Anthropic's announcement that it will embed invisible watermarks into text content processed by its AI model, Claude, has sparked strong backlash, with some users threatening to cancel their subscriptions.

On the 11th local time, Anthropic announced that it would insert watermarks into certain text content generated or edited by Claude to comply with the European Union's Artificial Intelligence Act, which took effect earlier.

Anthropic explained that by embedding machine-identifiable watermarks directly into the sentences, the watermark remains even if the text is copied, pasted, or modified.

The company plans to release a separate program later to identify these watermarks, which will technically allow schools to check for assignments or companies to check for reports containing sentences written by Claude.

Anthropic added that the technology is not foolproof, stating, "The detection of a watermark does not necessarily mean the entire text was written by Claude."

This means that even if a human writes a text themselves, watermarks could still remain if they had Claude proofread, translate, or summarize it.

This has raised concerns that individuals might face disadvantages, being misclassified as having produced AI-generated content simply because they used Claude to assist with their writing.

John Crickett, a software engineering educator, pointed out, "The issue of proving how much a human contributed in the process of claiming copyright could become complicated." Meanwhile, some users took to social media platforms like X and Reddit to post that they would cancel their paid Claude subscriptions, calling the move "a conspiracy against pure users."

A developer in France even released a tool on GitHub capable of removing Claude's watermarks.

On the other hand, counter-arguments emerged stating, "The backlash against watermarks is only because people tried to deceive others with AI." Some also argued that "it could reduce problems related to AI learning from content created by other AI."

Anthropic clarified, "Watermarks may not be detected if users wrote text themselves and only partially modified it with Claude, or if users extensively revised text generated by AI."

The company also explained that watermarks will likely be introduced in other companies' AI models as well due to the European Union's Artificial Intelligence Act.

Previously, Google also developed and applied its digital watermark, SynthID, to AI-generated content.

(Reported by Kim Jin-woo | Video by Na Hong-hee | Design by Lee Jung-joo | Produced by SBS Digital News)