SynthID Detector is now open to the public, Google DeepMind said Wednesday, letting anyone check whether online content was generated with artificial intelligence tools, including those made by outside companies.
Key Takeaways
Google DeepMind opened SynthID Detector to the public, allowing anyone to check online content for AI-generation signals
SynthID Detector scans images, audio, video and text and returns a probability score rather than a yes-or-no answer
The system can identify signals from Gemini, Imagen, OpenAI and Nvidia tools through integration agreements
Content from non-participating AI tools or content stripped of its watermark will not register
What SynthID Detector Checks
The tool scans images, audio, video and text for an embedded signal flagging AI generation. DeepMind’s post said it can identify content made with Gemini and Imagen, as well as tools from partners including OpenAI and Nvidia (NVDA) through integration agreements.
Users upload a file or paste a link into a web interface, and it returns a probability score rather than a yes-or-no answer.
The score reflects that watermark signals can weaken when content is cropped, recompressed or run through another tool.
SynthID is a watermarking system DeepMind began building in 2023. It embeds an imperceptible statistical pattern in AI-generated pixels or audio waveforms at creation.
The pattern survives most routine edits, resizing, screenshots and light compression, but not every edit.
Also Read: Data Center Bill Launch Targets AI Power Fight
The system works only when the tool that made the content embedded the signal. No law currently requires that, leaving a structural gap.
Content from non-participating AI tools, or content deliberately stripped of its watermark, will not register regardless of how well the detector performs on cooperating systems.
The release follows a year in which AI video and voice-cloning tools have outpaced most platforms’ ability to flag them, feeding deepfake fraud and election disinformation cases worldwide. DeepMind had limited verification to partner platforms and research institutions.
Opening it extends that checking capability to anyone with a browser.
On an ordinary day, SynthID Detector can check only for signals embedded by participating systems. It does not establish whether all AI-generated content can be identified.
What remains unresolved is how many model makers outside DeepMind’s partner circle opt in, since labs running closed models with no watermarking commitments remain outside its reach.
Read Next: Two Words Nearly Double A Language Model’s Math Score