Your days of having Claude quietly do your homework may be numbered.
Anthropic has begun rolling out invisible watermarks in Claude-generated text, part of a broader push to make AI-generated content identifiable. And unlike the little logo stamped on an AI-generated image, you won’t actually see this one.
The technology is based on SynthID Text, developed by Google DeepMind. Instead of adding hidden characters or metadata, SynthID subtly influences which tokens an LLM chooses while generating a response. Think of it as an invisible statistical fingerprint woven into the writing itself. The words still look completely normal to us, but a detector can recognize the pattern.
Google has already been using SynthID to watermark text generated by Gemini, and Anthropic is now bringing similar technology to Claude as it works to comply with transparency requirements under the EU AI Act.
And yes, this could get interesting for education.
A student could copy an AI-generated essay into Word, change a few sentences, and submit it without any obvious indication that an LLM helped write it. A statistical watermark is designed to survive at least some copying and editing, potentially giving schools something much stronger than today's notoriously unreliable "AI detectors."
But don't panic, ChatGPT fans — OpenAI hasn't announced text watermarking for ChatGPT yet. OpenAI currently uses SynthID for supported images and audio and has experimented with text watermarking, but has publicly acknowledged problems including circumvention and false positives.
Watermarking isn't foolproof either. Significant rewriting or translation can weaken these signals. Even Google describes SynthID as a building block rather than a silver bullet.
Still, something important just changed.
The AI era may be moving from Can we detect AI writing? to Can AI writing identify itself?
And that is a much bigger deal.
Thanks for reading and if you have a question, please leave a comment.
Comments
Post a Comment