Google Opens SynthID Detector to Everyone — Here's What It Catches

Google has opened its SynthID Detector to the public, letting anyone with a Google, Apple or ChatGPT account upload images, audio and video to check whether the file carries a SynthID watermark indicating generative-AI manipulation. The launch matters because AI-generated media has become photorealistic enough that spotting fakes by eye — the era of counting six-fingered hands — is largely over.

The standalone detector currently recognizes media generated by Google, OpenAI, Nvidia and Kakao. Apple, which announced it would support SynthID at its fall event alongside the iPhone 18 Pro and iPhone Duo, is not yet an official partner — but Google says that support is coming.

What does the detector actually detect?

SynthID is Google's watermarking system for AI-generated content, introduced in 2023. In late 2025, Google made it possible to ask Gemini whether an image was created with Google's AI. At the Google I/O developer conference this past May, the company expanded detection to video and audio content in Google Chrome and Google Search. The new public-facing detector consolidates that capability into a standalone tool anyone can use.

The technology works only when a SynthID tag is embedded at the moment of generation or AI editing. It does not analyze an image for visual tells of fakery. That design choice shapes both its strengths and its blind spots.

How well does it work in practice?

CNET's early testing found the tool largely accurate, with instructive gaps:

• A photo shot on an iPhone 18 Pro and edited with Apple's Reframe tool was correctly flagged, meaning the detector already reads Apple's SynthID tags even before Apple's support is officially listed.
• An untouched iPhone photo correctly returned no tag.
• Images generated from prompts in ChatGPT were correctly flagged as AI-generated.
• An image created entirely in Apple's Playground app using ChatGPT's image generation model was not identified — likely because Playground has not yet wired SynthID tagging into its output pipeline.

Google's own devices illustrate a two-pronged strategy for image authenticity. Photos captured on a Google Pixel 11 Pro carry C2PA metadata — the standard from the Coalition for Content Provenance and Authenticity — verifying the original capture, and gain a SynthID tag only when edited with AI. Apple has adopted SynthID but has not signed onto C2PA. Instead, it relies on its own Reference Image technology, which stores a cryptographically verified original of a photo — but only for shots taken in Reference mode in the iPhone Camera app.

What happens to the files you upload?

Google handles uploaded media with a clear data policy. On first visiting the site, users must agree to terms of service stating that uploaded files are processed in real time and deleted after results are displayed. A digital signature of the media is mapped to the user's identifier and deleted after 24 hours, according to those terms.

Why does this matter for the broader detection race?

The stakes are straightforward: as generative models produce photorealistic images, cloned audio and synthetic video at scale, provenance tooling is one of the few structural defenses against so-called AI slop and deliberate disinformation. Google framed the cross-company approach as essential. A Google spokesperson said in an email to CNET that working with other companies is crucial because no single company can solve the problem alone, and said Apple support has no specific launch date but will be available soon.

The competitive picture is fragmented. OpenAI runs its own verification tool, but it only detects OpenAI watermarks — the AI-generated samples CNET tested through Google's tool came up negative on OpenAI's site. A detection ecosystem split by vendor means no single checker covers everything, which is precisely why Google's multi-provider detector marks a step toward practical, everyday verification.

There is a hard limit, though: the burden still falls on the viewer. Detection only works if someone notices something looks off, downloads the media, opens the tool and uploads it. Watermarking cannot flag content from generators that never embedded a tag. Google's Pixel approach shows the complementary role of capture-time provenance — C2PA metadata verifying what a camera actually recorded — versus generation-time watermarks verifying what a model produced.

The direction of travel is nonetheless clear. SynthID tagging is spreading into more media pipelines — Google's own apps, Chrome, Search, and soon Apple's — and the detector gives the public a free, low-friction way to check for it. As more platforms embed watermarks at the point of generation, tools like SynthID Detector shift AI-content verification from expert forensics to something any reader can run in a browser tab.