ChatGPT Watermark: Text, Images, C2PA, and SynthID

OpenAI joined the C2PA steering committee and reached "C2PA Conforming Generator Product" status on May 19, 2026, launching a public verification tool at openai.com/verify on the same day.

• Images from ChatGPT, Codex, and the OpenAI API: C2PA Content Credentials metadata plus a SynthID watermark.
• Supported audio from ChatGPT Voice, GPT-Live, and the API: the same approach, added July 31, 2026, two days before EU AI Act Article 50 obligations took effect.
• A Content Provenance API, announced the same day, so third parties can check for these signals programmatically.
• Ordinary chat text: nothing. OpenAI describes text marking as a goal, not a shipped feature.

Both image layers have been in place since that May 2026 rollout, and OpenAI frames them as complementary. C2PA carries detailed context when it survives, and SynthID is meant to preserve a signal when the metadata does not.

The caveats are worth repeating in OpenAI's own words. "No detection method is foolproof," C2PA metadata "can be stripped, lost through uploads and downloads, or broken by transformations," and the Verify tool deliberately avoids a definitive conclusion when it finds nothing.

So no signal detected is not proof that an image or clip was not made with ChatGPT. It often just means the file traveled.

Years before its 2026 image and audio work, OpenAI reportedly built an internal ChatGPT text-watermarking system. Leaked internal documents, reported by the Wall Street Journal and covered by Tom's Hardware on August 5, 2024, described it as roughly 99.9 percent effective at flagging ChatGPT-written text.

OpenAI chose not to release it. The reported reasons stack up like this.

• Circumvention risk: a watermark that can be defeated invites a false sense of certainty.
• False positives: 99.9 percent accurate still means a meaningful volume of wrongly flagged writing at ChatGPT's scale.
• Disproportionate impact on non-native English writers, whose prose can read as more mechanical to these classifiers.
• Internal survey data suggesting about 30 percent of users would use ChatGPT less if the feature shipped.

This is not only a leak. OpenAI has separately and publicly discussed text-watermarking research and its risks, naming false positives, circumvention, and disproportionate impact on non-native English speakers, in its own content-provenance posts.

Nothing published since 2024 changes the picture. OpenAI's 2026 provenance documentation still says the goal is to expand provenance signals to all modalities including text, which describes an intention rather than a status.

A November 6, 2025 investigation, published on LessWrong and cross-posted to the EA Forum, checked Sora 2 videos against both OpenAI's own Verify tool and the open-source c2pa-rs CLI. Some videos carrying the visible OpenAI watermark had no detectable C2PA metadata at all, while some Pro-tier videos without the visible watermark did carry it.

That is close to the reverse of OpenAI's public claim that every video generated with Sora includes both visible and invisible provenance signals. No OpenAI response to the finding has surfaced.

Separately, a thread on OpenAI's own Developer Community forum, dated June 28, 2026, found that a PNG carrying valid OpenAI C2PA metadata gets that metadata stripped down to bare PNG chunks the moment it is re-uploaded into a ChatGPT conversation. The original file still verifies fine on openai.com/verify. OpenAI staff had not addressed the thread as of this writing.

Neither finding means OpenAI's provenance system is broken. SynthID is a separate pixel-level mechanism from C2PA metadata, and the Sora investigator's own framing treats the practical risk as low, since metadata is trivially strippable anyway.

What they do show is concrete and dated: real daylight between "every image is watermarked" as a claim and what people find when they actually check.