Claude now leaves an invisible mark on AI-generated text. media update's Venelize de Lange considers the limitations of this latest function and the complicated question of AI authorship this conversation presents.
On Tuesday, 11 August, Anthropic announced that any text generated by their Claude AI will carry an "imperceptible watermark". This comes after media outlets' widely raised concerns regarding AI's pollution of digital content.
Such concerns have helped drive regulatory efforts, leading to the eventual implementation of the European Union AI Act, which requires artificial intelligence companies to introduce ways of identifying AI-generated content.
Correspondingly, Anthropic assures that this watermark already applies to any new Claude models launched on or after Sunday, 2 August, with the company considering how to implement the function to older models as well.
The watermark is directly embedded into text and not visible to readers, meaning it will move with any copy and paste text. It could even still show up after some editing, but images and other supported files will carry digitally signed provenance metadata.
While the systems are not yet available to users and third parties, Anthropic plans to eventually update accessibility. The watermark will apply worldwide across Claude's products, with users unable to opt out.
Many users are discontent by this development, as the AI watermark will be present even when Claude was just used for proofreading and editing, not crediting the ideas and writing to the users themselves.
Anthropic clarifies that this watermark doesn't prove that Claude wrote the content, only that the content may have been processed by Claude. The watermark is also not a fail safe; a piece of writing without a detectable Claude mark does not guarantee human-written content. Thorough editing, paraphrasing, translation or combining texts can make the signal undetectable, and short passages may also lack enough text for a reliable signal.
Similar technology has reportedly been available to OpenAI for models such as ChatGPT for years. Yet, the company refrained from implementing it, citing concerns about false positives and the high likelihood of users circumventing the system and turning to competitors instead.
Meanwhile, nearly 20 companies have already signed the EU's transparency code by the end of July. Although Anthropic's watermark limitations might mean that AI detection cannot automatically become a verdict on authorship, these regulatory efforts do speak to a more responsible relationship with AI in the media industry.
Did you enjoy this article? Let us know in the comments section below.
*Image courtesy of Canva
**Information sourced from Claude, Business Insider, Fortune and Independent