Skip to main content
AI Watermark Removal

Detector

AI Watermark Detector

Point OpenAI's Verify tool at a ChatGPT image and it checks for OpenAI's own signals. Point it at a Midjourney image and it has nothing useful to say, not because the image is human-made, but because it was never built to look for anyone else's watermark. That's the logic behind every image, video, and audio watermark detector on the market: one provider's signal, one configuration, one modality. Google's own Gemini checker, per its help page as re-read on 2026-09-25, checks SynthID and Content Credentials on images, video and audio, and nothing on text. Text watermark detection runs on a different, score-based mechanism with its own dedicated page. Everywhere else, a clean result answers one narrow question, and it is never "was this made by a human."

By ignacio-sjPublished Revised Confirmed

Correction,

Two claims were wrong. The page said OpenAI's Verify tool checks Sora videos and that the November 2025 Sora investigation used it; OpenAI's Verify tool accepts images and audio only, and this site's records agree only that one of the investigation's two tools was the open-source c2pa-rs CLI. It also said OpenAI's audio watermark covers ChatGPT Voice and GPT-Live; OpenAI's own wording is ChatGPT and the OpenAI API, and it names neither product. Both are corrected.

All corrections

Key takeaways

  • Watermark detectors check one provider's known signal, not "AI-ness" in general. OpenAI's Verify tool checks its own C2PA metadata and SynthID watermark; Google's in-chat checker reads SynthID and Content Credentials on images, video and audio, never text.
  • Detectors aren't even reliably reachable. Google's standalone SynthID Detector portal launched 2025-05-20 with image detection working and video and text still "coming soon," and remained waitlist-gated to journalists and researchers as of the most recent independent check.
  • A detector needs to know which scheme, and whose configuration, before a result means anything. Point it at the wrong one and you get a false negative, not proof nothing is there.
  • Ordinary handling breaks genuine signals too. One OpenAI community thread documented C2PA metadata stripped to bare PNG chunks the moment an image was re-uploaded into a ChatGPT conversation, even though the original file still verified cleanly.
  • A clean detector result never proves content is human-made. It means one specific signal, checked by one specific tool, was not found.

Figure 1

Which AI watermark detectors the public can actually use

The column that decides everything is the third one. A watermark nobody outside the provider can check is documented rather than verifiable, and most text watermarking sits in exactly that position.

Provider or toolModalityPublic accessA positive result meansA negative result means
C2PA verifiersContent Credentials inspector, c2patoolFilesOpenA signed manifest is present and validatesNo manifest survived, which is the normal case
Google SynthID DetectorPortalImage, audio, video (no text)Access-gatedA SynthID signal was foundNo signal found, or the content is out of scope
OpenAI Verify toolAlso the Content Provenance APIImage, audio (no OpenAI checker for video)OpenImages: a C2PA and SynthID result. Audio: a SynthID resultNo signal survived, it was not OpenAI output, or it came from gpt-image-1 or gpt-image-1.5 through the API, which OpenAI's coverage excludes
OpenAI textGrainAdded 2026-10-06TextApproved researchers onlyAn OpenAI model likely generated or processed the textNot proof of human authorship: too short, edited, translated, or never marked
Anthropic ClaudeTextEligible organizations only, private preview since 2026-09-01Not obtainable by the publicNot obtainable by the public
xAI GrokImage, video (visible mark); text undocumentedNone publishedNo detector exists; the visible Grok watermark on images and videos is read by eyeNot obtainable
DeepSeekTextNone publishedNot obtainableNot obtainable
AI detectorsGPTZero, Turnitin and similarTextOpenThe prose resembles model output statisticallyThe prose does not. Neither is watermark evidence.

Method Compiled from provider documentation. Positive and negative meanings are stated from each tool's own described scope, not inferred. Every row re-checked against the status database on 2026-10-06.

Checked 2026-10-06

Detector guide

Whose watermarks does each AI detector actually check?

Confirmed

OpenAI's Verify tool looks for OpenAI's signals; Google's checker recognizes Google's. The stated boundary of each, down to Google's ten-checks-a-day cap.

Each watermark detector checks one company's own signal, not AI content in general. OpenAI frames its image-verification tool as checking for provenance signals its own tools produced, specifically C2PA metadata and SynthID watermarking, rather than judging whether an image looks AI-made.

Feed it a file OpenAI's tools generated and it reports whether those particular signals are still intact. That's the whole question it answers.

It does not accept text. OpenAI's text watermark, textGrain, announced on 2026-10-05 for EU ChatGPT and Codex text, has a separate detector that is limited at launch to approved researchers and expert organizations. There is no public way to check text for it.

Google states its boundary outright instead of leaving it implied. Its support page, relocated and re-read on 2026-09-25, scopes Gemini verification to images, video and audio, documents SynthID and Content Credentials checks, and attaches strict limits.

  • Images, video, and audio only. Not text.
  • Files capped at 100MB.
  • Video under 90 seconds.
  • Audio under an hour.
  • Roughly ten checks per content type every 24 hours.

That's not a check-anything-anytime tool. It's a rate-limited check for one company's own mark, and every video or audio detector built around a single provider's signal works the same way.

Why can't most people reach Google's SynthID Detector?

Confirmed

Ten billion images and video frames marked, against a checking portal that has stayed waitlist-only for journalists and researchers since May 2025.

Most people cannot reach Google's SynthID Detector portal because access to it has stayed waitlist-gated since it launched in May 2025. Google also runs a separate, standalone SynthID Detector portal, launched 2025-05-20 at Google I/O, announced for scanning image, audio, video, and text uploads for SynthID marks. No published Google route accepts text today; the portal and the Gemini app take image, video and audio only.

At launch only image detection actually worked. Video and text detection were promised "in the coming weeks."

Access has stayed waitlist-gated to journalists, media organizations, and researchers, with no public API. An independent check by journalist Henk van Ess on 2025-11-20 still found waitlist-only access, with in-app audio and video verification described as "hit-or-miss" while image checks held up reliably.

OpenAI's provenance checking arrived modality by modality on a similar pattern: image signals from 2026-05-19, then audio signals plus a Content Provenance API for third parties on 2026-07-31, two days before EU AI Act Article 50 took effect.

Why does a watermark detector find nothing on real AI content?

Confirmed

A detector needs to know which scheme and whose configuration before its output means anything. Without both, nothing found is just a mismatch.

A blank result usually means the detector was pointed at the wrong scheme, or the signal was stripped in ordinary handling, not that nothing was ever there. A watermark detector needs two facts before its output means anything: which scheme was used, and whose configuration of it.

Google's description of SynthID Text makes the underlying mechanic explicit. Detection there is score-based, checking whether token choices match a statistical pattern more than chance would predict, so without knowing which configuration to test against there is nothing valid to score.

Image, video, and audio detectors carry the same requirement even though the signal lives in pixels, waveforms, or frames instead of word choices.

Ordinary handling produces the same blank result with nobody removing anything on purpose. A 2026-06-28 thread on OpenAI's own community forum documented C2PA metadata stripped down to bare PNG chunks the moment an image was re-uploaded into a ChatGPT conversation, while the original file, downloaded straight from the source, still verified cleanly.

A detector run on that re-uploaded copy reports no signal, correctly, and tells you nothing at all about whether the image ever carried one.

What do independent findings say about Sora 2 and SynthID?

Community discussion

Two outside findings sit awkwardly against the official account: a Sora 2 metadata gap, and an uncorroborated claim of 91% SynthID removal.

Two outside findings complicate the official account, and the better documented one concerns Sora 2: a named author, posting in November 2025, reported that Sora 2 videos carrying the visible watermark had no detectable C2PA metadata via either of two verification tools, one of them the open-source c2pa-rs CLI, while Pro-tier videos without the visible watermark did. OpenAI's own Verify tool does not accept video.

The author's own framing is worth keeping: practical risk is low, since C2PA metadata is trivially strippable anyway. The credibility problem is the part that stands.

Separately, an independent researcher claims spectral-analysis techniques reach 90% detection and 91% removal of image SynthID marks. That work is not peer-reviewed and not corroborated elsewhere. Three days after the viral thread, a correction thread citing the researcher's own admission put the evasion rate at 16 percent, so treat the higher numbers as a claim that has already been walked back.

Does a clean detector result prove content is human-made?

Confirmed

Here's the sentence to keep handy when someone waves a detector result at you.

A clean detector result never proves content is human-made; it only narrows the possibilities. There's a real difference between a watermark detector and a generic AI detector. The generic kind infers whether content statistically looks AI-generated, from style or pixel artifacts, and returns a likelihood score.

Treating a "looks human" score as equivalent to "no watermark found by a scheme-specific check" overstates what either tool verified.

The same caution applies when the correct scheme-specific detector is used and comes back clean. OpenAI states plainly that no detection method it offers is foolproof, and that its provenance signals are not a guarantee content is accurate, unedited, or shown in its original context.

A missing signal has at least four ordinary explanations.

  • The content was never watermarked by that scheme.
  • A real watermark degraded through editing, compression, or a re-upload.
  • It came from a different generator entirely, watermarked or not.
  • The detector you used doesn't cover that provider's signal in the first place.

None of those add up to "proven human-made." A detector result narrows the possibilities; it never settles the question by itself.

FAQ

Can one detector find all AI watermarks?

No, and there isn't a universal one to reach for. Detection depends on matching the specific watermarking method and provider configuration used at generation time, and each provider's checker (OpenAI's Verify tool, Google's in-chat checker, the separate SynthID Detector portal) covers only that provider's own signal. A detector built for one scheme has no way to check for a different one, and some of these tools aren't even open to the general public yet.

If a detector finds nothing, does that mean the content is human-made?

No. A clean result only means one specific scheme's signal was not found by that particular check. The content could have come from a different generator, from one that doesn't watermark at all, or a real watermark could have degraded past recoverability through editing, compression, or something as routine as a re-upload. One OpenAI community forum thread documented exactly that: C2PA metadata stripped to bare PNG chunks after an image was re-uploaded into a ChatGPT conversation, even though the original file still verified cleanly.

Is Google's SynthID Detector portal open to the public?

Not as of the most recent independent check. It launched 2025-05-20 with only image detection working, while video and text detection were promised but not live. Access has stayed limited to journalists, media organizations, and researchers on a waitlist, with no public API, and a November 2025 independent check found it still waitlist-gated, with audio and video verification inconsistent even for those who had access.

Why do image, video, and audio watermark checks work differently from text?

Text detection for schemes like SynthID Text is score-based: a detector checks whether a passage's token choices statistically match a watermark configuration more than chance would predict. Image, video, and audio detectors look for a different kind of embedded signal, but the same requirement holds either way. The detector needs to know which scheme and configuration to test against, or the result is meaningless.

Next steps

  • Text detection uses a different mechanism from anything on this page. Start there if the thing you're checking is prose. How text watermark detection works
  • If someone has handed you a detector score and called it evidence, the distinction that matters is watermark versus style guess. AI watermark vs AI detector
  • Read the documented misfires before trusting any flag, including a Google checker bug that returned the wrong verdict across four tested sessions. AI detector false positives

Sources and citation status