AI Branding: Ben Thompson Critiques Anthropic's Plan

The European regulation requiring machine-readable labeling of AI outputs has been in effect since August 2 and has been signed by Anthropic, OpenAI, Google, Meta, and Microsoft, but not by SpaceX. Anthropic has not indicated that it is limiting this option to the European market, and the extension by other players to all users remains uncertain. In this context, the company announced the labeling of Claude's content. Analyst Ben Thompson opposes this and highlights side effects, particularly for editing and translation.

The EU has mandated machine-readable labeling since August 2

European Union regulations require that outputs from AI systems be labeled in a machine-readable format and detectable as artificially generated or manipulated. The transparency provision came into effect on August 2. Anthropic, OpenAI, Google, Meta, and Microsoft have signed the relevant section, unlike SpaceX. In the same movement, regulators are turning to watermarks to help the public identify content generated by models. Anthropic has not specified that this labeling is limited to the European scope, and it is not established whether other companies will extend it to all users.

For Ben Thompson, the requirements are excessive and ineffective

Ben Thompson rejects the principle of labeling and considers Anthropic's project to be both overly ambitious and not sufficiently developed. He expresses reservations about the European regulation, arguing that the transparency obligations imposed are too strict and may lack real effectiveness. He also questions the definition of "false," comparing a LLM that states a true fact to a person who invents a story. He particularly targets the addition of a watermark when Claude is used solely for proofreading. The EU regulation provides an exception for editing, but he argues that Anthropic cannot technically distinguish such usage. Thompson specifies that he uses LLMs for proofreading but applies each correction manually, deeming it unfair that grammatical corrections could lead to labeling an original text as AI-generated, a frustration he extends to translation. More broadly, he believes that labeling unduly attributes ideation to AI when models remain, in his view, tools wielded by humans, going so far as to compare the labeling requirement to asking a ballpoint pen to declare itself an author.

What Anthropic and OpenAI plan regarding labeling

Anthropic has announced that it will identify texts produced by Claude. Versions of Claude released to the market from August 2 would include this feature from their launch, and the company indicates that it is working to implement it on earlier models as well. It cited European regulation as justification for this measure. OpenAI is already labeling images and audio through Google DeepMind's SynthID and has promised to extend the process to text, although it has not published the details. For text, provenance markers are not visible to the naked eye; Anthropic and other companies have promised to release tools that act as AI verifiers. The labeling project has, for its part, deeply divided the tech community.