Google’s SynthID AI content detector expands globally, will work on Apple’s models ‘soon’

SynthID's public detector can confirm a watermark, but a clean result proves nothing about who made a file

Google opened its SynthID Detector to the public in October 2026, globally but in English only, according to 9to5Google, Unite.AI and Search Engine Journal. Users upload an image, video or audio file at synthid.com. The tool looks for Google's invisible watermark, including versions applied by OpenAI, NVIDIA and Kakao. Apple is promised, without a date. Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio since 2023. It also says built-in checks in Search, the Gemini app and Chrome handle over 1 million requests daily.

Can the SynthID Detector tell me a file is AI-generated?

Only partly. It confirms whether a file carries a watermark from Google or a partner tool. A missing watermark says nothing about whether another generator, or a person, made the file.

Google's own FAQ calls the tool "not a general AI detector". It recognises only media from companies that adopted SynthID. A positive result says the file was made or edited with Google AI. Search Engine Journal noted that a retouched photo and a fully generated one can get the same label. 9to5Google frames the tool as a way to spot human-made images, which it cannot do. Treat a hit as evidence and a miss as no information.

Does it help with AI-written text?

Not directly. The portal is described as checking images, video and audio, and none of the coverage says text uploads are accepted, though Gemini app text is reportedly watermarked at generation.

Unite.AI, relaying DeepMind's description, says the text mark works by adjusting the probability scores a model assigns to candidate words. That is marking at generation, unlike classifiers that guess from patterns afterwards. Nothing in the coverage says how it survives paraphrasing or humanizer tools. The robustness claims for images and video, covering cropping, filters and compression, come from Google's own page. Our earlier coverage of Quartz's AI-assisted output is the case in point: written text sits outside this tool.

How much should I trust Google's scale figures?

Very little for buying decisions. They are Google's own counts of what it has watermarked and how often users check, not measures of detection accuracy or of how much web content is AI-made.

In a May 2026 post Google cited more than 100 billion watermarked images and videos, 60,000 years of audio and 50 million Gemini app verifications. The new totals are cumulative output counts, tallied by Google in units of its choosing. The 1 million daily figure counts requests, not positive finds. None of the three outlets reports a false-positive or false-negative rate.

What should I ask my detection vendor?

Ask whether the product checks for provenance watermarks at all, what it reports when none is found, and whether accuracy figures come from independent testing rather than the vendor's own samples.

Apple is the loose end. 9to5Google's headline says SynthID will work on Apple's models. Google says only that Apple Intelligence images will carry a watermark soon, and Search Engine Journal notes there is no release date. Until then those images come back clean. Ask whether your vendor's image checks cover Apple output yet.

Frequently asked questions

Does it work outside English?

Availability is global, but Google's post references English only. 9to5Google expects wider support later, with no languages or dates given.

Should I upload the original or a screenshot?

Upload the best-quality copy you have. The site says that gives the most accurate result. DeepMind says image and video marks are designed to survive cropping, filters and compression.

When will Apple images be detectable?

Google says soon. No date has been published.

Sources: 9to5Google, Unite.AI, Search Engine Journal.