Anthropic to Fine-Tune Claude Amid Ongoing Scope and Criticism

Anthropic plans to add a watermark to the text outputs of several models via Claude to comply with European law. The company describes a process based on word choices, verifiable with a key. However, it has not yet clarified how it will handle certain minor modifications or adjacent uses, which has fueled criticism from professionals and the emergence of erasure tools on GitHub.

Anthropic has not defined the scope of watermarking

The company has not detailed how it will practically distinguish between "generated" text and related tasks performed with AI. It indicated that watermarks may or may not apply to minor modifications made with Claude. These gray areas raise very concrete questions, such as those posed by John McCarthy, an opinion writer at The Drum: should tasks like sorting and responding to proposals, automatic video subtitles, interview transcriptions, or even the use of a light filter in Zoom all be marked the same way as text written by a model? One position in the debate suggests that any adjacent editorial work using AI should not be equated with "generated" content.

Creators and publishers criticize the cost and provenance

Creators are questioning the provenance of the models and the resulting business model. Sara Simeone, founder of NoCodeLab.ai, claims that labs have built their wealth by scraping content without consent, then charging for regurgitated versions of that same content while demanding attribution that would trace intellectual property back to them. Critics also argue that large amounts of online content have been collected without compensating authors to train the systems. In this context, it is suggested that there is hypocrisy in marking the release of copied books that are then destroyed, with claims that a settlement was reached with authors. Beyond this, some describe a deeper reaction towards the labs: relying on tools that automate work, then placing a tracker on that work in a way perceived as risky, is felt as a betrayal; others counter that a company that has never been accountable cannot betray.

Anthropic announces marking to comply with the EU

Anthropic plans to begin watermarking the text outputs of several of its models accessible via Claude. The stated goal is to meet the transparency requirements of the European AI law, particularly Article 50. After an initial communication that was sparse on details regarding the functioning and reading of these watermarks, the company published guidelines indicating that the imprint is attached to specific word choices.

The described process: verifiable lexical choices by key

The mechanism relies on the probabilistic nature of the models that generate words one by one, based on the immediate context. Anthropic illustrates this point with synonyms like "cloudy" or "gray" in a weather description: choices deemed low-stakes that would engrave an imperceptible pattern for a reader but detectable by someone with the key. Technically, the random selection of the next word is slightly skewed: the key and a few previous words guide the drawing. The chosen words remain random; however, a sequence can be verified afterward to assess its consistency with what Claude would produce under the key, thus deducing a probability of origin.

Technologists contest the loss of control, erasure tools emerge

Several technical voices oppose the idea that these word choices would be inconsequential. John Gruber, co-creator of Markdown, argues that every term matters, that he wants the most precise formulations possible, and finds it unacceptable for a tool to sacrifice clarity or quality to insert hidden cues. While some users express discomfort with this marking, public outrage has grown, and projects aimed at removing these marks are multiplying on GitHub.