Google opens its SynthID AI-content detector to the public worldwide

TLDR

• Google has opened SynthID Detector, its AI content watermark checker, to the public worldwide, letting anyone check an image, video or audio file for free.¹²
• The tool can now identify watermarks from Google, OpenAI, Nvidia and Kakao, with Apple support coming soon, instead of only Google's own as before.¹³
• Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio since SynthID launched in 2023.¹
• Users get roughly 10 checks a day after logging in, a cap Google says curbs attempts to reverse-engineer the watermark.³
• The tool only catches content from its listed partners, missing Meta's own watermark standard and any unwatermarked open-weight AI output, and testing shows chatbots can still get the backstory wrong even when they correctly flag a fake.³⁴

Google opened its SynthID Detector website to the public on Tuesday, letting anyone upload an image, video or audio file to check whether it carries an invisible AI watermark.¹

Why it matters: until now, checking a suspicious file for an AI watermark mostly meant trusting a chatbot's answer, and this week's testing shows that answer can be confidently wrong even when the chatbot correctly spots a fake.⁴

Driving the news: the free site, SynthID.com, requires logging in with a Google, OpenAI or Apple account and gives each user about 10 checks a day.³ It had previously been limited to journalists, media professionals and researchers testing it since Google I/O last year, and before that the only way to check was asking Gemini directly.²³ The expanded detector now reads watermarks from all of Google's SynthID partners, Google, OpenAI, Nvidia and Kakao, with Apple's support coming soon, rather than only Google's own as before.¹³

By the numbers: Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio with SynthID since the technology launched in 2023, and that Search, the Gemini app and Chrome already handle over 1 million verification requests a day between them.¹

The other side: the detector only flags content from its named partners, so material made with Meta's own watermarking standard, open-weight models, or anything stripped of metadata will not be identified, and the site carries a disclaimer saying so.³ Gizmodo tested the gap with an AI-fabricated image of Texas attorney general Ken Paxton, Donald Trump and McDonald's mascot Grimace: asked directly, Gemini correctly flagged the image as AI-edited but then invented a fictitious "original" March 2024 photo and a dead link to support the claim, a reminder that catching a fake is not the same as explaining it accurately.⁴