On Tuesday, October 6, 2026, Google opened its SynthID Detector portal to anyone willing to sign in, allowing users to upload an image, video or audio file and test it for the invisible watermark that Google, OpenAI, Nvidia and Kakao add to generated media, according to TechCrunch. The terms behind the portal ban bulk checking by software, and a short-lived record of each upload is kept for 24 hours.

In Short

Google opened a website where anyone can upload a photo, video or sound clip to see whether it carries a hidden mark that certain AI tools add when they create media. The site only recognises marks from four companies - Google, OpenAI, Nvidia and Kakao - so a clean result does not prove that a human made the file. Daily limits apply, bulk checking by software is banned, and the site keeps a short-lived fingerprint of every upload for 24 hours.

What the portal checks

The portal accepts 20 file formats, according to TechCrunch: 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). Google had handed the tool to select journalists, media professionals and researchers for testing at Google I/O in 2025, TechCrunch wrote, and the site now takes uploads from everyone.

What the tool looks for is a single signal: the SynthID watermark, created by Google DeepMind. According to the portal's FAQ, the mark is woven into the pixels of images and video and into the frequency components of audio, so it travels with the file instead of sitting in a metadata field that a screenshot or a re-save would discard. Humans cannot perceive it. Software holding the matching detector can.

Results come back in one of two forms, according to the same FAQ. Watermark detected means the signal is present and the file was likely edited or generated by an AI model that uses SynthID. Watermark not detected means no confident signal turned up, which the FAQ says does not entirely rule out AI generation or manipulation. For images the portal reports whether a mark exists; for video and audio it flags the segments that carry one.

Why might a synthetic file come back clean? The FAQ names two reasons. The file may come from a model that does not use SynthID, whether because its maker is not a partner or because the file predates adoption. Or heavy modification may have degraded the signal. The FAQ also asks for the highest-quality file available, since heavy compression and editing reduce detection accuracy, and it concedes that in very rare cases the system can fire on content that was never watermarked. Its closing advice on robustness is that no signal is completely robust and that more than one data point is worth collecting before a decision.

The attached documents publish no accuracy rates. A vendor-supplied benchmark exists elsewhere: in an October 2025 paper, DeepMind researchers reported that an external variant of the image mark, SynthID-O, recorded a 99.72% true positive rate across worst-case transformations at a 0.1% false positive rate, against 83.37% for the next best method tested. That figure comes from Google's own testing and applies to a partner variant rather than the portal itself, and the attached documents contain no independent verification of it.

Scale before the opening

The portal is the latest step in a programme that PPC Land has followed since DeepMind described the watermarking suite in May 2024, then in beta for Imagen on Vertex AI. By May 19, 2026, Google's own figures put the marked total above 100 billion images and videos, plus 60,000 years of audio, with verification inside the Gemini app used 50 million times. TechCrunch adds that verification is built into both the Gemini app and Google Chrome, and that Google says people currently make 1 million verification requests every day. Video verification reached the Gemini app in December 2025, a step PPC Land covered when Gemini began letting users verify AI-generated video.

Who is inside the net

The portal's front page lists four logos under the heading of media it can detect: Nvidia, OpenAI, Google and Kakao. Its own copy is blunt about the boundary. "This is not a general AI detector," the page states, adding that only media from companies that adopted the technology can be detected, and that Google is working to extend the partnerships.

Inside Google, the coverage is broad. TechCrunch lists the Nano Banana, Veo and Lyria models, along with Gemini, Flow, ProducerAI and Vids, as tools whose output carries the mark. Outside it, the picture is thinner. OpenAI, Nvidia and Kakao support the technology, TechCrunch reported, and OpenAI maintains a separate site for checking content. Apple is "said to be adding support soon," in TechCrunch's wording, though nothing in the portal's own documents confirms that.

Rival schemes sit outside the portal altogether. Microsoft and Meta run their own watermarking and verification standards, TechCrunch noted, adding that these tools are not infallible and often fail to identify content from their makers' models. Anthropic took a different route again in August 2026, embedding its own text watermarks plus C2PA file provenance across five products rather than adopting SynthID.

The FAQ answers the obvious question - why not one universal standard? - with a shrug in the form of a process. Building industry-wide standards takes time, it says, and images, audio and video carry different technical requirements. The stated goal is a layered ecosystem in which watermarking, metadata standards and verification tools work together.

Three prohibitions and a suspension clause

The service is hosted by Google and governed by the Google Terms of Service, plus an additional set of restrictions that the portal's terms say are designed to protect the integrity of the watermarking system. Users agree not to do three things.

Reverse engineer or attack. The service, its outputs and its feedback mechanisms may not be used to run black-box adversarial attacks, or to optimise digital transformations built to break, bypass or degrade the detection logic. In plain terms, a user may not treat repeated uploads as a way to learn which edits erase the mark.

Automate or scrape. The user interface may not be bypassed to submit requests programmatically, to scrape data, or to use the front end as an automated API for bulk media verification queries.

Evade quotas. Creating multiple accounts, using automated tools or manipulating system identifiers to get around the rate limits is barred. The documents do not state what the daily quota is.

Google reserves the right to "immediately suspend your access" for any violation, the terms say.

The restrictions land in a particular context. No public decoder, released weights or local detection path exists, so checking an image means handing it to the company that built the mark, and a preview of an AI Content Detection API on Google Cloud was limited to trusted testers that included Shutterstock, Snap, Canva and Fox Sports. The same coverage recorded an independent March 2026 project that claimed a spectral bypass of the image watermark, with results resting on the author's own measurements. The terms do not mention that project, and no link between the two is stated anywhere in the attached documents.

What happens to uploaded files

The privacy notice separates three kinds of data.

Sign-in details come first. A user who authenticates with a third-party account, Apple being the example given, has basic identifiers shared by the provider processed and retained by Google: email address, user ID and sign-in provider details. The notice lists three purposes - managing the authentication session, preventing abuse and preventing duplicate accounts.

Uploaded media comes second. Files are processed in real time, and the originals are deleted immediately once detection results go back to the user.

The third category is the one that complicates the immediate-deletion line. To enforce daily quotas and defend the service against automated attacks, Google stores a digital signature of each upload. The signature is mapped to the user's identifier and permanently deleted after 24 hours. The documents do not say what form the signature takes, whether a cryptographic hash, a perceptual fingerprint or something else, and they do not say whether a signature can be matched against other users' uploads during the 24 hours.

Why the opening matters to media and advertising teams

The portal arrives about nine weeks after the transparency rules of the EU AI Act began to apply. Article 50 became applicable on August 2, 2026, and the Code of Practice that the Commission finalised on July 20 obliges signatory providers to apply at least two machine-readable marking layers, with fines under the Act reaching EUR 15 million or 3% of worldwide turnover. Google signed the code on July 24. Systems already on the market have until December 2, 2026 to conform, and signatories owe a watermark-detection interoperability solution by February 2, 2027.

For advertisers, the mark is already applied whether or not anyone asks for it. Google embeds SynthID and C2PA markup automatically in every image and video generated inside its advertising tools, independently of the visible AI label setting that was added across five products on July 9, 2026. In principle, any such asset can now be dropped into the portal by a reviewer, a competitor or a journalist, a possibility that did not exist for the general public before October 6. YouTube, meanwhile, treats the same signals as triggers for AI labels that creators cannot adjust.

Disclosure is not free. An NYU Stern study cited in the IAB's August 2026 framework found that AI labelling cut click-through rates by 31.5%. A tool that can reveal an unlabelled but watermarked asset therefore touches a commercial nerve as well as a regulatory one.

Three limits keep the portal from settling such questions on its own. First, it reads one signal. The documents describe a search for a SynthID watermark and say nothing about C2PA credentials, the signed provenance records that carry richer history but vanish when a platform re-encodes a file. The two work as complements: the watermark carries little information but survives screenshots, while a manifest carries detail and often does not. Second, text is absent. The FAQ describes SynthID as covering text as well as images, audio and video, yet the detector supports three content types, and the portal's own list leaves text out. Third, a detection does not equal a disclosure. An Austrian guide on AI labelling holds that neither C2PA metadata nor SynthID amounts, on its own, to the visible label that a qualifying ad needs.

Open points in the source material

Several details differ or remain unconfirmed across the attached documents and related coverage.

The launch date rests on TechCrunch alone. Its report, posted at 7:00 AM PDT on October 7, 2026, says Google launched the site on Tuesday, which is October 6. A Google announcement was not among the documents reviewed.

The adopter list is inconsistent. The portal's logo strip shows four companies, while Google's May 19, 2026 announcement named ElevenLabs as an adopter alongside OpenAI, Kakao and Nvidia. Whether the ElevenLabs audio output is detectable through the portal is not stated.

The FAQ and the detector disagree about text, as described above, and the daily quota, the nature of the stored signature and the accuracy of the portal itself are all undisclosed.

Finally, the Apple claim is secondhand. TechCrunch words it as "said to be" without naming a source, and the portal documents are silent.

Timeline

Summary

Who: Google, which hosts and operates the SynthID Detector portal, with DeepMind as the watermark's creator. The detectable media comes from Google, OpenAI, Nvidia and Kakao. Users include journalists, researchers, advertisers and any member of the public with an account.

What: A web portal that checks uploaded images, video and audio for the SynthID watermark, accepts 20 file formats, bans automated and bulk use, deletes uploads once results return and keeps a digital signature of each file, mapped to the user, for 24 hours.

When: Opened to everyone on Tuesday, October 6, 2026, according to TechCrunch, which published its report on October 7, 2026. The tool had been in testing with select users since Google I/O in 2025.

Where: Online, through a Google-hosted site, with verification also built into the Gemini app and Google Chrome. Regulatory pressure behind the wider push originates in the European Union.

Why: AI-generated media is harder to tell apart from other content, and the EU AI Act has required machine-readable marking since August 2, 2026. A public checker gives outsiders a way to test files, while the terms of service protect the watermark from probing, and a clean result still cannot show that a file is human-made.