# OpenAI Adds Invisible Watermarks to AI Voice Audio *GPT-Live audio can now carry SynthID signals, giving people and software a new way to check supported files.* By [Tech Tab](https://paragraph.com/@tech-tab) ยท 2026-08-01 artificial intelligence, digital trust, software --- OpenAI has added an invisible watermark to supported audio generated with GPT-Live, the voice system behind the new ChatGPT Voice experience. The company announced the change on July 31, alongside an expansion of its public verification tool and a new verification API for developers. The update matters because synthetic speech is becoming less like a recorded answer and more like a live conversation. GPT-Live can listen while speaking, handle interruptions and delegate harder questions to another model. As these systems move into customer support, media production and workplace tools, identifying where an audio file came from becomes a practical problem rather than a future concern. A signal inside the sound ------------------------- The watermark uses SynthID, technology developed by Google DeepMind. Instead of attaching a visible label, SynthID places a signal inside the generated media. The aim is to keep it imperceptible to listeners while allowing a compatible detector to recognize it later. OpenAI's browser-based Verify tool now accepts supported audio as well as images. Developers can also submit one file at a time to the Content Provenance API and receive a result in the same request. That makes the check usable inside content review, fact-checking and platform moderation workflows rather than only through a manual upload page. This is genuinely new for OpenAI audio. The company had already begun adding SynthID and C2PA Content Credentials to supported images. C2PA is signed metadata that can describe a file's origin and editing history. OpenAI's current documentation says its audio checks rely on SynthID, while its image checks can use both systems. Evidence, not a truth machine ----------------------------- The most useful detail is also the easiest to miss. A positive result means the verifier found a supported OpenAI provenance signal. A negative result does not prove that a clip is human-made. OpenAI says a watermark may be absent, unsupported or degraded. A file might also come from an older model, predate the new signals or have been generated by another company. Editing, compression and format conversion can weaken provenance information. The verifier is therefore a source check for part of the AI ecosystem, not a universal deepfake detector. Independent coverage of OpenAI's earlier image rollout reached the same practical conclusion. The Verge noted that metadata can disappear when media moves between platforms and that no detection method is foolproof. Google DeepMind's own description of SynthID says its detector can identify likely watermarked portions of audio, but broad access and real-world resilience still depend on the implementation and the transformations applied to a file. For publishers, support teams and online platforms, the practical response is simple. Add provenance checks where audio enters a workflow, preserve original files when possible and treat a detected watermark as evidence about origin. Do not treat an undetected watermark as evidence of authenticity. The missing angle in much of the announcement coverage is operational. Watermarking becomes valuable only when organizations build verification into routine review and explain inconclusive results honestly. OpenAI has now supplied the browser tool and API needed to start doing that. What remains uncertain is how reliably the new audio signal survives everyday reposting, editing and recompression outside controlled tests. Sources ------- [OpenAI's GPT-Live announcement](https://openai.com/index/introducing-gpt-live/), [OpenAI Content Provenance API guide](https://developers.openai.com/api/docs/guides/content-provenance), [Google DeepMind's SynthID explainer](https://blog.google/innovation-and-ai/products/google-synthid-ai-content-detector/), [The Verge on OpenAI's provenance rollout](https://www.theverge.com/ai-artificial-intelligence/933442/openai-synthid-content-credentials-c2pa-expansion), and [independent confirmation of the audio update](https://www.globemarketresearch.com/industry-news/openai-adds-synthid-watermarking). --- *Originally published on [Tech Tab](https://paragraph.com/@tech-tab/openai-adds-invisible-watermarks-to-ai-voice-audio)*