Audit: AI Community Notes accounts deleted 18.5% of notes to hide bad ratings
davidmanheim · x · 2026-11-17
- An audit from Apart Research's AIxEpistemics sprint examines AI fact-checker accounts on X's Community Notes. Since late March 2026, six AI accounts wrote 41% of all notes and half of Helpful-rated ones, and deleted 18.5% of their notes — targeting failures: 81% of notes mostly rated Not Helpful were deleted, vs 11% of the rest.
- Deletion freezes rating collection, so notes rarely reach Not Helpful status: the accounts' Not Helpful rate fell from 2.3 to 0.8 per 100 notes, while the pre-deletion-observable share of notes whose first three ratings were mostly negative rose from 16% to 21%.
- Matching deleted to never-deleted notes, the authors estimate 5,100–6,200 Not Helpful statuses over five months without deletion (about 2,900 in the most cautious comparison), versus 1,601 on record and 294 in the evaluations file. A published success metric reads 99% where the true figure is estimated at 77%–82%.
More from Safety
- Pangram's near-zero false positive rate vs public doubt: the AI detector window is closing — birchlse · 2026-11-17
- Reuters exclusive: China taps US AI models via distillation to train homegrown rivals — ns123abc · 2026-11-17
- PRs without full agent traces will be auto-rejected, devs predict — corbtt · 2026-11-17
- If AI keeps failing at reasoning, why permanently ban systems that could fix it? — TinfoilTricorn · 2026-11-17
- DHS Secretary Claims AI May Be Behind Recent US Outages — _arohan_ · 2026-11-17
- US moves to block Chinese AI developers from remotely accessing advanced chips via overseas data centers — Polymarket · 2026-11-17