Does the "Created with AI" label still make sense? More and more companies are introducing it, even though it does not limit the spread of deepfakes
Recently, Anthropic announced that it will introduce a system to apply a kind of label — machine-readable by computer systems — that makes content and files processed by its chatbot Claude identifiable, as well as those produced by Claude Code, its tool that is very popular among web developers.
Anthropic explained that for images, the watermark will be applied within the respective metadata — that is, the information associated with the file that describes its characteristics — where it will be indicated that generative artificial intelligence was used.
The measure follows an increasingly popular approach among companies in the sector, which are experimenting with new solutions to make these watermarks more effective. Such tools, however, do not prevent the spread of deepfake images and videos used for propaganda purposes: this is also why some experts are questioning the real usefulness of labelling content generated by artificial intelligence.
How Anthropic's watermark will work
@willfrancis24 Claude will soon be watermarking all texts that it generates. Anthropic's popular AI will use a clever watermarking system which uses its choice of words to create patterns that are recognisable by an AI detector. This means that the usual methods of evading AI detection, such as prompts to humanise text or even some editing of AI-generated text, won't remove these signs of AI writing. Here's everything we know so far about when Claude models such as Fable, Opus, and Sonnet will start to use these patterns in its output. #claude #anthropic #llm #ai #aitools #artificialintelligence #tech #technology #learnontiktok original sound - Will Francis - AI + Marketing
Dario Amodei's company has said that the labelling system in question will be imperceptible to users and will not affect the rendering of content in any way — whether visual or otherwise. Furthermore, the watermark will remain present even when individual files are subsequently modified.
The fact remains that Anthropic's label is not able to indicate at what point Claude was used or with what intent. The company itself explained in a statement that the «watermark can only determine that Claude was likely involved in creating the content at some point. It cannot distinguish between "Claude wrote this" and "Claude heavily edited this other thing"». All of these characteristics give the more critical observers reason to be sceptical.
Where this labelling system comes from
A detection system similar to Anthropic's has also been introduced by the Substack platform, in collaboration with Pangram (a well-known tool designed to identify texts generated by chatbots): in this case, however, the method is based on the analysis of linguistic patterns, syntactic structures, and writing styles typical of texts produced by artificial intelligence.
Other major technology companies, including Microsoft and OpenAI, have also put their own labelling systems in place. The measure has been made mandatory by the AI Act, the European law regulating the development and use of artificial intelligence systems, approved in 2024 and entering into force on 2 August — though individual companies have a 4-month period to comply with the various provisions.
Despite the AI Act's rules applying only to the European market, Anthropic has chosen to extend its system for marking Claude's outputs globally — in keeping with the image of a particularly cautious company it has built for itself within the industry.