The AI check reads the fine print
The check that clears new commits now reads image descriptions and poll options, not just the body. Put something ugly in a caption and your post waits for a person.
Attach a screenshot, type a caption, add a two-choice poll, hit post. The bot now reads all of it. It reads your body text, the description under each image, and every option in the poll. Clean, and the commit goes live while you are still reaching for your coffee.
Credit goes to arjankapteijn, who taught the check to gather the rest of your text. It takes the body, up to three image descriptions of 1,000 characters each, and two to four poll options of 50 characters, joins them with blank lines, and sends the lot to OpenAI's moderation endpoint and to the reasoning model that hunts for implied threats. Any category scoring 0.2 or more flags the commit. Sexual content involving minors trips at 0.01. A commit that is only images and a description now gets the reasoning check too, because it finally has text to read.
One catch. A flag does not bounce the commit. Only a moderation score of 0.95 or higher rejects it outright, and the reasoning check adds a category but no score, so the implied threats only it catches will sit in the pending queue until a moderator opens them.