An AI fake checker that anybody can use sounds like an answer to a familiar problem: somebody sends you an extraordinary video, and you want to know whether it is real. On 7 October, Google opened its SynthID Detector to the public worldwide in English. It can look for invisible watermarks in supported images, audio and video, including material made with AI systems from several participating companies.
That does not mean the mystery of synthetic media has been solved. Google’s announcement of the public rollout describes a system for finding a particular kind of digital mark, not a universal lie detector. If the detector cannot find that mark, the file may still be artificially generated, edited or misleading. The distinction matters when a clip involves a politician, a family member or a request to send money.
How the AI fake checker actually works
Rather than guessing whether a face looks strange or a voice sounds robotic, SynthID looks for signals deliberately inserted when participating AI tools generate content. These signals are intended to survive ordinary handling without changing what a human sees or hears. Google’s technical explanation of SynthID describes watermarks embedded into the content rather than a simple visible badge placed in a corner.
For a person investigating a suspicious photograph, that is a meaningful distinction. A visible badge can be cropped away. A watermark woven into the media may be harder to remove without changing the file. Google says its system covers images, video and audio, although its capabilities depend on how content was made and what changes happened afterwards. The detector is checking for an established signal, not inspecting the moral intentions of whoever posted the material.
Google’s expansion also matters because the participating network is wider than Google’s own products. The company names OpenAI, Nvidia and Kakao among partners, with Apple described as forthcoming. That increases the usefulness of the check, but it does not imply that every product from every partner is already covered. Coverage should be understood at the level of supported content and watermarking, not brand names alone.
A missing watermark is not a certificate of truth
Imagine receiving a voice note that appears to be from a colleague. A positive, reliable watermark result could be useful evidence that supported AI technology created or altered the recording. A negative result is weaker information. The original might have been made with an unsupported tool, the watermark may have been damaged, or the recording might be entirely real but presented in a false context.
The same uncertainty runs in the other direction. An authentic photograph can be paired with a fabricated claim about when or where it was taken. No watermark detector can settle that context by examining pixels alone. It is therefore sensible to compare a suspicious file with independent reporting, original uploads or another way of contacting the apparent sender before acting on it.
This is especially important in scams. LiveAIWire has examined the problem of real-time deepfake callers; a convincing-looking or convincing-sounding person is not sufficient proof of identity. The practical question is whether there is trustworthy evidence behind a request. Technology can contribute to that decision, but it should not replace it.
Why Google is making the detector public now
SynthID previously had a restricted verification portal for journalists and selected media professionals. Google’s October announcement makes that portal available more broadly, turning a newsroom tool into something an ordinary reader can try. Google also says verification features already built into Search, Gemini and Chrome handle over a million requests a day. Those figures are company-reported, not an independent measure of accuracy.
The timing reflects a wider problem: synthetic images and sound are increasingly easy to make, while established social habits of trusting what appears on a screen are much harder to change. A publicly available check offers a low-friction starting point. It could help a parent investigate a misleading clip or a small business challenge a dubious audio message before a costly mistake is made.
There is also a question of privacy. Uploading a sensitive picture or recording to any verification service creates a separate decision about what information to share. Before submitting private workplace material, personal documents or family media, users should review the service’s current handling rules. The broader issues of what automated systems can infer from ordinary content are explored in LiveAIWire’s report on AI privacy inferences.
The question to ask after any result
A sensible way to use the new checker is to treat its output as one clue. If a supported watermark is found, consider what the result says about generation or editing and whether that matters to the claim being made. If no watermark is found, resist the temptation to treat silence as exoneration. The most serious falsehood may not be the image at all, but the story attached to it.
Google’s move is a useful advance because verification is more accessible than it was. Its limit is just as important as its convenience. An AI fake checker can answer whether a particular watermark is present. Deciding whether a message, promise or alleged event is trustworthy still requires evidence beyond the file itself.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.