As of August 2026, some of the text and images produced by the world’s most popular AI tools carry a hidden mark identifying them as machine-made — and most people using those tools have no idea. Anthropic just switched on invisible watermarking for everything Claude writes; Google has been quietly marking Gemini’s output for a while; and a new EU law is forcing the whole industry to move. If you write for a living, grade papers, run a business, or just want to understand what’s happening to the text on the internet, AI watermarking is about to matter to you. Here’s what’s real, what’s hype, and the one distinction that trips up almost everyone.
What “Marking AI Content” Actually Means
There are two completely different mechanisms hiding under the umbrella term, and keeping them straight is the key to the whole topic:
- Invisible text watermarks. A statistical signal woven into the words themselves — undetectable to a human reader, not shown on screen, and (the makers say) with no effect on meaning or quality. It rides along inside the text when you copy and paste it.
- File provenance metadata. For generated images, audio, and some documents, a cryptographically signed tag attached to the file’s metadata — built on the C2PA (Coalition for Content Provenance and Authenticity) standard — stating what tool made or edited it. This is the same “Content Credentials” system camera makers and Adobe have been pushing.
Text watermarking is the genuinely new and interesting one, because text is far harder to mark than a picture. So let’s look at how it actually works.
How Invisible Text Watermarking Works
When an AI writes, it constantly picks the next word from a set of plausible options. After “the weather today was cold and…”, the model might choose “grey,” “overcast,” “damp,” or “wet” — several are equally good, and the pick is normally driven by a random seed. Watermarking hijacks that randomness: instead of a neutral coin flip, the choice is nudged by a secret key combined with the preceding words, so across a long passage the word choices form a faint, deliberate statistical pattern. Read one sentence and you’d never notice. Run the whole text through the matching key and the pattern lights up.
The dominant technique is SynthID-Text, published by Google DeepMind in Nature in 2024 (building on a 2022 proposal from computer scientist Scott Aaronson). Two things make it practical: DeepMind’s testing found no measurable drop in quality — human raters comparing watermarked and un-watermarked answers side by side couldn’t tell the difference — and it’s designed to back off in places where there’s no free choice, like exact factual statements, numbers, or code, where forcing a different word would break the output.
AI Watermarking in 2026: Who Marks What
Here’s where the three biggest AI providers actually stand as of August 2026:
| Text watermark? | Image/file provenance? | Public way to check? | |
|---|---|---|---|
| Claude (Anthropic) | Yes — from Aug 2, 2026, worldwide | Yes — C2PA on .png/.jpg/.svg | Detection API “coming soon” |
| Gemini (Google) | Yes — SynthID-Text | Yes — SynthID across formats | No public text-detection tool |
| ChatGPT (OpenAI) | No text watermark yet | Yes — C2PA / Content Credentials + SynthID for images & audio | N/A for text |
The pattern: image and audio watermarking is now widespread; text watermarking is not. Claude and Gemini are the two clear examples of major models marking their text. OpenAI joined the C2PA steering committee in May 2026 and marks images and audio, but has still not shipped a text watermark for ChatGPT — a reluctance widely attributed to competitive and user-retention concerns (a watermark only your product carries is a reason for users to pick a rival that doesn’t).
Why Now? The EU AI Act
None of this is happening out of pure goodwill. The driver is Article 50 of the EU AI Act and its Code of Practice on transparency of AI-generated content, which set an August 2, 2026 deadline for providers to make synthetic content machine-detectable. Because building one system for Europe and another for everywhere else is impractical, providers like Anthropic are applying the marking globally — which is why a law passed in Brussels now affects the text an AI writes for someone in Ohio. It’s the same “Brussels effect” that made cookie banners and USB-C charging worldwide defaults.
The Distinction Almost Everyone Gets Wrong
This is the most important section, so read it twice. There are two totally different things people call “AI detection,” and confusing them causes real harm:
- Watermark detection reads the deliberate signal the AI maker embedded, using that maker’s secret key. When it says “this is AI,” it is essentially certain, because it’s reading a mark that was intentionally put there. The catch: it only works if (a) the text was made by a model that watermarks, and (b) you have access to that provider’s detector. Today that means Claude (detector still rolling out) and Gemini (no public tool at all). It is not what your school or your editor is using.
- AI “detectors” — the Turnitin/GPTZero-style tools — do something completely different: they guess, using statistical hints like how predictable or uniform the writing is. They read no watermark. And they are alarmingly unreliable.
How unreliable? 2026 testing puts most detectors between 65% and 90% accuracy — with none clearing 80% across all categories — and false-positive rates high enough to be dangerous. They perform worst on exactly the writing that should score best: clear, well-structured, formal prose gets flagged as AI because logical paragraphs and consistent terminology look “machine-like” to a detector. Accuracy on mixed human-and-AI text collapses toward zero, and scores on scientific writing run 28–38 points lower than on humanities essays. The human cost is real: students falsely accused of cheating, freelancers losing clients over a bad score, an atmosphere of default suspicion — all from a tool that is, in high-stakes use, closer to a coin flip than a lie detector.
The takeaway: the reliable method (watermark detection) is provider-locked and barely available; the available method (AI detectors) is unreliable and should never be the sole basis for an accusation. Anyone treating a detector score as proof — in a classroom, a newsroom, or an HR office — is making a serious mistake.
What Can (and Can’t) Remove a Watermark
Text watermarks are robust, but not indestructible — by design:
- Survives: copy-paste, light edits, fixing a few words, formatting changes.
- Degrades or disappears: heavy paraphrasing, translation to another language, or a full rewrite where essentially every word is replaced.
- Never present: short snippets, and factual/numeric/code-heavy passages where the model had no interchangeable-word choices to encode into.
That cuts both ways. It means a genuine human edit of an AI draft can legitimately shed the mark — but it also means the watermark was never going to be a bulletproof cheating-catcher, and shouldn’t be sold as one.
What It Means for You
- Students & educators: the watermark could eventually give schools a genuinely reliable check — but only if the AI makers expose detection, and only for models that watermark. Right now, the tools schools actually run are the unreliable guessers. If you teach, do not hang an integrity case on a detector score. If you study, know that formal, well-organized writing is the most likely to be falsely flagged.
- Writers, marketers, freelancers: assume AI-assisted text you deliver may carry a detectable mark, and let that inform your disclosure. Most platforms and clients care far more about honesty and quality than about whether a tool was involved — get ahead of it rather than hoping a detector never fires.
- Businesses: if you operate in the EU, Article 50 transparency obligations are now live. C2PA provenance on generated images is becoming a genuine compliance and trust asset, not just a nicety.
- Developers: good news — generated code is largely un-watermarked, because exact syntax leaves no room for the interchangeable-word trick the watermark depends on.
The Bottom Line
AI text watermarking went from research paper to shipping reality in 2026, pushed by European law and led by Anthropic and Google. It’s real, it’s invisible, it doesn’t hurt quality, and it’s genuinely useful for provenance. But it is also fragmented (only some providers, only some content), provider-locked (you need their key to check), and defeatable by a thorough rewrite. The single most important thing to internalize: a watermark is near-certain proof, but the “AI detectors” most people actually use are not watermark readers — they’re unreliable guessers, and treating their output as proof is how innocent people get burned. As marking spreads and detection tools mature, that gap should narrow. Until it does, treat “this was written by AI” claims with exactly as much skepticism as the method behind them deserves.
Full disclosure, since it would be strange not to on this topic: AI tools help produce a great deal of what you read online in 2026 — quite possibly including articles you’ve read today without a second thought. That’s rather the point of all this — and exactly why knowing how the marking works, and how badly the detection can mislead, is worth your five minutes. For more on the AI hardware side of this shift, see our guide to the best AI glasses of 2026, and if you’re locking down your own digital footprint, our 1Password vs. Bitwarden comparison covers the security side.