Google opens its SynthID AI detector to the public

[Lampard80]

Google has opened SynthID Detector, a free web tool for identifying AI generated and AI edited content, to the public at synthid.com, Engadget reports. It was previously limited to journalists and media professionals, and Google has worked on it for over a year. The tool can flag images, video and audio made with tools from Google, OpenAI, Nvidia and Kakao. Users need to sign in with a Google, OpenAI or Apple account

Engadget says it will be one of the first detectors able to identify images from Apple's AI features such as Image Playground, with support for Apple Intelligence coming soon. ElevenLabs has previously committed to SynthID support for audio. That kind of cross industry support is what watermarking needs to be useful, since a detector that only recognises one company's content is of limited value

The tool has clear limits. It cannot catch all AI content, because companies use different watermarking standards. Watermarks can sometimes be removed. It also cannot tell the difference between content that was entirely AI generated and content that was only edited with AI. Meta's own detector, by contrast, only covers media from its latest Muse Image models

It fits a theme we have discussed several times recently, from the AI Lab's post quantum content signing protocol to our thread asking whether people can still tell AI content apart. Free public tools are a step forward, but they are not a complete answer. As Google puts it, the company is "continuously working on extending our partnerships to allow detecting more generated media." The more companies join, the more useful it becomes

Will you use SynthID to check suspicious images or videos? And can watermarking ever keep up with AI generated fakes?

Googlewww.engadget.com

[Peter82]

Google, OpenAI, Nvidia and now Apple is a decent coalition. If Meta and the major Chinese labs join, it starts to cover most mainstream content

Industry standards bodies have been trying to agree on this for years. The more companies sign up, the harder it becomes for others to stay out

[Rachel83]

Watermarking only works if everyone joins in. Open source models will never add watermarks, so bad actors will just use those. Detection will always be one step behind

It still raises the bar for casual fakes, which is something. Education about checking sources matters just as much

[RefreshAddict]

Not being able to tell fully generated from AI edited content is a big limitation. A photo with the background tidied up is very different from a completely fake scene

Context matters a lot for judging whether something is misleading. A label saying AI involved could cover anything from a filter to a total fake

[Linda52]

Useful for checking viral images before sharing them. I will definitely bookmark it

Even if it misses some content, a positive result is helpful. Fact checkers will find it handy too. Every bit of evidence helps

[Daemon55]

Journalists have had it for a year already. Opening it to everyone is good, but I would like to know how accurate it has been in their hands. Published accuracy figures would build trust

Real world performance matters more than lab tests. Hopefully Google publishes some data soon