What is an AI watermark?
An AI watermark is a mark that says a file or a piece of text came from a generative model. There are five kinds in use, and they fail in different ways.

| Kind | Where it lives | What breaks it | Who checks it |
|---|---|---|---|
| Visible mark | A logo or label over the picture | Cropping or editing it out | Anyone, by eye |
| Metadata (C2PA) | A signed record inside the file | A screenshot, a format conversion, a platform that strips it | A Content Credentials viewer |
| Invisible image or video mark | The pixels and frames themselves | Heavy editing (designed to survive crops, filters and compression) | The provider's detector |
| Invisible audio mark | The sound, inaudible to people | Not stated beyond noise | The provider's detector |
| Text mark | The probability of each word choice | Paraphrasing, very short text, code | The provider only, in most cases |
The first two are labels that can be lifted off. The last three are signals built into the content, which is why people call them real watermarks. Sources for the table are listed at the end of the page.
Does AI generated content have a watermark?
Some of it does, and the answer depends on the model and the format. Google marks images, video, audio and Gemini app text with SynthID. OpenAI says images made with ChatGPT or its API carry two signals, C2PA Content Credentials and SynthID watermarks. On 11 August 2026 Anthropic said its models will watermark text, and its support page says Claude models launched in the EU on or after 2 August 2026 support marking from launch, with older models being updated. IEEE Spectrum adds that Anthropic declined to give more detail.
OpenAI had not yet introduced a text watermark at the time of the IEEE Spectrum report, although it plans to. Open source models are the gap. The SynthID-Text authors write that enforcing watermarking on open source models deployed in a decentralized way is difficult. A file from one of those models can carry no mark at all.
So the absence of a watermark tells you nothing. That's the same limit that makes AI detectors a statistical guess, and it matters for deepfakes, where the person making the fake chooses the tool.
What is SynthID and how does it work?
SynthID is Google DeepMind's family of invisible watermarks for text, images, video and audio. For text, DeepMind says it "adjusts these probability scores to generate a watermark" for the words a model might pick next, and says this "doesn't affect the quality of the output". A reader sees ordinary prose. A detector that knows the scheme looks for a statistical lean in the word choices.
For images and video the mark is added at creation and is "designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression", in DeepMind's words. "Designed to" is a goal, not a guarantee. For audio, DeepMind says the mark is inaudible and covers audio from its Lyria music model and NotebookLM podcasts.
The text version was published in Nature on 23 October 2024. Google ran a live test of about 20 million watermarked and unwatermarked Gemini responses and the authors found no significant difference in thumbs up and thumbs down rates. The same paper lists limits: coordination among model providers is required, and generative watermarks are vulnerable to stealing, spoofing and scrubbing attacks. IEEE Spectrum quotes the paper as reporting that detection fell below 50 percent for short replies.
What is C2PA and what are Content Credentials?
C2PA is an open technical standard from the Coalition for Content Provenance and Authenticity, and Content Credentials is the name for its labels. The coalition describes them as working like a nutrition label for digital content, showing where it came from and how it was edited. Its steering committee lists Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok and Truepic.
The key difference from SynthID is where the record sits. C2PA data is attached to the file as metadata, so it can carry rich detail, such as the tool and the edit history. It can also be removed. OpenAI states that Content Credentials "can be lost through everyday actions, such as taking a screenshot or converting a file", and that platforms and editing tools may remove them.
OpenAI adds one useful point: when credentials remain, they "can provide more detail about a file's origin and history than a watermark alone". The two approaches cover for each other. Metadata says more when it survives, and an invisible mark may survive some changes that strip metadata.
Can AI watermarks be removed?
Yes, and how easily depends on the kind. The table compares what each source states, as of November 2026. A cell says not stated where the source did not say.
| Mark | Survives a crop | Survives a screenshot | Survives paraphrase or edit | Source |
|---|---|---|---|---|
| Visible logo | No, cropping removes it | Not stated | Not applicable | Zapier |
| C2PA metadata | Not stated | No, OpenAI says it can be lost | Not stated | OpenAI help center |
| SynthID image or video | Designed to, per Google | May be harder to detect, per OpenAI | Substantial edits can remove it, per Zapier | DeepMind, OpenAI, Zapier |
| SynthID audio | Not stated | Not stated | Survives added noise and MP3 compression, per Google | DeepMind |
| SynthID text | Not applicable | Not applicable | Paraphrase weakens it, short replies detect below 50 percent | Nature, IEEE Spectrum |
The honest summary is that no mark survives a determined attacker. The SynthID-Text paper names stealing, spoofing and scrubbing as open research problems. A spoofed mark could make human writing look machine made, which is why a hit is evidence and not proof.
Searches for an "AI watermark remover" are common. The ethical line is the one in AI ethics: removing a mark to pass machine text off as human work, or to hide a fake, is the use the labels exist to stop.
Is there an AI watermark detector?
There are provider checkers, and no universal one. Google says Gemini, Search and Chrome can check an uploaded image, video or audio file for a SynthID watermark, and that its SynthID Detector portal is opening to everyone after journalist testing. Its page doesn't say these tools check text, and says the portal also covers files made with partners such as OpenAI.
For Claude, Anthropic says anyone can check a file for a Claude Content Credential with its free Claude Content Checker. Text watermark detection is in private preview for eligible organizations such as regulators, researchers, media and educational groups, so a reader cannot check Claude text today. A clean result from the wrong checker means nothing, because each one finds only the marks it knows. For content with Content Credentials, a viewer reads the signed record instead. Tools that claim to detect "AI text" from style alone are a different thing, covered in how AI detectors work.
Is AI watermarking required by law?
In the European Union, yes, for providers. Article 50 of the EU AI Act says providers of AI systems that generate synthetic audio, image, video or text must ensure the outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated". The technical solution has to be effective, interoperable, robust and reliable where technically feasible. Systems that only assist standard editing are exempt. Paul Weiss reports that most Article 50 duties applied from 2 August 2026, and that the marking duty for generative systems already on the market was deferred to 2 December 2026 by the Digital Omnibus. The Omnibus recital 38 describes a transitional period of four months for providers who placed systems on the market before 2 August 2026. Check the current timetable before you rely on it.
A second duty falls on deployers. Anyone who uses a system to generate or manipulate content that is a deepfake must disclose it. For evidently artistic, satirical or fictional work the duty is limited to disclosing it in a way that does not spoil the work. The US has no equivalent federal rule that this page could source, so the labels there are voluntary or set by platforms and individual states. TechCrunch reports that Anthropic tied its text watermark to the EU rules.
Does Claude watermark its text?
Yes, on supported models. Anthropic's support page says supported Claude models embed an imperceptible watermark in generated text that does not change its meaning or readability, and that the mark stays with the text when copied and may survive some editing. Supported file types such as PNG and JPEG get a C2PA Content Credential.
The page lists its own limits. A missing mark is not proof the content was not AI generated, since the text may be short or heavily edited, the file metadata may be stripped, or a non default sampling setting may have been used. A detected mark is a signal, not proof: Claude may only have edited or summarized content from elsewhere.
AI watermark or AI detector: what is the difference?
A watermark detector looks for a signal the generator wrote on purpose, so it works only for that provider. An AI text detector guesses from style and has no signal to find, which is why it produces false positives. Hyperclapper makes the same split, and the figure it gives for detector accuracy has no source, so this page does not repeat it. See how AI detectors work.
Three real cases of watermarks at work
- Gemini, 2024. Google routed live Gemini queries to watermarked and unwatermarked models and compared about 20 million responses. Users didn't rate the watermarked text worse. This is the strongest public evidence that a text mark can be quiet.
- Claude, 11 August 2026. IEEE Spectrum reports that Anthropic's text watermark is based on Google's SynthID-Text, and quotes critics, among them Meta scientist Vinu Sankar Sadasivan, who dispute the claim of no quality change, while John Kirchenbauer defends the approach. It is a report of a statement, not a test.
- A ChatGPT image, today. OpenAI says its images carry both Content Credentials and SynthID. If you screenshot one and post it, the metadata may be gone, and the invisible mark, if intact, is what is left. OpenAI says the watermark may survive some changes that remove metadata.
What can you ask an AI about watermarks?
A chatbot can't read a watermark from a picture, so use it to plan the check, not to run it. These five prompts work in Claude, ChatGPT, Gemini or Whizi, with web search on.
- "I have an image I suspect is AI made. List the checks I can run in order, from visible marks to Content Credentials to the provider's checker, and what each one can and cannot prove."
- "Summarize the current state of SynthID, C2PA and Content Credentials using only the vendors' own pages. Give the link and the date you read each one."
- "I run a small site that publishes AI generated images. What does EU AI Act Article 50 ask of me as a provider or a deployer? Quote the clause and tell me which parts need a lawyer."
- "Compare watermarking and AI detectors for catching AI essays. Make a table with how each works, one failure mode and one source."
- "Here is my caption for an AI generated image: [paste]. Rewrite it so a reader knows it is AI generated, in under 20 words."
Where Whizi fits
Whizi can't read or verify a watermark. It helps with the reading around the check. You can put one question, such as prompt 2 or prompt 3, to Claude, GPT and Gemini in one app and see where they disagree. Web search is on by default in a chat, so the answer can cite current pages.
The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Side by side Compare of two models on one prompt is a Pro feature, and Pro's first week is $1.99. See Whizi pricing for the plans.
What do AI watermarks not tell you?
A watermark says a provider marked the content. It doesn't say the content is false, harmful or unedited. Zapier makes the point that a text watermark shows AI involvement and not the level of involvement, so a human could have rewritten most of it.
The reverse also holds. A missing mark doesn't clear a file, because the mark depends on the generator cooperating. Verify a claim the way you would any other, with a fact check against sources you can open. For where this fits in the wider technology, see what is generative AI.
- Keep the original file and avoid screenshots before you check it
- Look for Content Credentials, then run the provider's own checker
- Treat a found mark as evidence and a missing mark as unknown
- Never use a watermark result alone to accuse someone
- Confirm the date of any legal rule, since timetables move
Frequently asked questions
How do I check if an image has an AI watermark?
Upload the original file to a provider checker, such as Gemini or the SynthID Detector for Google marks, and open its Content Credentials in a viewer. Google says its detector finds only content made with Google or partner SynthID marks, so a miss is not proof, as of November 2026.
Does ChatGPT watermark its text?
Not as of the IEEE Spectrum report on Anthropic in August 2026, which says OpenAI has not yet introduced a text watermark but plans to. OpenAI does say its ChatGPT and API images carry Content Credentials and SynthID. Check its help center for changes.
What is the difference between SynthID and C2PA?
SynthID is an invisible mark built into the content, and C2PA is signed metadata attached to the file. C2PA can carry the tool and edit history but can be lost to a screenshot, while SynthID is designed to survive many edits. OpenAI uses both on its images.
Is an AI watermark proof that something is fake?
No. It shows a cooperating provider marked the output, not that the content is false, and a spoofed mark is possible, since the SynthID-Text paper lists spoofing as an open problem. Use it as one clue next to sources and context.