Anthropic’s Text Watermarking Proves AI Companies Do Not Care at All About Writing

Key Points:

• Anthropic will watermark future versions of its AI, Claude, by subtly altering word choices in generated text rather than adding visible or hidden markers, allowing detection through a proprietary algorithm without affecting readability.
• The watermarking method changes the randomness source used to select among synonyms deemed interchangeable by the AI, which critics argue overlooks the nuanced meaning and intentionality behind human word choice.
• Anthropic claims watermarking does not impact text quality or creativity, citing internal tests and a Google study where user feedback showed negligible differences between watermarked and unwatermarked outputs.
• Critics and writing experts contend that treating synonyms as fungible ignores the craft of writing and devalues language, highlighting a fundamental disconnect between AI probabilistic text generation and human authorship.
• The controversy underscores broader concerns about AI-generated content, including ethical issues around training data and the limitations of evaluating AI writing quality through simplistic metrics like user thumbs-up/down feedback.