Google has opened SynthID Detector, its verification portal, to everyone worldwide. Upload an image, video or audio file at synthid.com and it scans for a SynthID watermark — an invisible digital signal embedded at the point of creation, which the eye cannot see but detection software can read.
The service is available in English globally, built on partnerships with OpenAI, Nvidia and Kakao, with Apple expected to follow. Google frames it as a tool against deepfakes and AI-driven misinformation.
Read the mechanism before trusting the result
How useful the portal actually is turns on this distinction, and it is not what most coverage will say.
SynthID Detector does not detect whether something was made by AI. It detects whether a SynthID watermark is present. Those are different claims, and the difference runs entirely one way:
- A positive result is strong. If the watermark is there, the file came from a participating tool.
- A negative result means almost nothing. It could be a real photograph. It could equally be AI from a non-participating generator, or a participating one whose watermark was destroyed by a screenshot, a re-encode, a crop or a download site.
Anyone who reads "not detected" as "this is real" has the tool exactly backwards — and that is the reading it will mostly get.
Where this sits among the other attempts
Three approaches to the same problem have now appeared within a week, and they are not competing so much as covering different ground.
OpenAI's statistical watermark on ChatGPT text addresses writing. TikTok's C2PA content credentials attach a signed record that platforms read. Google's SynthID is a signal inside the pixels or the waveform, and the Detector is the public reader for it.
All three share the same weakness, which is that they are opt-in at the point of creation. None of them can say anything about a file made by a tool that chose not to mark it.
What it means in Bangladesh
The portal is free and needs no account, which makes it genuinely usable here — but English-only matters more than it sounds. A Bangla caption on a manipulated image is not the thing being checked; the file is. So the language limit affects the interface, not the capability, and a journalist in Dhaka can use it today on any image regardless of its language.
The realistic use is narrow and worth being clear about. Before a newsroom or a fact-checker publishes "this image is AI-generated", the portal can sometimes confirm it. It can never confirm the opposite. For the far more common Bangladeshi case — a recycled old photograph presented as a new event, or a real photograph with a false caption — SynthID is silent, because nothing about those files is synthetic.
That case is still best handled the old way: a reverse image search and a look at when the picture first appeared.
What a newsroom should do with it
The portal is most useful when it is treated as one input rather than a verdict, and the discipline is simple enough to write on a wall.
A positive result can be published: the file carries a watermark from a named tool, and that is a fact. A negative result cannot be published as anything at all, because the test does not distinguish between a real photograph and a synthetic one that was screenshotted. Writing "checked and found authentic" on the basis of a negative is the single mistake this tool makes easy.
There is also a practical trap. Images arrive in Bangladeshi newsrooms through WhatsApp and Facebook far more often than as original files, and both re-encode on upload. By the time a picture reaches a reporter it has usually passed through at least one process that strips exactly the signal the Detector is looking for — so a negative result on a forwarded image is close to meaningless.
Source: Google's announcement, via Prothom Alo's Techbarta.
