AI Watermark Detector

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.

Google states the same boundary outright instead of leaving it implied. Its support page says Gemini verification "can currently only recognize content created by Google AI tools," and the limits attached are strict.

• 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.

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, built to scan image, audio, video, and text uploads for SynthID marks.

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.

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.

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 OpenAI's Verify tool or the c2pa-rs CLI, while Pro-tier videos without the visible watermark did.

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, so treat the numbers as a claim to watch rather than a fact.

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.