The Stratechery analyst called the EU's transparency rules 'overburdensome' and argued watermarking punishes anyone who uses an LLM for proofreading or translation.
Tech analyst Ben Thompson tore into the European Union AI Act's transparency requirements, calling the logic behind mandatory AI watermarking "clearly absurd."
Anthropic this week said it would add watermarks to text and files generated by Claude models launched after August 2. The AI company cited the EU AI Act, which requires that "outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated." Anthropic, OpenAI, Google, Meta, and Microsoft have all signed the relevant section of the regulation. SpaceX has not.
In his widely read "Stratechery" newsletter, Thompson said the transparency requirements are both overburdensome and unlikely to be strong enough. He took particular issue with Anthropic adding a watermark even when Claude is used only for proofreading. The EU regulation allows an editing exception, but Thompson argued there was no practical way for Anthropic to distinguish such usage.
"I am sympathetic to the impetus behind watermarking: wouldn't it be better to know what is fake and what isn't?" Thompson wrote. "What, though, is 'fake'? If an LLM states a true fact, and a human a fable, does it matter that the latter doesn't have a watermark?"
He added: "I'm glad I never developed the habit of copy-and-pasting proof-reading runs (I use LLMs for proof-reading, but manually make every individual change in my text editor), but it hardly seems fair that anyone who wants to fix their grammar runs the risk of having their original content labeled as AI; the same frustration applies to translation."
OpenAI, which currently watermarks images and audio using Google DeepMind's SynthID, has pledged to expand the process to text output. The exact timeline and technical details have not been released.
The core dispute goes beyond implementation. Thompson argued the premise of AI watermarking – that the tool, not the human, deserves credit for the output – is flawed. "From this perspective," he wrote, "to insist on watermarking is no different than insisting that a ballpoint pen advertise itself as the author, a concept that is clearly absurd."
Anthropic has not said it would limit watermarking to Europe. That leaves open the possibility that users everywhere could see Claude output marked as AI-generated. The same question applies to OpenAI, Google, and Meta, which have all signed the transparency section of the EU regulation but have not detailed global enforcement plans.
Regulators are turning to watermarks, traditionally used on images and video, as a tool to help society distinguish AI-generated text from human writing. Unlike visual watermarks, the provenance markers for AI text are not perceptible to the naked eye. Anthropic and others have promised to release detection tools that function as AI checkers.
Thompson's critique cuts to a deeper question the regulation does not answer. If a human uses an LLM to fix grammar or translate a paragraph, whose work is it? The EU rules treat any machine involvement as a trigger for labeling. Thompson calls that a category error.
Drafted by a large language model from the source reporting linked above, then screened by automated publishing checks. It is not read by a journalist before publication. Some articles cite our Alpha Score. Verify prices and figures against the original source. Educational coverage, not personalized advice.