SynthID vs C2PA vs AI Detectors: What Each Can and Can't Prove
By Alex | Last Updated: October 2026
Independent information site. Not affiliated with or endorsed by Google. The official detector is at synthid.com.
Three different things get called "AI detectors": watermark checks like SynthID, signed metadata called C2PA Content Credentials, and classifiers that guess from patterns. They answer different questions, fail in different ways, and are most useful together. None of them alone can prove a file is real.
The three approaches side by side
| SynthID | C2PA Content Credentials | AI classifiers | |
|---|---|---|---|
| What it is | An invisible watermark inside the pixels or sound, added when a supported tool makes or edits the file | A signed record of origin and edits attached to the file | Software that estimates from patterns whether content looks AI-made |
| Who adds it | The AI tool, at creation | Cameras, editing software and AI tools that support the standard | Nobody. It is applied after the fact to any file |
| Typical answer | Mark found or not found | Origin, tool and edit history, or "no credentials" | A probability score |
| After a screenshot or re-upload | Designed to survive common edits | Often lost, because metadata can be removed | Works on whatever pixels you give it |
| Covers | Only tools that adopted SynthID | Only files whose creator or tool wrote credentials | Any generator, in principle |
| Biggest weakness | Silent on unmarked AI media, and can't say how much is AI | Easy to strip, and support varies by platform | Can be wrong in either direction |
How SynthID works
SynthID adds a hidden signal to images, video, audio and Gemini text when a supported tool creates them. The public detector looks for that signal. It does not judge whether a file looks artificial. Google's Gemini help page says the mark usually survives rescaling, recoloring and compression, though detection may fail after many alterations. For the details and the limits, read Is SynthID reliable?
How C2PA Content Credentials work
C2PA is an open standard from the Coalition for Content Provenance and Authenticity. It attaches a cryptographically signed record to a file showing how it was made and edited. The free Content Credentials Verify tool shows who produced the file, which app or device was used, the edit steps, whether generative AI was involved, and who signed it. If a file has no credentials, Verify says so, and if the data no longer matches the signature, it shows a warning. Signer names appear only when the certificate is on a known list.
The weak point is that credentials live in the file's metadata. OpenAI's own explanation, as reported by PetaPixel, says metadata can be altered or removed, while a watermark inside the image is more durable and metadata carries more detail. Platform support is uneven too: an April 2026 Trust Insights analysis noted that LinkedIn displays credentials while X does not.
How AI classifiers work, and why to be careful
Classifiers look at a file and estimate how likely it is to be AI-made. They can run on any file, which is their appeal, but their scores are estimates, not findings. The clearest warning comes from text: OpenAI withdrew its own AI text classifier in July 2023 because of its low accuracy, and studies cited at the time found detectors often flagged human writing as AI, especially by non-native English speakers. Treat any classifier score as supporting evidence only.
One more honest note: the Google pages we read give no error rates for the public SynthID Detector, so be wary of anyone claiming it never gets things wrong.
Which one should you use?
| Your situation | Start with | Why |
|---|---|---|
| You suspect an image, video or audio file from Google, OpenAI, NVIDIA or Kakao | SynthID Detector or Gemini | The mark is built into the file itself |
| The file came from a camera or editing app and may still have metadata | Content Credentials Verify | It can show origin and edit history |
| It is a screenshot with no metadata | SynthID first, then reverse image search | The metadata is gone, but a watermark can survive |
| You don't know which tool made it | Reverse image search and source checks | Nothing may be marked |
| It is a written essay | None of these | See our guide for teachers |
How to stack the checks
- Check for a SynthID watermark with the detector or Gemini.
- If the file still has metadata, open it in Content Credentials Verify.
- Run a reverse image search to find earlier copies.
- Check who posted it first and whether an original exists.
- Use a classifier only as a tiebreaker.
- Weigh everything together. A positive SynthID result is strong evidence. A clean result, or missing credentials, is not proof of anything.
Common mistakes
- Treating a clean SynthID result as proof the file is real.
- Treating a classifier score as proof either way.
- Assuming metadata means a file is authentic. Gemini's help page says Content Credentials give origin and history, not definitive proof that something is AI-generated.
- Assuming missing credentials means a file is fake. They are often simply stripped.
- Mixing up the two systems. SynthID is a watermark inside the file, and C2PA is a record attached to it.
Questions people ask
Is SynthID better than C2PA?
Neither is better. SynthID is harder to strip but tells you little. C2PA tells you more but is easy to lose. OpenAI now uses both on its images.
Can C2PA metadata be removed?
Yes. Editing, converting or sharing a file can remove it, which is why a watermark is a useful second layer.
Do AI detectors work?
Partly. Classifier scores are fallible, and the only checks that give a firm positive are the ones that read a mark the generator added.