Claude Is Shipping an AI Text Detection API You Can Use Yourself

Anthropic is going to ship a text detection API that anyone can call. Thariq Shihipar, an engineer on Anthropic's Claude Code team, confirmed it in a thread posted the day after the company disclosed that new Claude models embed invisible watermarks in the text they generate.

That second announcement is the more consequential of the two. A watermark nobody outside Anthropic can read is a policy commitment. A watermark plus a public detection API is a working system, and it changes how platforms, publishers, employers, and schools are able to evaluate written work.

We wrote about the watermark itself when Anthropic announced it. This post is about the detection layer: what it will realistically be able to tell anyone, what it will not, and how we are preparing for it at WriteHuman.

All Claude generated text will have this embedded watermarking. For example, you could check if a PR was generated by Claude Code.

Thariq Shihipar (@trq212), August 11, 2026

What Anthropic Actually Confirmed

Anthropic's help center documentation, published alongside the watermark disclosure, stopped short of promising a public interface. It said the company was working to enable users and third parties to detect the marks, and that details on detection mechanisms would arrive in forthcoming technical documentation. Early news coverage repeated that framing, and several outlets noted that the detection tooling had not shipped and that Anthropic had not said who would be allowed to use it.

Shihipar's thread went further. He said Anthropic will also be shipping a "text detection API that you can use yourself," and he gave a concrete example of the kind of question it is meant to answer: checking whether a pull request was generated by Claude Code.

He also placed the work in its regulatory context, describing it as part of working with the EU AI Act and noting that other labs are adding similar watermarking. The transparency obligations in Article 50 of that law became enforceable for newly launched systems on August 2, 2026, which is the same date Anthropic uses as the cutoff for which Claude models carry marks at launch.

A Provenance Detector Is Not the Same as an AI Detector

Every AI detector on the market today, including ours, works by inference. It examines the text, measures statistical properties of the writing, and estimates the probability that a machine produced it. It has no access to the moment of generation, so it is making an informed judgment about evidence left behind in the prose.

Anthropic's detector will work on a completely different principle. It is not assessing writing style at all. It is looking for a signal that Anthropic deliberately placed into the text during generation. When that signal is present and intact, the answer approaches certainty, because the party checking the text is the same party that planted the mark.

That is a genuine improvement in one narrow direction and a genuine limitation in every other. Accuracy on marked Claude text should be excellent. Coverage of anything that is not marked Claude text is zero.

What the Claude Detection API Will and Will Not Catch

Based on what Anthropic has published so far, the detector should identify text produced by Claude models launched on or after August 2, 2026, across the surfaces the company lists: the Claude Platform API, Claude, Claude Code, Claude Cowork, Claude Tag, and Claude accessed through AWS, Google Cloud, and Microsoft Foundry. Anthropic says the mark travels with the text when it is copied and pasted, and that it may survive some editing.

What it will not catch is a longer list. It cannot see output from GPT, Gemini, Llama, Mistral, or any other model, because Anthropic can only mark its own generations. It does not yet cover older Claude models, which the company describes as still in progress. It cannot work on passages too short to carry a reliable signal. And by Anthropic's own account, it will not reliably detect text that has been substantially rewritten, paraphrased, translated, or mixed into other writing.

This is the part most people are going to get wrong, so it is worth stating plainly. A Claude detection API is not an AI detector. It is a Claude detector. Anyone who runs a document through it, sees no result, and concludes that a human wrote the document has learned almost nothing.

Anthropic's Own Limitations Still Apply

Anthropic published a limitations section with the announcement, and it is unusually candid. It makes two points that cut in opposite directions.

The first is that a detected mark signals that content was processed by Claude, but is not fully conclusive. Anthropic notes that Claude may not be the original author at all, because people routinely use it to proofread, translate, summarize, and convert files. Output can carry a Claude mark even when the underlying ideas, text, and data came from a person.

The second is the reverse. Anthropic states that the lack of a detected mark does not mean content was not AI generated. The text may have come from a model released before marking was supported, it may be too short, its metadata may have been stripped by a format conversion or a screenshot, or it may have been produced on a surface where that marking type was not supported.

Those two paragraphs are the most important thing Anthropic published, and they are the ones most likely to be skipped once there is an API returning a result. A detection tool that gets used without its caveats will do more damage than the watermark ever could.

What This Changes for People Who Write With Claude

The practical effect is that a Claude draft can now be checked by anyone holding an API key, including a platform you submit work to. Going forward, the safe assumption is that anything produced by a current Claude model carries the mark.

The decision facing writers has not changed, but the stakes are more concrete than they were a week ago. If you are comfortable publishing Claude's raw output, nothing about this announcement stops you. If you are not, treat the response as a draft and rewrite it into something that is genuinely your own work.

What matters is being honest about what rewriting means. Swapping a handful of adjectives leaves Claude's sentence architecture completely intact. Anthropic names heavy editing and paraphrasing, not light touch-ups, as the transformations after which its mark may stop being detectable. Those are very different operations on a piece of text, and only one of them produces a genuinely different version of the writing.

How WriteHuman Is Preparing for the Claude Detection API

We are not going to make claims about a tool that does not exist yet. Nobody outside Anthropic has called this API. Any product promising a specific result against it today is guessing, and we would rather say that clearly than sell certainty we do not have.

Here is what we are actually doing.

We will test it on day one. The moment the endpoint is available, we will run WriteHuman output through it and publish what we find. That is the only way anyone learns how the detector behaves on rewritten text, and we would rather run that test in public than have our users discover the answer on their own work.

We will add it to what we measure. We publish HumanizerBench, our own public monthly benchmark of humanizing tools. If Anthropic's detector becomes callable, we intend to add it as a measured target alongside the statistical detectors we already track, and report the numbers on the same cadence.

We will keep our own detector honest about its scope. Our free AI detector estimates whether text reads as machine generated across models. A Claude provenance check answers a much narrower question. Those are two different signals, and we would rather show both and explain the difference than blend them into a single number that hides which question is being answered.

What We Still Do Not Know

Several things remain genuinely open, and they will determine how much any of this matters.

We do not know when it ships. Shihipar said more details would be shared during the rollout, but no date has been given.

We do not know who gets access. This is the central unresolved question, and it cuts both ways. An openly available endpoint lets anyone verify Anthropic's claims independently, and it also lets anyone measure their way around the watermark. A gated one prevents outside verification of how well the system actually works.

We do not know what it returns. A binary answer, a confidence score, and a passage by passage breakdown are three very different products with three very different failure modes.

And we do not know how it behaves on partially rewritten text. That is the question that matters most to the people who use WriteHuman, and it can only be answered by running the test once the endpoint exists.

The Practical Takeaway

Anthropic is now building both halves of the same system: a mark inside the text, and an interface for reading it. The first half was the announcement everyone covered. The second half is the one that makes it real.

What that system will do well is tell you whether recent Claude output has passed through a piece of text in close to its original form. What it will not do is tell you whether a person wrote something, whether AI was involved at all, or what happened to a draft after it left Claude.

Our position has not changed. Use Claude for the draft if it helps you. If you would rather not publish text in the exact shape Claude produced it, rewrite it properly, review the result yourself, and publish the version you actually want your name attached to. That was the right approach before the detection API, and it will still be the right approach after it.