Check text for hidden characters and AI writing, and images and audio for OpenAI's own C2PA and SynthID marks — in one place.
OpenAI uses different signals for text, images and audio. Here is each one, and how this detector handles it.
| Signal | What OpenAI does | What this detector does |
|---|---|---|
| Text — textGrain | Invisible statistical watermark in word choice. Rolling out to ChatGPT and Codex for users in the EU; opt-in for API customers worldwide. | AI scoreOpenAI’s text detector is limited to approved researchers, so no public tool can read textGrain yet. tropa-3 scores the writing itself, with or without a watermark. |
| Text — hidden characters | Not an official OpenAI watermark, but zero-width characters, homoglyphs and unusual spaces show up in copied and “humanized” text. | CheckedEvery text check scans for them, tells you how many were found and removes them before scoring. |
| Images | C2PA content credentials plus an invisible SynthID watermark on supported ChatGPT and API images. | CheckedEvery image is checked with OpenAI’s own Content Provenance API. A confirmed mark is shown as “Verified by OpenAI”. |
| Audio | Invisible SynthID watermark on supported OpenAI voice output. | CheckedThe audio check asks OpenAI’s Content Provenance API whether the clip carries that watermark. |
When people say ChatGPT hides a watermark in its text, they usually mean invisible Unicode characters: zero-width spaces, look-alike letters from other alphabets, unusual spaces or full-width characters. They are not OpenAI's watermark — textGrain lives in the choice of words, not in hidden symbols — but they are real, and they matter.
They slip in when text is copied between apps, and “humanizer” tools add them on purpose to confuse AI detectors. Our text check counts each type, tells you what it found, and removes them before scoring, so the result reflects the writing rather than the tricks.
On 5 October 2026 OpenAI announced textGrain, an invisible statistical signal in the model's word choices, in response to the EU AI Act. It is coming to eligible ChatGPT and Codex output in the EU over the following weeks, and API customers can switch it on worldwide.
By OpenAI's own numbers, at a 1% false-positive rate the detector finds about 95% of watermarked 400-token passages and about 80% of 200-token passages — and replacing a quarter of the words drops that to 17%. The detector is limited to approved researchers for now.
So for most ChatGPT text you will meet in practice, a watermark check alone won't answer the question. That is why every check here also runs tropa-3, which scores the writing itself. Full details in our textGrain explainer.
Four steps, and the same detector handles text, images and audio.
Paste ChatGPT text into the detector above, or upload an image (PNG, JPG, WebP) or an audio clip of up to 60 seconds. Use the original file, not a screenshot, where you can.
For text, the result lists any zero-width characters, homoglyphs or unusual spaces it found. They are removed before the text is scored.
For images and audio, the result shows “Verified by OpenAI” when OpenAI confirms its C2PA credentials or SynthID watermark. No mark does not mean human-made.
tropa-3 gives every text a score from 0 to 100, with or without a watermark; with Pro it also highlights the sentences that read as AI-written. Treat it as evidence to review, not as proof.
A watermark is strong evidence when it is there. An AI detector covers everything the watermark never reaches. You want both.
| ChatGPT text watermark (textGrain) | AI detector (tropa-3) | |
|---|---|---|
| Which text it covers | Eligible ChatGPT and Codex output for EU users, plus API text where the developer opted in | Any text, from any model, any region, any date |
| Text written before October 2026 | Not watermarked | Covered |
| Claude, Gemini, Llama and others | Not covered (other companies use their own marks) | Covered |
| After heavy editing or translation | Signal weakens sharply (25% of words replaced → 17% detected) | Score reflects the edited text; humanized text is a known hard case |
| Who can run it today | Approved researchers only | Anyone, on the website or via API |
| What a result means | An OpenAI system generated or processed the text | How strongly the writing reads as AI-generated |
Images from ChatGPT and the OpenAI API can carry two marks. C2PA content credentials are signed metadata attached to the file: they record that an OpenAI model created the image and when. SynthID is an invisible watermark woven into the pixels themselves.
The difference matters once an image has travelled. Screenshots, most social platforms and many editing apps drop the metadata, so the C2PA credentials are gone. SynthID is designed to survive more of that, though heavy compression or tight cropping can weaken it as well.
Our image check asks OpenAI's Content Provenance API about both marks and also runs our own detector on the pixels, so an image without any mark still gets a score. See the AI image detection API if you want to run the same check in your own product.
Look at essays and theses with evidence you can talk through with the student.
Academic AI detector →Screen submissions and freelance copy before it goes out under your name.
For publishers →Review cover letters and take-home tasks without accusing anyone on a single score.
AI detection in hiring →Flag AI-generated profile photos, reviews and voice messages at scale through the API.
Deepfake detector →Looking for the tells in the writing itself? Read signs of AI writing.
Partly, from October 2026. OpenAI announced textGrain on 5 October 2026: an invisible statistical watermark in the model’s word choices. It is rolling out to eligible ChatGPT and Codex output for users in the EU, and API customers anywhere can opt in. Outside the EU it is not a default, and older ChatGPT text carries no watermark at all.
Not directly, for now. OpenAI’s textGrain detector is limited to approved researchers and expert organisations, so no public tool can verify it yet. What you can do today is score the text with an AI detector: tropa-3 reads the writing itself, so it works on ChatGPT text with or without a watermark.
Those characters are not OpenAI’s official watermark. But zero-width characters, homoglyphs and unusual spaces do turn up in text copied from chat interfaces and especially in text run through “humanizer” tools. Our text check counts them, shows you what it found and removes them before scoring.
Yes. OpenAI adds C2PA content credentials and an invisible SynthID watermark to supported images. Screenshots and re-saves often strip the C2PA metadata; the SynthID watermark is more robust, though heavy compression or cropping can weaken it too. We check every uploaded image with OpenAI’s Content Provenance API, so upload the original file where you can.
No. A missing watermark proves nothing on its own: the text may be from before the rollout, from outside the EU, from the API with watermarking off, edited, translated, or from another company’s model. That is why every check also includes an AI score from tropa-3.
No. textGrain is part of the word choices themselves, so copying, pasting or reformatting leaves it in place. Hidden Unicode characters are different: some editors strip them on paste, others keep them. Rewriting or translating the text is what weakens a statistical watermark, according to OpenAI.
No. Em dashes, “delve” and similar habits are features of a writing style, not a watermark, and plenty of people use them too. On their own they prove nothing. An AI detector looks at many such signals together instead of a single tell.
Only if the developer opts in. Since 5 October 2026 API customers can switch on textGrain for select models; it is off by default. Text from apps built on the API is therefore usually unwatermarked, which is another reason to rely on an AI score as well.
Often only partly. A screenshot or a social media upload usually removes the C2PA content credentials. The SynthID watermark is built to survive more edits, but heavy compression or cropping can weaken it. Upload the original download where you can; our image detector also scores the pixels themselves, with or without a watermark.
You can try it without an account, and verifying your email gives you 1,000 free credits. Text costs 1 credit per word and an image 300 credits. For regular checks, Pro covers unlimited website checks; developers can use the same checks through our API.
No. A watermark shows that an OpenAI system generated or processed content, not how much a person contributed or who used it. Treat any result, ours included, as evidence to review rather than as proof.
What OpenAI announced, how well it works by its own numbers, and what you can check today.
Anthropic’s SynthID-Text watermark and C2PA metadata, and which of the two you can verify.
Google’s watermark for images, audio and video — and what it covers.
Check whether text was written by ChatGPT, GPT-6 and other models.
Disclaimer: WasItAIGenerated is an independent AI detection service and is not affiliated with, endorsed by, or officially connected to OpenAI or Google. ChatGPT, Codex and textGrain are trademarks of OpenAI; SynthID is a trademark of Google DeepMind. Image and audio provenance checks use OpenAI's public Content Provenance API. References to third-party announcements are based on publicly available information as of October 2026.
Paste text, or upload an image or audio file. Hidden characters, OpenAI's marks and an AI score in one result.
Try it free1,000 free credits after sign-up — no credit card needed