On 7 October 2026, Google opened the SynthID Detector to everyone, in English, at synthid.com. The portal checks whether an image, video, or audio file carries a SynthID watermark from Google or partners that have adopted it, including OpenAI, Nvidia, and Kakao. A missing watermark is not proof that a clip is human. The tool is not a general SI detector, and it does not replace a platform label.
What did Google open on 7 October?
Pushmeet Kohli, VP of Science and Strategic Initiatives at Google DeepMind, published We're making it easier to identify AI-generated content globally on 7 October 2026. The post says Google is expanding the SynthID Detector so anyone can use it, "available globally in English starting today," at synthid.com.
This is a wider door, not a new watermark. SynthID has embedded an imperceptible signal in generated images, video, and audio since 2023. Kohli writes that Google has watermarked over 180 billion images and videos, and 240,000 years of audio. Search, the Gemini app, and Chrome already run built-in checks that "now regularly handle over 1 million requests daily." The new fact is the standalone portal, open without a newsroom waitlist.
The earlier portal was not public. On 20 May 2025, the same author announced SynthID Detector and invited journalists, media professionals, and researchers to join a waitlist. The 7 October 2026 post calls that release "an early version" for media professionals, then says access now extends to everyone. A creator does not need a press badge.
That May 2025 announcement also named Gemini, Imagen, Lyria, and Veo as models already carrying SynthID. That is a 2025 list, not a line on today's result. The public detector reports a signal. It does not name the model.
What can the detector actually see?
The live portal says it in one sentence: "This is not a general AI detector." It can only flag media from companies that have adopted SynthID. The 7 October post names Google's own tools and partners "including OpenAI, NVIDIA, Kakao, and soon, Apple." The FAQ on synthid.com, read on 9 October 2026, names the adopters as Google, Nvidia, OpenAI, and Kakao. Apple is "soon" on the blog. It is not in that FAQ sentence. Do not treat a clip from an Apple model as covered.
The detector analyzes three types. For an image, it looks for a watermark in the file. For video, it looks for segments that carry the watermark. For audio, it does the same by segment. Uploads accepted on the portal are JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, and GIF for stills; MP4, MOV, and WEBM for video; WAV, MP3, OGG, FLAC, AAC, and M4A for audio.
SynthID also describes a watermark in text. The detector's list of what it analyzes does not include a text file. Do not upload a caption or a script and expect a verdict.
A result of "watermark detected" means a SynthID signal is present, and the file was likely edited or generated by a SynthID-enabled model. It does not name the prompt, the account, or the product. On 23 September 2026, Google said every clip generated with Omni 1.1 in Google Vids includes an imperceptible SynthID watermark in the frames. That is a product promise about Vids. The portal still answers with the signal, not with the name Google Vids.
What does "not detected" actually mean?
The second result is easy to over-read. "Watermark not detected" means no confident SynthID signal was found, so the file was not likely edited or generated by a SynthID-enabled model. The same FAQ paragraph says this "does not totally rule out AI generation or manipulation."
A miss is the expected answer when the file came from a model that never adopted SynthID, from a partner model before that partner turned the watermark on, or when heavy edits degraded the signal. The page also says that in very rare cases SynthID can falsely trigger on content that was not watermarked. One upload is a clue. The FAQ tells you to collect more than one point of evidence before you decide what the file is.
Google says the mark is built to survive filters, colour shifts, and ordinary lossy saves, and that it is not infallible. Enough transformations can make a detection miss. SynthID is "one important layer in a broader approach to content provenance — not a standalone guarantee."
Upload the highest-quality file you still have. The portal says heavy compression and editing reduce accuracy. A screen recording, a screenshot, or a file that has passed through several apps is a worse specimen than the original export.
How does this land for a creator in Southeast Asia?
The 7 October post says the tool is global and in English. A creator in Jakarta, Surabaya, Manila, Bangkok, or Ho Chi Minh City opens the same URL. Google did not announce a Bahasa Indonesia, Filipino, Thai, or Vietnamese interface. You translate the result for a client. The buttons will not follow the language of the video.
Many edits in this region never touch the named partners. CapCut, Kling, and effects inside TikTok or Instagram are not on the adopter list published on synthid.com. A "not detected" result on that kind of file is what the FAQ describes, even when the picture is obviously generated. Running the portal to clear a CapCut edit before you post it does not clear the edit.
Chat apps recompress. A Gemini still forwarded in a thread is not the file the model wrote. Ask for the original export. The portal wants the highest quality you can get.
Platform labels are a different system. YouTube, Instagram, and TikTok still expect a disclosure when realistic generated media is undeclared. A SynthID hit does not file that disclosure, and a miss does not cancel it. The label rules are a separate change, covered in What actually changed for creators using SI in 2026. When the question is a checkbox in an upload form, read the platform page. The watermark will not tick the box.
Read the privacy notice before you upload a client's face or voice. Google says the file is processed in real time and deleted once the result returns. A digital signature is kept for 24 hours, tied to the user, for quotas and abuse checks. A third-party sign-in shares basic account identifiers with Google. The extra terms forbid probing how to break the watermark and forbid bypassing the page for bulk checks. One export at a time is the described use.
What should you do with a file this week?
If you generated the still, the clip, or the track inside a Google product, export the original and run it through synthid.com before the file enters a chat thread. You are checking that a watermark is present. You are not checking that the picture is allowed to be posted. For a Google Vids clip made with Omni 1.1, Google has already said the watermark is in the frames. A detected result on your own export is a sanity check. A miss on that same original is a reason to look at how you exported it, not a reason to tell a client the clip is human.
If the file came from OpenAI, Nvidia, or Kakao, the same check is in scope, because those names are on the live FAQ. If the file came from anywhere else, including a tool you use every day in Jakarta, do not use the portal as proof. Say which tool you used, in the caption and in the platform's own disclosure field.
If you are reviewing someone else's clip, read "detected" as "a SynthID partner likely touched this file" and "not detected" as "no confident signal." Neither line names the person, the prompt, or the country. Do not post a screenshot of the portal as proof a video is fake or real. After a chat app and a second export, you are measuring compression. The detector is open. It is still a watermark reader with a short partner list, an English interface, and a result you explain yourself.