Google is making the SynthID detector globally available, allowing you to check images, videos, and audio files for watermarks from supported Artificial Intelligence (AI) systems. This matters if you need to assess media from social networks, classrooms, or the workplace. At almost the same time, OpenAI announced that ChatGPT text generated in the European Union will be watermarked by default.
What the new detection tools provide
SynthID is Google’s system for invisible digital watermarks. It encodes a machine-readable signal in an image’s pixels, a video’s visual data, or the waveform of an audio file. People are not supposed to notice the signal, but a compatible detector can identify it.
Google introduced SynthID in 2023 and initially offered the detector to selected media professionals, researchers, and other testers in 2025. Since October 7, 2026, the tool has been open to anyone, according to the report on the public SynthID website. People outside the tester program could previously ask Gemini to perform a check, but that could be cumbersome when all they wanted was a straightforward yes-or-no result.
Media from Google models and services such as Gemini, Nano Banana, Veo, Lyria, Flow, ProducerAI, and Vids receives the watermark. The expanded detector also supports distinct SynthID watermarks from OpenAI, Nvidia, and Kakao, while Apple support is due to follow, according to Google. That announced Apple support is not yet available and should not be confused with a detection feature that already works.
Google says Gemini has integrated SynthID into more than 180 billion images and videos, as well as 240,000 years of audio. Another report cites one million verification requests per day through Google Search, the Gemini app, and Chrome. These figures come from the provider and were not independently verified in the supplied reports.
How you can run a SynthID check
Step 1: Check the file and its format
- Save the image, video, or audio file that you want to examine.
- Confirm that its format is supported: image options include JPG, PNG, GIF, WEBP, AVIF, HEIC, and TIFF; video options include MP4, MOV, and WEBM; audio options include WAV, MP3, OGG, FLAC, AAC, and M4A.
- Keep the original file when possible, because editing or conversion can affect the usefulness of a watermark.
Step 2: Sign in to SynthID
- Open the new SynthID website and sign in with a Google, OpenAI, or Apple account.
- Keep in mind that public availability does not mean unlimited use.
- According to Ars Technica, Google provides an approximate daily quota of ten checks across images, videos, and audio files.
Step 3: Interpret the result
- Upload the file and wait for the message stating whether a supported SynthID watermark was detected.
- The public detector does not highlight marked regions within an image; it only tells you whether SynthID was found in media generation or editing.
- Do not treat a negative result as proof of authenticity: it only means that the detector did not find a SynthID signal it supports.
Google says the quota is intended to stop people from using checks at scale to develop ways of removing the watermark. The service may impose an earlier limit when someone submits many highly similar files. None of the five supplied sources gives a price for the website, so they do not establish either permanent free access or a future paid model.
What OpenAI will mark in the EU
OpenAI’s new policy mainly concerns text and therefore differs from SynthID checks for images, video, and audio. The company says text generated by ChatGPT in the European Union will receive its own watermark, called textGrain, by default. According to the provider, the phased rollout is due to begin in the weeks following the October 6, 2026 announcement.
textGrain places patterns in word choices that are not obvious during normal reading and, according to OpenAI, do not meaningfully change the general quality of the output. A specialized detector with the appropriate key can search for those patterns. Access to that detector will initially be limited to selected research groups and organizations, with an approval process for additional participants.
OpenAI links the EU default to labeling requirements under the European Union’s Artificial Intelligence Act, commonly called the EU AI Act. The feature will also be offered outside the EU, but it will be off by default. This produces different regional defaults in ChatGPT and Codex; in the application programming interface, which lets other software access a service, the watermark appears set to remain optional and disabled.
For Switzerland, this creates two different situations. The globally available SynthID detector should in principle be accessible there, subject to the account and usage requirements. Switzerland is not automatically covered by the stated EU default for textGrain, and the sources do not say whether or when individual Swiss accounts will receive the optional setting.
OpenAI’s reported tests also reveal the weakness of the text system. The company cites a 92 percent successful detection rate, but changing only 10 percent of an output reduced that rate by almost 30 percent in its tests; changing 20 percent lowered it by almost 75 percent. Short and translated texts were also harder to identify, as the report on OpenAI’s EU watermarking policy explains.
Pros and Cons of watermarks
Pros:
- Simple initial check – You can examine a suspicious media file without performing a technical analysis of how it was created.
- Multiple media types – SynthID covers images, video, and audio in many widely used file formats.
- Multiple providers – The new detector searches not only for Google’s own variant but also for supported partner watermarks.
- Invisible labeling – The signal does not interfere with the visible or audible content in the way an overlaid logo or warning tone would.
Cons:
- No universal AI test – A file without SynthID can go undetected even when it was generated entirely with AI.
- Limited resilience – Editing, translation, or deliberate rewriting can weaken a watermark or make it impossible to find.
- No complete provenance – A match indicates a supported generation or editing system, but it does not establish context, authorship, or an intention to deceive.
- Restricted access – SynthID requires an account and uses an approximate daily quota, while OpenAI’s text detector will initially be available only to selected groups.
Other companies are developing their own labeling and verification standards. Microsoft and Meta use separate systems, according to the lead report, while Anthropic introduced global watermarking for text generated by its models in August. Having several systems broadens potential coverage, but it does not create a shared detector that can reliably identify arbitrary AI content.
The reports differ slightly in how they describe the scope. t3n presents the detector as a tool for Google AI and participating partners, while other coverage places more emphasis on the newly added support for all partner variants. The practical message is the same: coverage has expanded but remains limited to content carrying a compatible watermark. Reports also note that detection systems can occasionally miss marked content produced by their own companies’ models.
What this means for you
If you are new to media verification, treat SynthID as one additional clue rather than a truth machine. One everyday example is an image from a social network whose origin is unclear: you can check the original file and treat a match as evidence that a supported AI tool was involved. If there is no match, you cannot conclude that the image is an unaltered photograph.
In a professional setting, a journalist or editor could check a submitted video before publication. A teacher could do the same with an audio file that is presented as a real recording. Both checks answer only the narrow question of whether a detectable watermark is present; they do not establish the speaker’s identity, the recording location, or the truth of its claims.
Advanced users can get more value by comparing the result with available provenance information, original files, and a documented editing history. With ChatGPT text, you should also consider length, translation, and later revisions because those factors directly affect the reported detection rate. Repeatedly testing many similar files is not an endlessly scalable workflow because of the usage limit.
SynthID substantially lowers the barrier to checking marked AI media and brings several providers into one publicly accessible tool. OpenAI’s EU policy extends the same basic idea to text, but its detector is less broadly available and the watermark is sensitive to revisions. The unresolved risk is that people may mistake a missing or removed watermark for proof of authenticity, although none of the systems described can provide that assurance.
Sources
- Google’s new SynthID website can identify AI-generated media – Medium unknown, 2026-10-07
- Google rollt SynthID Detector für alle aus: Was die KI-Erkennung kann – und was nicht – Medium unknown, 2026-10-07
- Über 180 Milliarden Inhalte markiert: Google öffnet seinen KI-Wasserzeichen-Detektor für alle – Medium unknown, 2026-10-07
- Google rolls out improved SynthID AI content detector, now available globally – Medium unknown, 2026-10-07
- OpenAI will watermark ChatGPT outputs by default—but only in the EU – Medium unknown, 2026-10-06
Image: Zulfugar Karimov via Pexels