Audio now carries a SynthID watermark
Provenance is the record of where a piece of content came from and how it was created or edited. OpenAI has been building this out for images since May 2026. The July 31 update takes it beyond images.
The three things the July 31, 2026 update changed
The scope is "supported audio"
Scope is the easy thing to misread. What the announcement says is "supported audio generated with OpenAI tools," explicitly including both ChatGPT and the OpenAI API. Which is to say it does not state that all generated audio is uniformly covered. No list of what counts as supported appears in the announcement either. For now, whether a given audio file is in scope has to be checked case by case.
Audio watermarking is not new territory for OpenAI. The company has used visible watermarks in Sora and an audio watermark in Voice Engine, and says it has continued to test and research accuracy and reliability over time.
Update July 31, 2026: We’re expanding this work beyond images. Supported audio generated with OpenAI tools, including through ChatGPT and the OpenAI API, now includes SynthID watermarking. Our public verification tool will now allow for verification of supported audio files in addition to images. We’re also introducing API access for verification, so developers and organizations can incorporate provenance checks into their own workflows. / We have used visible watermarks in Sora and an audio watermark in Voice Engine, and have continued to test and research accuracy and reliability over time. — From the expansion to audio, the verification tool's scope, the API access, and the prior watermarking work
What API access to verification means
The other change that lands in practice is the API. Developers and organisations can now incorporate provenance checks into their own workflows. Anyone accepting uploads can verify provenance as part of processing rather than through manual inspection.
The public verification tool itself checks whether an uploaded image — and, since July 31, supported audio — contains provenance signals including Content Credentials (the C2PA name for provenance information) and SynthID, and verifies whether it was generated on ChatGPT, the OpenAI API, or Codex.
We’re also introducing API access for verification, so developers and organizations can incorporate provenance checks into their own workflows. / We are now previewing a public verification tool that will help people verify whether an uploaded image was generated on ChatGPT, the OpenAI API, or Codex, by checking if it contains provenance signals, including Content Credentials and SynthID. — From the API access for verification and what the public verification tool checks
Why watermarks and C2PA are layered together
Provenance is not one mechanism. OpenAI layers C2PA metadata — the cross-industry standard for recording where content came from — together with the SynthID watermark. The reason is that each alone leaves gaps.
Why the two layers
| Layer | Strength | Weakness |
|---|---|---|
| C2PA metadata | Carries detailed information: origin, how it was created or edited, who signed it | Stripped or lost through uploads and downloads; broken by format changes, resizing, screenshots |
| SynthID watermark | Relatively durable through transformations like screenshots, so the signal survives | Cannot carry as much information as metadata |
Metadata comes off easily
C2PA is a standard that uses metadata and cryptographic signatures to carry information about a piece of media along with the content itself — where it came from, how it was created or edited, and who signed that information. It is not foolproof. It can be stripped, lost through uploads and downloads, or broken by transformations like file format changes, resizing, or screenshots.
SynthID fills that gap. It embeds an invisible watermarking layer that leaves a signal even when the metadata does not survive. One layer supplies detail, the other supplies durability.
C2PA metadata is an important foundation for provenance. It helps content carry information about where it came from, how it was created or edited, and who signed that information. But metadata is not foolproof. It can be stripped, lost through uploads and downloads, or broken by transformations like file format changes, resizing, or screenshots. / SynthID embeds an invisible watermarking layer that complements C2PA metadata-based approaches. / These two systems reinforce each other. C2PA helps content carry detailed context; SynthID helps preserve a signal when metadata does not survive. Watermarking can be more durable through transformations like screenshots, while metadata can provide more information than a watermark alone. — From the role and fragility of C2PA metadata, what SynthID adds, and how the two layers reinforce each other
The limits of verification, and the labelling duties that began the same week
This is the part that matters. What the mechanism can and cannot establish is distinctly asymmetric.
What the verification tool can and cannot tell you
"Not detected" is not an acquittal
OpenAI states its own caution explicitly. When no metadata or watermark is detected, the tool will not make a definitive conclusion about whether the content was generated with OpenAI tools — because provenance signals can in some cases be stripped.
Scope is limited too. At launch the tool only handles content generated by OpenAI; anything made with another company's AI cannot be assessed. OpenAI says it aims over the coming months to support cross-industry efforts to make verification possible across platforms.
No detection method is foolproof, so we take a cautious approach in cases when detection fails. If no metadata or watermark is detected, for example, the tool will not make a definitive conclusion about whether the image was generated with OpenAI tools since provenance signals can in some cases be stripped. / At launch, the tool is limited to content generated by OpenAI. In the upcoming months, we aim to support cross-industry efforts to make verification possible across platforms. — From how failed detection is treated and the limits on scope
It lands alongside new labelling duties
Two days after this update, on August 2, 2026, the EU began enforcing its labelling obligations for AI-generated content, and California's AI Transparency Act became operative. Both require provenance disclosure in machine-readable form.
OpenAI's announcement offers no regulatory rationale. The timing overlaps, but with no official statement citing regulation as the reason, reading it that way goes further than the record supports. What holds is the factual overlap: the side recording provenance and the side requiring disclosure moved in the same window. Anything past that is inference.
These transparency obligations, applicable from 2 August 2026, complement other rules like those for high-risk AI systems or general-purpose AI models — From the statement that the EU transparency obligations apply from 2 August 2026
This bill would delay the operation of the California AI Transparency Act until August 2, 2026. — From the statement that the California AI Transparency Act becomes operative on August 2, 2026
Official announcements in English can render differently depending on the language toggle, which makes dates and qualifying words harder to pin down. Converting the page to Markdown before reading is the safer way to check specific passages against the original.
Conclusion: a watermark proves presence, never absence
The July 31, 2026 update extended OpenAI's provenance work to audio. Supported audio from ChatGPT and the API carries a SynthID watermark, the public verification tool handles audio, and verification can be called from the API. The design layers C2PA metadata with the SynthID watermark so that one survives when the other does not. But verification only establishes that a signal was there. Nothing detected does not mean not AI-generated, and content from other companies' AI cannot be assessed at all. So it is no basis for clearing anything. Build it in as a tool that answers one question — whether this came from OpenAI — not as one that settles true or false.
