Anthropic’s new watermarking feature for AI-generated text has sparked a debate among users of its Claude chatbot. While some see it as a move towards transparency, others question if European regulations truly necessitate such measures.
According to a Business Insider analysis, Anthropic’s implementation might exceed the requirements set by the European Union. Specifically, the Article 50(2) of the EU dictates that AI-generated content need not be marked if the technology is used for minor editing or doesn’t significantly alter the user’s original intent.
Despite this, Anthropic’s watermark, which stays intact even when text is copied or translated, is being applied broadly. Caroline De Cock of the Lisbon Council remarked, “Basically, Article 50 called for a scalpel to catch deepfakes and mass synthetic generation, but Anthropic is using a sledgehammer that hits simple proofreading and translation.”
Anthropic asserts that the watermark is a step towards compliance with the EU’s AI Act and is being implemented globally due to technical constraints.
While the transparency of AI watermarks has been appreciated by some, the move has also led to backlash. Some users have canceled their subscriptions, and there are new tools designed to remove these watermarks.
Concerns arise that the watermark might be mistakenly seen as an indication that Claude generated the text, even if it was only used for editing. This is because the watermark influences Claude’s word choices to create identifiable patterns, suggesting its involvement in the text creation.
In cases of minimal editing, the watermark is less likely to appear, according to an Anthropic spokesperson, as there’s less for it to attach to.
Anthropic’s approach seems to align with its “safety-first AI developer” brand identity, possibly aiming to reduce regulatory risks by placing the onus of label interpretation on users and external platforms, as noted by Caroline De Cock.
A Wave of AI Watermarks
The EU’s regulations mandate that new AI models launched in Europe from August 2 onward must include watermarks, while older models have until December to comply. This is prompting other tech giants to develop methods for identifying AI-generated content.
Google, for instance, employs its SynthID technology globally to invisibly mark AI-generated outputs. OpenAI plans to expand its labeling to text but hasn’t specified a date, while Meta is working on tools to identify AI-generated content, focusing currently on images.
As companies encourage AI use, there’s growing concern over the potential attribution of AI-generated work, which could lead to questions about authorship. Milos Rusic of Deepset warned, “These systems can be useful evidence, but I would be very cautious about turning them into an automated judgment about authorship or misconduct.”
