Invisible watermark against AI fraud cracked within a week

On 2 August, Anthropic introduced an invisible watermark on text generated by Claude, as an implementation of Article 50 of the European AI Regulation. The watermark subtly adjusts word choice using a cryptographic key, so that generated text can later be statistically identified without the reader noticing anything.

Within a week, several projects had already appeared on GitHub to remove the watermark. One of these, now with thousands of stars, removes the marking from text and from other file formats. The creators themselves conclude that a statistical watermark does not provide reliable proof of origin. An Anthropic employee also acknowledged that it is not a watertight solution, but rather a first step.

Dutch educational organisations have been warning for some time

Kennisnet writes about assessment and generative AI that detection does not yet provide a reliable basis for a judgement on fraud. Npuls observes that attention to AI in education now continues throughout the year, no longer limited to a single month, and that schools are looking for an approach that goes beyond after-the-fact checks. SURF has long advocated for greater transparency about algorithms in education instead of a black box that has to prove something after the fact.

The common thread is always the same. Detection and watermarks attempt to provide a technical answer to a question that is not technical. A pupil who paraphrases well will evade any detector, while a pupil who writes a little less skilfully is more likely to be suspected even though nothing wrong has been done.

Transparency instead of detection

At EduGPT, we consciously choose not to go down that path. Instead of trying to prove after the fact whether a text was created with AI, EduGPT makes the use of AI visible in advance. A pupil or student who has worked with an AI Companion does not need to hide it, and a teacher does not have to guess.

This transparency removes the shame that often surrounds AI use, and that is precisely where detection fails. As soon as AI use becomes a normal, visible part of the learning process instead of a secret that needs to be uncovered, the conversation between teacher and pupil about what was created with AI and what was thought up independently becomes much more valuable. If a school also consciously integrates AI Companions into the curriculum, it is clear exactly what assistance was provided and which part can therefore be attributed to the pupil themselves. In this way, the teacher’s judgement becomes traceable, rather than dependent on a watermark that turned out to be cracked within a week.

An example of how you can give AI a place in your curriculum can be seen in the video below:

Source

Anthropic, announcement of the watermark for Claude text, August 2026, anthropic.com. Kennisnet on assessment and generative AI in secondary education, kennisnet.nl. Npuls on AI in education throughout the year, npuls.nl. SURF on transparency of algorithms in education, surf.nl.