en
РусскийHow to read the result
What each Synth ID Detector verdict means, how much to trust an estimate, what the bot cannot see and which official checkers to use for a manual check.
The bot's answer has up to four parts: a verdict, the generator, the signs found in the file and a note about what the check cannot prove. This page explains each.
The verdicts
| Verdict | What it means |
|---|---|
| AI-generated | The file carries a label of AI generation: a trainedAlgorithmicMedia marker, a named AI generator or Stable Diffusion and ComfyUI parameters. |
| AI-edited | A real photo was extended or changed by AI: the file has a compositeWithTrainedAlgorithmicMedia or compositeSynthetic marker. |
| No AI labels, but Content Credentials | The file has a C2PA creation and edit history, and it does not name AI. |
| No AI labels found | The bot found no labels. This is not a proof that the file is real. |
| Likely AI-generated (image) | No labels, but a neural network sees signs of the SynthID watermark. An estimate. |
| Likely an AI-generated voice | No labels, but a neural network hears signs of synthesized speech. An estimate. |
Labels against estimates
A label is something the generator wrote into the file itself. When it is there, the verdict is firm: the file says it is AI.
An estimate comes from a neural network, and the bot shows it only at 90 percent confidence or more and only when there are no labels. It is a probability, not a fact, and every such answer carries a note that explains the limits. For the SynthID watermark, the bot's own test found about 9 in 10 Gemini images, and the network occasionally fires on other generators. For a voice it finds about half of synthesized speech and occasionally fires on a real voice.
What the bot cannot see
- An unlabeled image from a generator without SynthID: Midjourney, Flux, Qwen, Stable Diffusion and others, after the labels are removed.
- The picture of a video: only its soundtrack and file labels are checked.
- Unlabeled music, and singing: the voice check is made for speech.
- Any file whose labels were removed by a screenshot, recompression, a messenger or sending a photo with compression.
Check by hand
When a result matters, confirm it with the official checkers. The bot links them under an answer that has no labels:
- SynthID Detector by Google, which also covers OpenAI, NVIDIA and Kakao watermarks.
- OpenAI Verify for ChatGPT images.
- The AI Speech Classifier by ElevenLabs for ElevenLabs voices.
The bot does not call these services for you: open them yourself.
Questions
The bot says AI-generated. Can it be wrong?
A label is written by the generator, so a wrong label is unlikely, but a label says how the file was made, not whether a person edited it afterwards. An estimate can be wrong either way, which is why it is called likely.
The bot found nothing, but I am sure it is AI. Why?
The file may have lost its labels, or the generator does not write any. Trust your own checks too, and use the official checkers.
Does the bot store the verdict?
The bot keeps a short record of each check: the file name, type and size, and the result with the signs found. See the privacy policy.