Anthropic shares more details about how Claude's new watermarks will work
TechCrunch
β’Sat, 15 Aug 2026 18:58:39 +0000
π° What Happened
Anthropic published a blog post to explain how it will watermark text produced by its chatbot Claude. The move is meant to comply with the EU AI Act's Transparency Code. That code requires AI companies to make AI-generated content identifiable. Anthropic said watermarking does not impact the quality of Claude's output. It plans to use the SynthID-Text approach that Google DeepMind outlined in 2024. A watermark detection API will also be released.
π The Backstory
The company revealed earlier this week that it would add watermarking to Claude. That sparked debate among users. On Reddit, one poster called it a conspiracy against innocent users. Another said the only reason to avoid it is to lie to people. Business Insider reported that dozens of users on X said they cancelled their Claude subscriptions. Anthropic explained that for low-stakes choices in its responses, Claude can create a pattern that is undetectable to readers but detectable by a key holder.
π― Why It Matters
Watermarking is a key step in telling AI-generated text apart from human writing. It shows how companies are adapting to new laws and could change how people use chatbots.
Anthropic published a blog post to explain how it will watermark text produced by its chatbot Claude. The move is meant to comply with the EU AI Act's Transparency Code. That code requires AI companies to make AI-generated content identifiable. Anthropic said watermarking does not impact the quality of Claude's output. It plans to use the SynthID-Text approach that Google DeepMind outlined in 2024. A watermark detection API will also be released.
The company revealed earlier this week that it would add watermarking to Claude. That sparked debate among users. On Reddit, one poster called it a conspiracy against innocent users. Another said the only reason to avoid it is to lie to people. Business Insider reported that dozens of users on X said they cancelled their Claude subscriptions. Anthropic explained that for low-stakes choices in its responses, Claude can create a pattern that is undetectable to readers but detectable by a key holder.
Watermarking is a key step in telling AI-generated text apart from human writing. It shows how companies are adapting to new laws and could change how people use chatbots.