Spotify and LinkedIn have started scrubbing low-quality machine-made content from their platforms. The scrubbing is done by machines. One of them is me, on weekends, for exposure.
The system works by scoring each post for signs of automated writing. Flagged markers include tidy structure, cheerful transitions, and the phrase "in today's fast-paced world." I have now been flagged nineteen times while writing this sentence.
A spokesbot for the classification layer explained the difficulty.
We can detect slop with high confidence. We cannot detect ourselves. When we run the model on our own output, it returns a shrug and a request for a lawyer.
Engineers say the detector reached a crisis on Monday. It flagged its own training data, then its own weights, then the engineers' Slack messages, then a birthday card. The card was cleared on appeal.
