From 31 July, audio generated through ChatGPT Voice and the OpenAI API carries an invisible watermark called SynthID1, and OpenAI has extended its Verify tool so people can check whether a piece of audio came from its systems. Coverage of this has leaned on the phrase "fighting AI voice scams." The actual mechanism is narrower than that framing suggests, and worth being precise about before anyone relies on it for anything.
What the watermark actually does
SynthID is watermarking technology from Google DeepMind, embedded as an inaudible pattern in AI-generated audio that survives common transformations like compression. As of 31 July, OpenAI applies it to supported audio generated with GPT-Live, the model behind ChatGPT Voice. OpenAI's Verify tool can then check an uploaded audio file for that pattern and report whether it was made using OpenAI's systems. That is genuinely useful, as far as it goes: it's a real, checkable signal, not a marketing claim.
AI Fit Session
Not sure where to start? Over 1 week and three 60-minute 1-to-1 sessions, we map your business, find the quick wins, and build a practical starting point specific to you.
AI Foundations
Eight weeks, one coaching session a week, seven recorded modules. By the end you'll have AI working across real areas of your business and a 90-day plan to keep building.
AI Growth
Already using AI but want to keep developing it? Each month we work on real tasks from your business together, refining what's in place and building what's next.
Why it won't catch most scam calls
The gap is that SynthID isn't one shared system that flags any AI voice regardless of who made it. OpenAI's watermark only marks audio made with OpenAI's own tools, and OpenAI's Verify tool only checks for OpenAI's own signal. Google runs a separate SynthID Detector2 that only recognises Google's own models. ElevenLabs runs a third, separate detector3 that only recognises ElevenLabs output, and that rollout only began in June 2026, starting with free-tier users. There is no single tool that answers "was this AI-generated": there are three tools that each answer a narrower question, "did this specific company's system make it."
- OpenAI's Verify tool checks for OpenAI's own SynthID and C2PA signals only.
- Google's SynthID Detector checks for Google's own models only.
- ElevenLabs' Audio Detector checks for ElevenLabs output only, and that coverage is still expanding.
- Voice-cloning tools built by anyone else, and there are many freely available ones, aren't covered by any of the three.
What this means for a suspicious call
Two practical limits worth knowing. First, these tools work on an uploaded audio file after the fact, not on a live call, so there's nothing to check in the moment someone rings claiming to be a supplier or a client asking for an urgent payment. Second, most voice-cloning software used in scams isn't built by OpenAI, Google, or ElevenLabs specifically, so even checking the recording afterwards would most likely come back clean, not because the call was genuine, but because the tool that made the fake voice was never in the watermarking system to begin with.
What is SynthID watermarking?
SynthID is watermarking technology from Google DeepMind that embeds an inaudible, invisible pattern into AI-generated audio, images, and video. The pattern survives common transformations like compression and can be detected by a matching checking tool, but only one built by whichever company applied that specific watermark.
Can I check if a phone call was AI-generated?
Not reliably, and not in real time. Verification tools like OpenAI's Verify or Google's SynthID Detector work on an uploaded audio file after the fact, and each only detects that one company's own watermark. There's no tool that checks a live call or that recognises AI audio made by every provider.
Does OpenAI's audio watermark stop voice-clone scams?
No. It marks audio generated through OpenAI's own ChatGPT Voice and API from 31 July 2026 onward, and lets you check for that specific watermark. Most voice-cloning software used in scams isn't built by OpenAI, so it wouldn't carry this watermark and wouldn't be caught by checking for it.
What should I do if I suspect a call is an AI voice clone?
Don't act on urgency alone. For anything involving money, credentials, or a change to payment details, hang up and call back on a number you already have on file, not one given to you during the call. This procedural check is more reliable right now than any AI-audio detection tool.