Every reply Claude writes has carried an invisible signature since Aug. 2, 2026 — Anthropic wove a watermark into the output of every model it released from that date on, worldwide, with no way to turn it off. It's the kind of real, checkable signal a reader might assume finally answers the question a search bar gets asked constantly: is this AI? It doesn't, not yet. Anthropic's own mark has no public reader anyone outside the company can use, Google's equivalent for Gemini is stuck in a year-plus-old journalist waitlist, and the tools most people already reach for instead — GPTZero, Turnitin, the free checkers a search turns up — don't read a watermark at all. They guess. Knowing which of those three things you're actually looking at, and what each one has and hasn't earned the right to tell you, is the actual gap this guide closes.
Two different kinds of evidence, and only one of them is checkable today
Watermarking and detecting solve different problems, and the industry has spent three years proving it can build one far more easily than the other. A watermark is planted by the model itself at the moment it writes — Claude and Gemini both work by subtly steering word choices toward a pattern only the lab's own key can read back out, the same trick Google's SynthID uses on images. That's real, mathematically grounded evidence, if you can actually get a machine to check it for you. A detector instead looks at finished text from an unknown source and estimates, statistically, how AI-shaped its phrasing is — no key, no cooperation from whichever model wrote it, just a guess. OpenAI has tried building both, and its own history is the cleanest evidence of how far apart they sit: it shipped a detector in January 2023 and killed it seven months later over accuracy, then reportedly built a watermark accurate to 99.9% of its own output — and has kept it unreleased for two years over who it might falsely implicate.
Check a piece of text in this order
- The fastest and most reliable check is also the easiest one to skip: ask directly, or ask to see the drafting process — an outline, a messy first pass, a version with mistakes still in it.
- Google Docs' Version History and Word's Track Changes both show whether a document grew in increments over time or arrived as one finished paste.
- As of August 2026, Claude and Gemini both weave an invisible pattern into the words they generate at the moment of creation — but neither lab has shipped a public tool for a reader to check it. Anthropic's own help center says it is only "working to enable" third-party detection, with no date; Google's SynthID Detector is limited to a waitlist for journalists and researchers, over a year after it opened.
- Tools like GPTZero, Turnitin, and Originality.ai don't read any lab's watermark — they estimate a probability from writing style, which is a fundamentally weaker kind of evidence. A Stanford study found these detectors misclassified non-native English writing as AI-generated far more often than native English writing.
- "GPTZero returned an 85% AI-likelihood score" is a fact. "So this was written by AI" is a guess wearing the first sentence's clothes — especially given how often that first sentence is wrong.
Which of those checks matters most depends on the situation — grading a student's essay carries different stakes than wondering about a stranger's post, and the order worth working through them changes accordingly.
Start where the real evidence is
What each system can actually tell you
Two mechanisms, answering different questions
| Lab-embedded watermark Claude, Gemini | Third-party detector GPTZero, Turnitin, Originality.ai | |
|---|---|---|
| What it actually measures | A statistical pattern the model wove into its own word choices at generation time | A guess about word predictability and sentence rhythm in text of unknown origin |
| Public checker available today | No — both labs describe detection tools as planned, not shipped, as of August 2026 | Yes — free and paid tools anyone can run right now |
| Documented error pattern | Untested at public scale; no independent accuracy figures published yet | Misclassifies non-native English writing as AI-generated far more than native writing |
| What a clean result actually proves | Nothing yet checkable — there's no public reader of the mark to return a result | That this specific tool's guess came back low — not that a human wrote it |
None of this makes the underlying question go away — it narrows what any given check is entitled to claim. A lab-embedded watermark, once a public reader for it exists, will be strong evidence for the model that planted it and silent on every other model. A clean detector score is weaker than most people treat it as, because the tools most people already reach for were never reading a watermark in the first place — they were reading style, on a false-positive rate that isn't evenly distributed: reporting on the Stanford study found real international students named and disputing flags from exactly this kind of tool, well before anyone had a watermark to check instead.
Where this check goes wrong
Five ways this check gets called settled when it isn't
None of this makes catching AI-written text effortless — it makes the claim checkable, which is the more honest goal. Our guide to checking whether an image is AI-generated walks the same discipline for pictures, using two real watermark systems already in wide use, and our guide to checking whether a phone call is AI-cloned covers the same question for a voice you can't just run through a detector mid-call; the dictionary has short entries for terms used here — watermark, hallucination, model weights — if any of it needs unpacking further. The story behind Anthropic's move, including why writers are objecting to it, is here.
- Anthropic and Google now weave real watermarks into AI-generated text — but neither offers a public checker yet.
- Third-party detectors like GPTZero and Turnitin don't read a watermark — they guess from writing style.
- Those guessing tools misclassify non-native English writing as AI-generated far more than native writing.
- OpenAI's own 2023 detector correctly caught only 26% of AI text and wrongly flagged 9% of human writing.
- Caveat: no tool here proves human authorship — a clean result only means this check found nothing.