We analyzed this claim: is it true that
Google watermarks all audio generated by Gemini 3.8 Live with SynthID?
Launch materials confirm that audio generated by Gemini 3.8 Live carries Google's SynthID watermark, though technical details remain partly undisclosed
Google watermarks its live artificial intelligence voices
Synthetic speech now carries an inaudible digital signature to help listeners verify its origin
In a nutshell
Google now weaves an inaudible digital watermark called SynthID into all spoken audio generated by its latest conversational model, Gemini 3.8 Live. First developed for music in 2023, the technology embeds a hidden signature into sound frequencies so files remain identifiable even after they are edited or compressed. While Google lets users scan clips for these marks, it keeps the underlying code private, meaning independent researchers cannot yet test how easily the protection can be bypassed.
Highlights
- Google launched SynthID for audio on November 16, 2023, to watermark computer-generated music.
- Google introduced a detector on May 20, 2025, that scans uploaded recordings to find watermarked sections.
- Google released Gemini 3.8 Live on September 15, 2026, confirming all its spoken responses carry the hidden stamp.
- Google paired SynthID with verbal consent checks on September 23, 2026, for tools that clone real human voices.
- Google expanded watermarking on September 24, 2026, to cover both synthetic speech and animated video avatars.
From the Editor’s Diary
A digital watermark can prove when a trusted company made a sound, but it cannot stop rivals from releasing unmarked voices that sound just as real.
Who's involved
Google DeepMind
Google's artificial-intelligence research arm that created the SynthID audio watermarking system
goal → make artificial speech recognizable without changing how it sounds to listeners
Google
the technology firm selling Gemini software to the public, businesses, and software developers
goal → release talking artificial-intelligence tools widely while letting people identify machine-made voices
Journalists, researchers and content verifiers
investigators checking sound recordings to see whether a computer made or altered them
goal → find dependable tools that confirm whether a spoken recording is genuine or fake
Gemini developers and enterprise customers
companies and software programmers building tools with Google's speech-generation software
goal → create realistic voice programs that carry built-in proof of their artificial origin
Voice performers and rights holders
actors and speakers whose real voices can be copied by speech computers
goal → stop people from copying their voices without permission
In short
Listeners can now check whether spoken audio came from Google's latest artificial intelligence, because the company weaves an inaudible digital fingerprint into every clip its newest model speaks.
This makes it likely that automated tools will increasingly detect AI-generated audio clips across the internet before they spread widely.
That outcome is uncertain, however, because independent researchers cannot yet test the secret code behind Google's detector to measure how often it misses an altered file or mistakes human speech for a machine.
How it unfolded
Watermarks expand from pictures to sound
DeepMind first applied SynthID to audio on November 16, 2023, for its Lyria music maker, embedding hidden signals into sound charts rather than pictures. That step established the basic method Google uses today to stamp live voices.
Google moves watermarks into spoken words
Google moved the stamp from music into speech engines, making it standard equipment across its talking products. In May 2025, Google unveiled an online scanner to spot these marks in uploaded recordings, and soon after mandated the mark for its Gemini 2.5 voice models.
Public tools arrive to check live voices
In early 2026, Google allowed regular users to upload sound files and ask whether its systems had made them. Watermarking then expanded to Gemini 3.1 Flash Live, showing that automatic stamping was already standard practice before the newest model arrived.
Gemini 3.8 makes the protection routine
On September 15, Google launched Gemini 3.8 Live for quick spoken chats and complex verbal reasoning, confirming all its audio is stamped with SynthID. The move kept an existing rule in place rather than inventing a new one.
Watermarks guard voice clones and video faces
As artificial voices became easier to copy from short recordings, Google paired watermarks with verbal consent checks and digital credentials. It also added moving video avatars, embedding SynthID into both the picture and the sound.
Where things stand
Google's newest conversational artificial intelligence, Gemini 3.8 Live, automatically marks all the spoken audio it produces with SynthID. The system works by embedding an inaudible digital code across the audio frequencies so the mark can survive background noise, file compression, and playback changes. The public can also use Google's verification tools to check whether an audio file contains this secret signature.
Significant blind spots remain. Because Google keeps the underlying code secret, independent scientists cannot verify how easily the mark can be erased or how often the detector makes mistakes. A watermark confirms that Google generated the sound, but its absence does not prove that a voice is human or that another company's computer did not make it.