Anyone who uses the internet will have noticed an increasing number of AI-generated images, videos and text posts on social media. Some are obviously made using AI, but others are harder to spot. Indeed, the use of chatbots by students to produce homework has become a major problem in academic institutions. In July 2026 a professor of economics at Brown University became suspicious when his 86 students managed a 96% class average in midterm exams. When he moved the final exam to be strictly in-person, the average dropped to 48.6%. Only two students scored within 10% of their midterm score. AI images and videos of people, which even a couple of years ago would frequently feature human bodies with extra fingers, and even the odd extra limb, now appear very realistic to all but the trained eye.
Governments are gradually reacting with legislation. China passed a law in September 2025 that mandated “both explicit visible labels and implicit digital watermarks/metadata on all synthetic text, audio, images, and video”. The European Union has followed, with Article 50 of the EU AI Act taking effect in August 2026, with serious penalties for non-compliance. How does it work?
There is a standard for content credentials, published by the Coalition for Content Provenance and Authenticity, which has steering members including Amazon, Google, Sony, Meta and the BBC. This standard has been picked up by many AI providers, including Anthropic. Generated files such as images or videos include digitally signed provenance metadata.
For text, the solution that Anthropic has used is quite interesting. Rather than just putting an easily deleted tag at the end of the text it generates, Anthropic has turned to mathematics. A large language model generates text one token at a time. If it is generating a sentence like “the cat sat on the …”, then it assigns a probability to the most likely next word. The word “mat” might have a high probability (say, 35%) based on its training data, but other words are quite possible, such as “floor” (18% probability), “chair” (12%) or “bed” (9%). A text watermarking scheme divides the vocabulary, in a way determined by a secret key, into two groups, called “green” and “red” tokens. So, suppose that “mat” was red, “floor” was green, “chair” was green and “bed” was red. The algorithm slightly increases the probabilities of the green tokens before sampling. So maybe the word “mat” normally had a probability of 35%. After the watermarking is applied, this probability would be reduced slightly, to 28%. Similarly, “floor” (a green token) would be bumped up, say from 18% to 23%. Likewise, the word “chair”, another green token, may be increased from 12% to 15%, while the red token “bed” might be reduced from 9% to 7%. The choice of which tokens are red or green changes based on a secret key, but over hundreds of generated tokens, the output would contain more green tokens than chance would predict. The algorithm can then look at generated text and spot that the generated tokens appear according to the key, and so can detect whether they are AI-generated. It is important not to bias the probabilities too much, or the quality of the output will be affected.
A detector takes a piece of finished text, reconstructs which tokens would be “green” at each position, and counts how often the author selects them. A piece of human-written text might be expected to contain 50% green tokens, but a watermarked model might produce perhaps 60%. Over a long enough piece of text, the difference becomes statistically significant. Of course, someone could take a piece of AI-generated text and edit a few words, but light editing would not necessarily fool the watermarking algorithm. Naturally, very heavy editing could do that, so the system is not foolproof. It took just one day after the Anthropic announcement about introducing watermarks, for a “AI watermark remover” tool to appear on the internet.
The AI vendors have been nudged into providing watermarks by governments, but watermarking clearly has public support. A large Pew survey in 2025 found that 76% of US adults said it was “extremely important” or “very important” to know whether text, images or videos were made by humans or by AI, and only 12% felt that they could tell the difference themselves. A separate survey by Statista in 2025 found that around 70% of respondents “struggle to trust online information because they cannot tell whether content is real or AI-generated.”
Surveys of public opinion on AI have found a hardening of attitudes towards AI. Watermarking may be one small step in rebuilding some of that trust.
