On 7 October 2026 Google expanded its SynthID Detector to everyone, globally, in English. Anyone can now upload an image, video or audio file and check whether it was made with AI from Google or its partners — Google names OpenAI, NVIDIA and Kakao, with Apple coming. It is a genuinely useful tool. It also answers a narrower question than most people will assume.
How SynthID works
SynthID is a watermark. When a participating model generates an image, clip or audio track, it embeds an imperceptible signal in the content. The detector looks for that signal. If it is there, you have strong evidence the file came from a participating generator — and Google says the mark is built to survive common edits like resizing and compression.
That is the key phrase: participating generator. A watermark is only present if the tool that made the file chose to add it.
The blind spot: "no watermark" is not "real"
A clean SynthID result tells you one thing: this file does not carry a SynthID mark. It does not tell you the file is authentic. Content with no SynthID watermark includes:
- Open-source models run locally — anyone can generate images or clone voices on their own computer with no watermark.
- Generators outside the partner list, which may use their own marks, C2PA labels, or nothing.
- Face swaps and edits made with tools that never watermark — a real photo with a swapped face or an inpainted detail.
- Files where the mark was degraded by heavy re-encoding, cropping or re-recording a screen.
People committing fraud have every reason to pick tools in this list. That's why a watermark check works best as a positive signal: a hit is meaningful, a miss is not.
What forensic detection does differently
Forensic detection doesn't need the generator's cooperation. It inspects the file itself:
- Provenance — reading and validating any C2PA Content Credentials in the file.
- Metadata and structure — camera data, encoder fingerprints, signs of re-saving.
- Compression analysis — Error Level Analysis to highlight regions that were pasted in or regenerated.
- AI model review — looking for visual, temporal or acoustic artifacts typical of generation.
Forensics is probabilistic: it gives a likelihood and evidence, not a cryptographic yes/no. Watermarks are near-certain when present but silent when absent. They complement each other.
Side by side
| SynthID Detector | Forensic detection (VerifyAI) | |
|---|---|---|
| Needs a watermark in the file | Yes | No |
| Strength of a positive result | Very strong | Probabilistic, with evidence |
| Meaning of a negative result | Only "no SynthID mark" | Low likelihood across several checks |
| Non-partner and local models | Not covered | Checked by artifact analysis |
| Edits to real photos | Only if made by a watermarking tool | Localised by compression heatmap |
| Evidence record for a case file | — | File hash, custody log and certificate |
So which should you use?
Both. If you just want to know whether a picture came from a mainstream AI tool, Google's detector is free and quick — run it. If a decision depends on the answer — an insurance claim, an ID check, a rental listing, a voice note asking for money — a missing watermark isn't enough. You need a check that looks at the file itself and leaves a record you can show someone later.
That is what VerifyAI is built for: provenance, metadata, compression and AI review in one pass, with a SHA-256 fingerprint and a shareable certificate. Try it free above — no sign-up — or see seven signs of AI-generated images to check by eye.
FAQ
If SynthID finds no watermark, is the file real?
No. It only means no SynthID mark was found. Many generators never add one.
Does VerifyAI read SynthID watermarks?
No — SynthID detection is Google's own tool. VerifyAI reads C2PA Content Credentials and runs forensic checks, so the two are worth using together.