Pushmeet Kohli still remembers the moment the numbers crossed a threshold that made even him pause. Over 180 billion images and videos. Another 240,000 years of audio. All carrying an invisible mark inserted at the moment of creation by Google’s AI systems and now a growing list of partners.

Today that mark, known as SynthID, moves from specialist tool to public utility. Anyone with an internet connection can visit synthid.com, upload a file, and learn whether it carries the hidden signal from models built by Google, OpenAI, NVIDIA, Kakao or, before long, Apple. The Google blog post announcing the change lands with deliberate understatement. Yet its implications stretch across newsrooms, courts, advertising agencies and scientific databases.

The detector itself arrives three years after SynthID first appeared in 2023. Early versions stayed behind closed doors or limited beta programs for journalists and researchers. Last year Google gave a narrow group of media professionals access to a prototype. Now the gates swing open. The site accepts common image formats from JPEG to HEIC, video files in MP4 and WEBM, and audio in everything from WAV to M4A. One click, one login through a Google, Apple or OpenAI account, and the system returns its verdict.

But. The verdict comes with sharp boundaries.

SynthID only detects content watermarked by participating models. It says nothing definitive about files created outside that circle. A clean result does not prove an image is real. It simply means no recognized watermark was found. Kohli, VP of science and strategic initiatives at Google DeepMind, acknowledges the gap in the announcement. “As AI tools become better and easier to use, having clear context about the content you encounter online helps you make informed decisions about what to trust,” he wrote.

That context now flows through multiple channels. Google Search, the Gemini app and Chrome together process more than one million verification requests each day. The new public portal joins them as a standalone destination. And the technology itself has spread. OpenAI embeds SynthID in its image outputs. NVIDIA applies it to certain foundation models. Kakao joined the effort. Apple plans to follow.

The watermark lives inside the data. For images and video it alters pixel patterns in ways invisible to the human eye yet readable by the detector. Audio versions modify the waveform. The signal survives cropping, compression, resizing and even screenshots in many cases. Tests published earlier this year by Ars Technica showed the mark holding up after hundreds of compression cycles. Simple edits rarely destroy it. Yet running an image through a second generative model often does.

This weakness matters. Bad actors can launder content by re-generating it. The detector highlights regions of an image or video where the mark appears strongest, giving investigators visual clues. Still, the system remains silent on whether a file was merely edited with AI tools rather than created from scratch. It cannot spot generations from non-participating models that dominate open-source communities.

Industry watchers point to these limits while praising the progress. A TechCrunch report published hours after the announcement notes that previous detectors were model-specific. Gemini could only spot Google watermarks. OpenAI’s checker recognized only its own. The new portal reads across participants, a rare display of coordination among fierce competitors.

Google has not stood still. It combined SynthID with the C2PA content credentials standard, layering invisible marks with metadata that provides additional provenance information. Visible “sparkle” watermarks in Gemini outputs became optional earlier this year after creators complained they interfered with professional work. The invisible signals stayed mandatory.

Real-world impact has already surfaced. In July a hoax image of Senator Mitch McConnell hooked up to hospital tubes spread across Reddit and X. Snopes used SynthID to confirm it was AI-generated. The watermark survived multiple screenshots and platform compressions. That single case offered proof of concept for fact-checkers who had grown tired of chasing shadows.

Yet success stories remain exceptions. Most synthetic content online carries no mark at all. And the arms race continues. Researchers have attempted to reverse-engineer the system. One developer claimed to have done so with a few hundred generated images and signal processing techniques. Google disputed the claim, insisting the core technology holds.

The company also extended the concept far beyond media. On September 30 it introduced SynthID Bio. The same idea, applied to protein sequences and 3D molecular structures designed by AI. A DeepMind technical post describes how the watermark is embedded directly into amino acid choices or atomic coordinates. Laboratory tests on proteins targeting VEGF-A, the SARS-CoV-2 spike protein and PD-L1 showed watermarked versions performed as well as unmarked ones. Binding affinity, hit rates and sequence diversity remained intact.

The biosecurity implications are immediate. Synthetic biology databases could soon require proof that submitted designs carry a verifiable mark. Regulators worry about AI-designed pathogens or toxins slipping into open repositories. SynthID Bio offers one mechanism to trace origin without compromising function. Early collaboration with Stanford’s Hie lab and the Arc Institute has already watermarked genomes of bacteriophages designed by advanced genomic models.

Pushmeet Kohli and his team present these expansions as part of a larger pattern. Watermarking is not a finished product. It is iterative defense. Each new attack vector prompts refinements. Rate limits on the detector, currently around ten checks per day for similar images, exist to prevent adversaries from systematically probing and bypassing the system.

Competitors pursue their own paths. Microsoft and Meta developed separate standards. Anthropic adopted a version of Google’s open-source text watermarking for Claude. The European Union’s AI Act pushes all major players toward machine-readable labeling. Interoperability remains the stated goal even if technical and commercial frictions persist.

For news organizations the public detector arrives at a tense moment. Deepfakes grow more convincing. Advertising creatives experiment with AI-generated assets at scale. Scientific publishers face questions about AI-designed molecules. A tool that works across four major AI developers plus one imminent addition gives journalists and moderators something concrete to check. But it also forces them to explain its shortcomings to audiences expecting simple truth-or-fiction answers.

One million daily requests already flow through Google’s integrated verification features. That volume will almost certainly rise now that the barrier to entry has vanished. Synthid.com requires no special credentials beyond a standard login. The interface is deliberately plain. Upload. Wait. Receive a confidence score and, for images and video, a highlighted map of where the signal lives.

Google frames the rollout as one piece of a broader effort to restore trust in online media. Others see it as table stakes. The technology cannot stop determined actors from generating unmarked content or stripping marks through regeneration. It can, however, raise the cost and complexity of deception. And in an information environment where speed often beats accuracy, even modest friction matters.

Kohli’s blog post ends without fanfare. No promises of perfect detection. No sweeping claims about solving disinformation. Just a new portal, a larger partner list, and an updated scale of deployment. The numbers keep climbing. Billions more watermarked assets will emerge in the coming months. Apple’s integration will widen the net further. Whether the public learns to consult the detector before sharing the next viral image remains an open question.

Yet the infrastructure now exists at global scale. For the first time, ordinary users can query the provenance of AI-generated media from some of the most powerful models in existence. The signal is there. The decision to look belongs to everyone else.