Anthropic has started rolling out invisible watermarks on text generated by its Claude models, joining Google and OpenAI in deploying content provenance technology ahead of the EU AI Act’s transparency requirements. The watermarking, which uses Google DeepMind’s SynthID-Text approach, is designed to travel with text when it is copied and pasted, surviving basic edits.
The watermarking applies to all Claude models launched after August 2, 2026, across the full product line: the Claude API, Claude, Claude Code, Claude Cowork, and Claude Tag. Older models are being updated over the coming months. While the change is driven by the EU AI Act’s Code of Practice, Anthropic is applying it worldwide from the start.
How the watermark works
Large language models generate text by predicting one token at a time, choosing from a probability distribution of likely next words. Traditional watermarking approaches add hidden characters or tweak vocabulary. SynthID-Text works differently — it modifies the source of randomness the model uses when making low-stakes choices between equally reasonable alternatives.
When the model has multiple plausible next tokens, a secret key combined with the preceding text influences which one gets selected. Over a long enough output, this leaves a detectable statistical pattern that human readers cannot perceive. Anyone with the decoding key can check whether a piece of text was generated by Claude.
Anthropic says the approach has no practical impact on output quality, generation speed, or cost. No additional tokens are needed and the watermark is invisible to the end user.
Where it stops short
The watermarking is not applied uniformly. For purely factual statements where only one answer is correct, such as “2 + 2 = 4”, the system backs off to avoid compromising accuracy. Code gets reduced watermarking because exact syntax matters for execution — though comments and less constrained sections may still carry the signal. Short snippets are harder to detect reliably since the statistical pattern needs enough text to become statistically significant.
Light editing and proofreading will not remove the watermark, but a full rewrite where every word is replaced probably will. The system can confirm Claude was involved in generating text, but it cannot distinguish between “Claude wrote this” and “Claude heavily edited this.”
For images, Anthropic is using C2PA provenance metadata rather than embedded watermarks. PNG, JPG, and SVG files generated by Claude carry cryptographically signed metadata that traces back to the model.
A detection API is coming
Anthropic plans to offer a detection API that lets users estimate the likelihood that a given piece of text was generated by Claude. This gives organizations, content platforms, and regulators a tool for auditing AI-generated content without needing access to the model itself.
The move puts Anthropic alongside other major labs working on content provenance. Google already uses SynthID across its generative AI products, and OpenAI has deployed watermarking in DALL-E images. What makes Anthropic’s implementation notable is the scope — it covers the full product line from day one and applies globally, not just in the EU.
What this means for developers
If you build on Claude’s API, watermarking happens automatically and does not change how you integrate. No new headers, no opt-in, no performance regression. The detection API will be relevant if you run a platform that needs to identify AI-generated text at scale, such as content moderation pipelines or academic submission checks.
The practical limitation to keep in mind is that the watermark is probabilistic, not deterministic. Short outputs, highly constrained responses, and heavily edited text may not carry a detectable signal. For critical provenance use cases, this should be one tool in a broader toolkit rather than a silver bullet.
How it compares to what others are doing
Google was first to deploy SynthID publicly, starting with image watermarking in 2023 and extending to text in 2024 across Gemini. OpenAI has taken a more cautious approach, publicly discussing text watermarking research since early 2023 while holding back on full deployment over concerns about user adoption and circumvention. Meta embeds watermarks in images generated by its Imagine tool and has experimented with text-level detection classifiers.
Anthropic’s approach sits somewhere in the middle. It uses the same fundamental technology as Google (SynthID-Text) but applies it more broadly — covering the entire product line globally rather than limiting it to specific surfaces or regions. The decision to offer a separate detection API also goes beyond what Google and OpenAI currently provide for text, giving third parties a way to verify provenance without access to the generating model.
None of these systems are foolproof. A motivated adversary with enough context and compute can strip or spoof most known watermarking schemes. The goal across the industry is less about absolute prevention and more about raising the cost of passing AI-generated content off as human-written — making it harder at scale even if determined bad actors can work around it.