Claude's watermark is a compliance artefact, not a cheating detector

Anthropic has said the text Claude produces will carry a hidden label marking it as AI-generated, and that files Claude generates, including Word and PowerPoint documents, will carry an additional watermark embedded in the file itself ^[1][2]. The company took the step after signing on to the European Union's AI Act, which obliges providers of systems that create synthetic audio, video, images or text to identify those outputs as AI-generated ^[3][4].

The important distinction is who is doing the asserting. AI-detection software infers authorship from the text, and reports have found those tools patchy and far from foolproof ^[5]. A watermark is a signal the producer deliberately inserted at generation time. It is not an image laid over the page but is invisibly embedded through the text, and reading it requires a detection mechanism that Anthropic has not yet released ^[6][7].

The limits are unusually well documented by the vendor. Anthropic warns its watermarking cannot tell whether text was human-written ^[8]. It does not work on small samples ^[9]. It cannot say whether AI was used to proofread, suggest edits or improve a draft ^[10], and it cannot tell whether a different system, such as DeepSeek, wrote the text ^[11]. Like detection, it can be evaded using other AI tools, and the absence of a marker does not establish that AI was not used ^[12][13]. The signal is therefore one-directional: a hit is evidence about one vendor's output, while a miss carries no information at all ^[14].

That asymmetry matters because the base rate is high. A 2026 UK survey found 95 percent of undergraduates use AI and 94 percent use it in their assessments, a gap of about one percentage point between any use and assessed use ^[15][16][17]. A Turnitin report in July found more than 53 percent of submissions from Australian university students run through its system used some form of AI ^[18]. When almost everyone is using the tools, a marker that confirms use answers a question nobody needed answered and cannot separate permitted use from unauthorised use.

Institutions have already been retreating from the detection model. An increasing number have abandoned the software over accuracy, bias and procedural fairness, and after successful legal challenges from students overseas ^[19]. Some Australian educators argue detection-based approaches to integrity are not the way forward, because the focus falls on policing whether AI was used rather than on what the student learned ^[20]. Jason Lodge of the University of Queensland has put it as a shift from looking for evidence that students are cheating to looking for evidence that learning has occurred ^[21].

The policy responses so far pull in the other direction. New South Wales moved last week to ban take-home assignments for Year 12 to stop AI use in assessments ^[22]. A return to in-person assessment does not work for students studying fully online, and does not prepare them to engage with AI ^[23]. The higher education regulator's position is that students need preparing to engage actively, responsibly and ethically in a society where AI is everywhere, and that judging learning requires assessment designs that are fair, inclusive and fitted to the situation, including tasks that connect and build in complexity across a whole degree ^[24][25][26].

Two things to watch. First, whether Anthropic ships the detection mechanism, and to whom, since a watermark nobody can read is a compliance record rather than a usable check ^[7]. Second, whether OpenAI releases the text watermarking tool it has built but withheld, reportedly in part over concerns it could disadvantage groups including non-native English speakers ^[27][28].