Anthropic said this week it will start watermarking content generated by Claude, joining a small but growing list of AI companies trying to make it easier to identify machine-generated text and images. The company shared technical details on how the system works, but left some key questions unanswered.

The watermarking applies to both text and images created by Claude. For text, Anthropic uses a statistical approach that subtly biases word choice in ways humans can’t detect but that show up in analysis. For images, the company embeds invisible markers that survive common editing operations.

That last part matters. Early watermarking schemes failed because you could strip them out by cropping an image or running it through basic filters. Anthropic says its system can handle JPEG compression, resizing, and “light editing,” though the company didn’t define what counts as light versus heavy.

What happens to code

The most interesting wrinkle is how watermarking affects code generation, which is one of Claude’s main use cases. Code has less flexibility than prose. You can’t bias word choice when there’s only one correct function name or when syntax requires specific keywords.

Anthropic says it will apply watermarking to code “where possible” but acknowledges the constraints. That probably means comments and variable names get watermarked, but core logic doesn’t. It’s unclear whether that leaves enough signal for reliable detection, especially in short snippets.

The company didn’t say whether developers can opt out of watermarking for code. That could matter if the watermarking introduces even small biases toward less efficient patterns or non-idiomatic naming.

Detection without a registry

Anthropic will offer a free detection API that anyone can use to check whether text or images were watermarked by Claude. That’s different from the registry approach some researchers have proposed, where all AI companies would report to a central database.

The detection API only works for Claude content. It won’t flag text from ChatGPT, Gemini, or other models. That limits usefulness for platforms trying to enforce AI disclosure rules, since they’d need to check against multiple APIs and still wouldn’t catch everything.

The company claims low false positive rates but didn’t publish numbers. False positives are the real test. If the system flags human writing as AI-generated even 1% of the time, that’s a problem for any high-stakes use case.

Why now

Anthropic isn’t the first to try this. Google announced watermarking for its models earlier this year, and OpenAI has discussed it without committing to a timeline. But deployment has been slow, partly because watermarking has real tradeoffs.

Adding watermarks uses compute, which means higher costs and slower generation. It also constrains output in ways that could make the model less useful. Anthropic apparently decided those costs are worth it, likely because regulatory pressure is building.

California is considering legislation that would require watermarking for AI-generated content. The EU’s AI Act includes transparency requirements that watermarking could help satisfy. Companies that implement watermarking now get ahead of mandates and can shape how the requirements work.

What it doesn’t solve

Watermarking helps with attribution but doesn’t address the underlying problems people care about. It won’t stop deepfakes. Anyone trying to spread misinformation will just strip the watermarks or use a model that doesn’t apply them.

It also won’t settle copyright disputes. Knowing that an image was AI-generated doesn’t tell you whether the training data was properly licensed or whether the output infringes on someone’s work.

The most realistic use case is platform moderation. A site that bans AI-generated content could use watermark detection as a first-pass filter, though it would need human review for anything important since false positives and stripped watermarks will happen.

Anthropic says the watermarking will roll out “in the coming months” but didn’t give a specific date. Developers using the Claude API will be able to see whether watermarking is enabled for their requests, which suggests there might be an option to disable it for certain use cases.

The company published a technical paper with more details, but it’s light on the adversarial testing. We don’t know how the watermarks hold up against someone actively trying to remove them, which is the scenario that matters most. Until independent researchers can test that, the real-world usefulness remains uncertain.