Rate LinkedIn posts for AI slop and bait

01 · Role in the system

What Jev does here

No generative model writes a verdict: Jev answers questions like was this written by an AI, does it use em dashes to punch up clauses, does it end on a lesson for everyone, does it show typing residue - plus four about bait: engagement bait, curiosity hook, rage bait. Fixed weights, learned from about 4,900 labeled posts, combine the thirteen AI answers into one score; clicking a badge shows every answer and which ones pushed the post toward AI or human. Cut-offs were fixed on training posts to flag two and five percent of real authors; on a held-out set the author reports 0.3 percent of pre-ChatGPT real authors flagged at the strict band and 67 percent of unseen-vendor AI posts caught (95 percent of plain AI posts, 53 percent of polished rewrites), at about four cents per thousand posts. The four bait questions ride in the same request, so the chip costs no extra call.

02 · Control boundary

Where Jev sits

Each post is one batched request of typed questions; deterministic weighted scoring maps the answers to verdict bands whose cut-offs were fixed on labeled training posts.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Detection numbers are author-reported from the author's own held-out splits, not independently reproduced here.
• Lightly AI-edited human posts mostly pass as human, and formulaic human writers get Uncertain more often than average.
• Post text goes to OpenRouter or TypeSafe under their no-training policies; the optional fact-checker is a separate Ollama pipeline.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• guybrush1984 ↗Community · github · public · checked 2026-09-20