AI Watermarks Won't Solve the Cheating Problem Either

Anthropic is embedding hidden watermarks in Claude's text and files to flag AI-generated content. Even the company admits the watermark can't prove a student wrote something themselves.

August 19, 2026
AI Watermarks Won't Solve the Cheating Problem Either Education

Summary: Anthropic has begun embedding hidden watermarks in Claude's text and generated files, part of its compliance with the European Union's AI Act, which requires AI providers to identify synthetic outputs. But Anthropic itself acknowledges the watermark can't confirm text was human-written, doesn't work on short samples, and can't detect other AI tools used for editing or a different model entirely. Meanwhile Turnitin data shows more than half of scanned Australian university submissions show some AI use, and a growing number of institutions are abandoning detection technology altogether over accuracy and fairness concerns.

aiwatermarkingacademicintegrityanthropiccheatinghighereducationaidetection

Watermarking sounds like the clean fix detection software never was. It isn't.

Anthropic recently began embedding hidden watermarks in text generated by Claude, along with an additional watermark hidden inside files like Word documents and PowerPoint presentations.

The company added the feature after signing onto the European Union's AI Act, which requires AI providers to identify synthetic outputs. That's a compliance move as much as a cheating deterrent, and the distinction matters for how much weight schools should put on it.

What the Watermark Actually Does

The watermark isn't visible on the page. It's embedded invisibly in the text itself, and reading it requires a separate detection tool Anthropic hasn't released yet.

OpenAI built something similar for ChatGPT years ago. It's never shipped publicly, reportedly over concerns the tool could unfairly flag certain writers, including non-native English speakers, more often than others.

The Loopholes Are Already Built In

Anthropic's own documentation lists what the watermark can't do, and the list is long.

It can't confirm text was actually written by a human. It doesn't work reliably on short samples.

It can't tell whether AI was used just to proofread or suggest edits rather than write the whole thing. And it says nothing about text generated by a different AI system entirely, like DeepSeek.

Watermarking can also be stripped out using other AI tools. That's the same evasion problem that already undermines detection software.

It's the same weakness behind the tool that flagged a UC Berkeley professor's AI-assisted op-ed a few weeks ago, a story MAIN covered as an example of how unreliable these tools remain even when institutions rely on them for real decisions.

Detection Tools Are Being Abandoned Anyway

The scale of the underlying problem makes any single technical fix look small. A 2026 UK survey found 95% of undergraduates use AI, and 94% use it specifically in assessments.

Turnitin reported in July that more than 53% of scanned submissions from Australian university students showed some form of AI use.

Faced with those numbers, a growing number of institutions have dropped AI detection software entirely, citing accuracy problems, bias, and successful legal challenges from students who contested false flags.

A Different Approach Entirely

Some researchers argue the whole framing is wrong. University of Queensland educational psychology expert Jason Lodge put it directly.

"Stop looking for evidence that students are using these tools to cheat and shift our emphasis to looking for evidence that learning has occurred," Lodge said.

That shift is already showing up in policy guidance. Rather than policing every assignment for AI use, some regulators are pushing toward assessments that build in complexity across an entire degree.

Alongside that, guidance increasingly calls for at least one "secure," identity-verified task per course, like an oral presentation or supervised in-class demonstration.

What This Means for Miami

The University of Miami's current approach, leaving AI policy to individual instructors rather than one blanket rule, looks more defensible in light of this. No detection or watermarking technology available today can reliably prove what a specific piece of writing was made with, so leaning on instructor judgment and clearly stated expectations may be more honest than pretending a tool can settle the question.

For South Florida's colleges and universities, the practical lesson from both this story and the Berkeley professor's case is the same. Investing in detection or watermarking infrastructure is likely to buy less certainty than it promises, and redesigning how assessments actually verify learning is the harder, but more durable, fix.

That redesign work is already underway elsewhere, and it doesn't have to mean abandoning take-home work entirely. A single verified, in-person component per course, paired with assignments that build on each other across a semester, gives instructors a way to confirm understanding without needing a piece of software to vouch for authorship it can't actually guarantee.

← Back to MAIN