Detection
The model recheck, explained
A vision-language model on Cloudflare Workers AI answers a fixed rubric about a post and returns a probability with notes. Pro triggers it on request; the answer is cached for everyone.
Updated
What runs
For images, a vision-language model is shown the image and asked for a structured answer: a probability that it is AI-generated and notes about hands, text, textures, lighting artefacts, a stock-photo look and uncanny symmetry. For text, a language model applies a rubric to the writing. Both run on Cloudflare Workers AI; the answer is parsed defensively and shown in the card as Model recheck with the badge Likely at 0.5 or above.
When it runs
Only when a Pro user chooses Recheck with model in the card. It never runs automatically. Pro accounts get 300 rechecks per rolling 24 hours.
How it combines
- Recheck and classifier both present: 60% classifier, 40% recheck.
- Recheck alone: the recheck probability is the score.
- Any trusted provenance signal outranks both.
Shared result
The recheck is written to the verdict cache against the content hash and returned to anyone who later meets the same content, on any tier, through the Community row. The cache never records who asked. That is why the free tier often shows recheck rows on widely shared posts. See Community reports and the verdict cache.
What it is not
A verdict. Models describe what they see and are sometimes confidently wrong. The notes are shown so you can decide whether the reasoning holds up.