OpenAI Introduces Invisible Watermark to Identify AI-Generated Voice Recordings
OpenAI has started embedding an inaudible digital watermark in supported audio files created using ChatGPT Voice and the OpenAI API, effective from July 31, 2026. This new feature aims to help identify voices generated by OpenAI's AI tools without displaying a visible label to listeners. The watermark technology, based on SynthID developed by Google DeepMind, embeds a digital signal within the audio file that does not noticeably affect sound quality but can be detected by compatible verification tools to confirm the file was produced by an OpenAI-supported system.
The update applies to audio generated through GPT Live, which powers ChatGPT's new voice conversation experience, as well as applications producing voice via the OpenAI API. OpenAI clarifies that only files created with supported tools and containing the watermark are identifiable, not all AI-generated audio online. Alongside this, OpenAI has expanded its public verification tool, previously focused on images, to allow users and organizations to upload supported audio files for authenticity checks. An API is also available for news sites, content platforms, and companies to integrate automatic verification into their systems.
This development follows OpenAI's May announcement adopting SynthID and the C2PA standard for marking AI-generated content, initially targeting images. The current extension to voice recordings aims to help detect deepfake audio, voice impersonation, and mistakenly presented recordings as genuine. However, OpenAI notes the system is not foolproof; heavy editing, format conversion, re-recording, or processing can damage the watermark, and absence of detection does not prove human origin. Users do not need to activate the watermark manually as it is embedded automatically in supported outputs. Recipients of suspicious recordings can verify them using OpenAI's tool, but for sensitive cases, additional verification through other communication channels is recommended.
