Detection
The generator classifier
A pixel-level classifier that works when metadata was stripped and names the likely generator. Pro triggers it; the cached result is shown to everyone.
Updated
What it is
A classifier that looks at the pixels of an image rather than its metadata and returns a probability that the image is AI-generated, plus per-generator scores: Nano Banana, Imagen, GPT, Flux, Midjourney, Ideogram and others. The card shows it as Generator classifier with the badge Likely when the probability is 0.5 or more, and names the top generator with "classifier" as the basis.
Two providers are supported, Sightengine and Hive. Which runs depends on our configuration; both return the same kind of answer.
Why it exists
Social platforms strip metadata on upload, and Google's SynthID survives but cannot be read by anyone but Google. For a stripped Google, Midjourney or Flux image the classifier is the only detector that can still say anything useful.
Who pays and who sees it
Classifier calls are paid, so they are budgeted:
- Pro accounts trigger up to 600 classifier calls per rolling 24 hours.
- Free accounts trigger none by default, but see the cached result whenever anyone has already triggered it on that image. Popular posts therefore usually carry a classifier row for everyone.
- A global daily budget caps total spend; when it is spent the row says the check was skipped.
Results are cached by content hash and shared, without recording who triggered them.
How it combines
With no provenance signal the classifier probability is the score. If a model recheck also exists the two are blended, 60% classifier and 40% recheck. A trusted Content Credentials signature or watermark match still outranks it.
Accuracy
Classifiers are probabilities. They can miss heavily edited or downscaled generations and can flag stylised photography, renders and illustrations. The card shows the probability so you can weigh it. See Why false positives happen.