text was human-written”. It doesn’t work on small samples of writing. It cannot tell whether AI was used to proofread the work, suggest edits or make improvements. It cannot tell whether a different AI system, such as DeepSeek, wrote the text.

OpenAI has withheld its text watermarking tool, in part, due to concerns it could negatively impact certain groups, including non-native English speakers.

AI use has been normalised

We know AI use is now extremely common among students. The tools are readily available and free to use.

A found 95% of undergraduate students use AI, with 94% using it in their assessments.

Unauthorised AI use is most common in written assessments completed outside of class. Educators have been using AI detection software to try and screen work for cheating.

A from software company Turnitin revealed more than 53% of submissions from Australian uni students run through its system used some form of AI.

However, an increasing number of institutions have abandoned this technology amid , and successful from students overseas.

What do we need to do instead?

So it would be unwise for teachers or institutions to rely on watermarking or detection tools to guard against cheating. These technologies can provide a sense of false security. But where does that leave us?

Some “detection-based approaches to integrity are not the way forward”. This is because the focus is on policing whether AI was used, rather than on what the student has learned.

Or as University of Queensland educational psychology expert has also , we need to:

stop looking for evidence that students are using these tools to cheat and shift our emphasis to looking for evidence that learning has occurred.

What happens now?

Advice on alternatives to the “detect and punish” approach has been steadily growing over the past three years. Some of this is via the . It notes:

• we need to prepare students to engage actively, responsibly and ethically in a society where AI is everywhere
• understanding what students have learned requires different types of assessment that are fair, inclusive and suited to the situation.

Concerns about cheating and AI-use have also resulted in calls to return to in-person assessments. But these for those who study fully online. They also don’t prepare students to engage with AI.

The higher education regulator has also how assessments could change in practice. This includes assessments that are connected and build upon each other throughout a whole degree, increasing in complexity. It also includes at least one “secure” assessment task in every unit or subject across a degree program.

Secure tasks can include oral presentations, in-class tasks and supervised practical demonstrations. They must also include mechanisms that verify the person completing the assessment is the enrolled student.