How to check media for AI with Google's SynthID Detector

Google's SynthID Detector was built for newsrooms. On October 7, 2026, Google DeepMind opened it to anyone with an account, in English, everywhere. The new portal is synthid.com. Upload an image, a video or an audio file, and it tells you whether the file carries a SynthID watermark, the invisible signal the company has been embedding in AI output since 2023.

The catch is in the wording. The page carries its own disclaimer, in plain text beside the upload button: "This is not a general AI detector." A positive result is real evidence. A blank result is not a certificate of authenticity, and people will quote a blank result as one anyway.

What the SynthID Detector actually checks

SynthID is a watermark, not a label glued to the outside of a file. It is written into the pixels of an image or video and into the frequency components of audio at the moment of generation. DeepMind's page says the mark is designed to survive cropping, filters, frame-rate changes and lossy compression, and that audio keeps it through added noise, MP3 compression and speed changes. Because the signal lives inside the content, it travels further than an attached label: a screenshot or a re-save that strips metadata will not strip a SynthID watermark.

The Detector looks for that one signal. Upload a file and it returns one of two readings, according to the portal's FAQ as reported by PPC Land: watermark detected, or watermark not detected. For images it reports whether a mark is present. For video and audio it flags the segments that carry one, so a partly generated clip can come back marked on only part of its timeline.

Google built the first version for a restricted group. It announced the Detector at Google I/O in 2025 with testing access for journalists, media professionals and researchers, TechCrunch reports, and the DeepMind page says the company spent a year collecting their feedback before opening the door.

What it will not tell you

Verification results get repeated in news stories and moderation calls, so the boundary matters.

A clean reading proves nothing. The portal only recognises marks from providers that have adopted SynthID: Google, OpenAI, NVIDIA and Kakao, with Apple "said to be adding support soon," in TechCrunch's wording. Everything from outside that circle, images from Meta's tools, Microsoft's, Midjourney, and every open-weight model anyone runs on their own hardware, comes back "not detected" whether or not a machine made it. Ars Technica puts the consequence plainly: plenty of AI content will not be identified as AI here.

Google's own help page for the Gemini app, support.google.com, uses the same caution. If a SynthID watermark is not detected, it says, the image or video was not created or edited by Google AI, "but it could have been created by other AI systems." The portal's FAQ goes further, saying a blank reading does not entirely rule out AI generation or manipulation, and naming two reasons a genuinely synthetic file can pass clean: it came from a model that never applied SynthID, or heavy editing degraded the signal. The FAQ also concedes the system can fire on content that was never watermarked, in rare cases.

The Detector reads one kind of signal. OpenAI runs a separate checker at openai.com/research/verify, and that tool looks for two things: C2PA Content Credentials and SynthID watermarks, reporting which one it found. The C2PA manifest is the richer record, a signed history of where a file came from. It is also the fragile one, because it lives in metadata that any re-encode can drop. SynthID carries almost no information but survives more. Google's portal documents describe a search for the SynthID watermark and say nothing about C2PA credentials, PPC Land reports.

Text is missing too. SynthID covers writing from Gemini, by adjusting the probability of each word as it comes out. The Detector does not analyse text. It supports images, video and audio only.

Why Google opened it now

The timing tracks a regulatory deadline. The EU AI Act's transparency rules, Article 50, became applicable on August 2, 2026, obliging providers to mark synthetic media in machine-readable form, and Google signed the associated Code of Practice on July 24, per PPC Land.

Scale explains the confidence. The blog post says Google has watermarked over 180 billion images and videos and 240,000 years of audio. Verification already sits inside Search, the Gemini app and Chrome, which the company says handle more than 1 million requests a day.

Google is not alone in the record. The Verge reported the May 2025 test launch, and PPC Land notes Google's May 19, 2026 announcement named ElevenLabs as an adopter alongside OpenAI, Kakao and NVIDIA.

What you need to run it

You need an account. Ars Technica reports the site accepts Google, OpenAI and Apple logins, and that each user gets "a daily usage quota of approximately 10 image, video, and audio checks." The exact figure is not printed on the page, and Google engineers told Ars the cap exists to stop people from tuning tools that strip the watermark. The terms of service ban automated queries, scraping and anything aimed at probing the detection logic, and Google reserves the right to suspend an account for breaking them.

The upload is temporary. The privacy notice says original files are deleted once results come back, but Google keeps a digital signature of each upload, tied to your user ID, for 24 hours to enforce the daily quota.

PPC Land lists 20 supported formats: 11 for images (JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF and GIF), three for video (MP4, MOV and WEBM) and six for audio (WAV, MP3, OGG, FLAC, AAC and M4A). TechCrunch reported the same set. The portal asks for your highest-quality copy, because compression and editing pull detection accuracy down.

The tool answers one question and declines the other. It can confirm that AI touched a file. It cannot certify that a file is real.

To try it, sign in at synthid.com, upload the highest-quality copy of the file you have, and read a blank result as "no mark found," not "not AI."