On July 31, OpenAI began embedding Google DeepMind's SynthID watermark into every piece of audio produced by its GPT-Live voice model, covering both ChatGPT Voice and the OpenAI API. Alongside it, the company opened a verification API that lets outside parties query programmatically whether a given audio file carries the mark. The timing lines up with EU transparency rules that take effect on August 2.

An Inaudible Mark, and a Way to Check It From Outside

Two things changed. Audio generated by GPT-Live through ChatGPT Voice and the API now carries a watermark that no human ear can pick out. And the presence of that watermark can now be queried from code.

OpenAI took a similar step on May 19, when images created through ChatGPT, Codex, and the API began carrying both C2PA signed metadata and a SynthID watermark. The verification method released at that point, however, was a public web page: a person uploads a file and reads the result. Useful for individuals, but impossible to wire into someone else's processing pipeline.

The new verification API fills that gap. A newsroom screening submitted audio, a platform filtering uploads at scale, a compliance team reviewing outbound recordings — these are the intended cases. A watermark that exists but that nobody can query does very little. It becomes a system only when a third party can ask the question.

The Mark Goes Into the Frequency Picture, Not the Waveform

Image watermarking works by adding tiny perturbations to pixel data after the image is produced. Audio does not allow that approach. For voice, SynthID operates on the spectrogram — the picture of how an audio signal's frequency content shifts over time.

The process runs in roughly three stages. The waveform is converted into a time-frequency representation, a watermark pattern is written into that representation, and the waveform is then reconstructed from the modified spectrogram. Where the pattern goes is decided by psychoacoustic masking, which exploits the fact that the human ear cannot perceive quiet sounds sitting close to louder ones. It is the same principle that lets MP3 discard data without an obvious drop in quality: the signal hides where nobody is listening.

That placement makes the watermark unusually durable. It reportedly survives MP3 compression, changes in playback speed and pitch, and even the round trip of playing audio through a speaker and re-recording it with a microphone. Metadata attached to a file would typically vanish at the first of those steps.

EU Disclosure Rules Kick In on August 2

The timing is not a coincidence. Article 50 of the EU AI Act applies across all 27 member states from August 2. It requires providers to mark AI-generated audio, images, video, and text in a machine-readable form so the content can be detected as artificially generated.

Penalties for non-compliance run up to 15 million EUR (about 2.7 billion yen) or 3 percent of worldwide annual turnover, whichever is higher. Systems already placed on the market before August 2 get a grace period on this specific marking obligation, which extends to December 2.

※1 EUR = 182 JPY (as of August 1, 2026)

The text asks for solutions that are effective, interoperable, robust, and reliable as far as is technically feasible. Metadata alone strips off too easily; a watermark alone carries too little information. OpenAI's decision to stack C2PA and SynthID together reflects a judgment that a single layer would struggle to meet that bar.

The Scope Is GPT-Live Audio, and It Stops There

The watermark applies to GPT-Live output. GPT-Live launched on July 8 as a full-duplex voice model that can listen and speak at the same time, and it replaced Advanced Voice Mode as the default behind ChatGPT Voice. OpenAI says more than 150 million people use ChatGPT's voice and dictation features each week. Two tiers exist: GPT-Live-1 for paid plans and GPT-Live-1 mini for free users.

The flip side is that anything not produced by GPT-Live carries nothing. Output from another vendor's speech synthesis, or a voice clone built with an open-source model, will return no signal from the verification API. And no signal is not proof that the audio is a real human voice. The API answers exactly one question — whether an OpenAI-origin watermark was found — and it is not a general-purpose synthetic speech detector. Anyone deploying it needs to hold that distinction clearly.

The strength of the watermark itself is also debated. Researchers have reported removing the SynthID signal at a high rate, and Google has disputed those figures. Even so, weakening a signal is not the same as erasing it, and the two-layer design means that a successful attack on SynthID may still leave C2PA metadata intact, or the reverse.

What exists today is the plumbing for asking the question. Whether platforms actually wire it into their content pipelines is a separate decision, and it belongs to them.

Summary

On July 31, OpenAI applied SynthID watermarking to GPT-Live audio generated through ChatGPT Voice and the API, and released a verification API for querying it. The mark is embedded in the spectrogram using psychoacoustic masking and survives MP3 compression and re-recording. Behind the move sits Article 50 of the EU AI Act, effective August 2, with penalties reaching 15 million EUR (about 2.7 billion yen) or 3 percent of global turnover. Detection, however, is limited to OpenAI-origin audio, and the absence of a signal proves nothing about authenticity.