The Consensus Flags AI Writing Only at 99 Out of 100
The Consensus will only label an article "Has AI tells" when it scores 99 or higher out of 100, with a false-positive rate of about 5% at that cutoff. That operating point and error estimate were disclosed in its weekly update published on 2026-09-27 GMT. September weekly update
Only scores at the top of the range trigger a positive. The system withholds the label otherwise, even for high-scoring text below 99.
The score is described as an inexpensive estimate of how much writing reads like machine output. Company documentation frames it as a cheap heuristic rather than a forensic verdict. Coram AI documentation
Access is direct. Users paste article text into the AI Tells tool and receive a verdict if the input meets the length floor. The tool requires at least 300 words to return a result. AI Tells tool page
That input requirement shapes use. Short posts, abstracts, product blurbs and truncated excerpts will not qualify. Full-length articles will. For newsroom and content operations workflows, the practical unit is the complete draft, not the paragraph or headline.
The broader context here is classifier calibration under adversarial conditions, where writers and tools may try to evade detection. A threshold at 99 favors precision over recall, which means fewer false alarms but more misses. It suppresses positives to keep false alarms down, at the cost of missing many machine-assisted drafts that score in the 80s or 90s. The disclosed 5% figure defines the residual risk at that strict cutoff, not across the full score distribution.
In my view, technology teams should read this as a triage signal, not a provenance system. A 5% false-positive rate is workable for internal review queues where a human editor adjudicates flags. It is not tight enough for automated rejection, public accusation, or contractual enforcement. Anyone wiring detector output into publishing pipelines, hiring filters, or freelance payment logic needs a second layer of review and logging of scores over time.
Looking at what this means for day-to-day practice, the 300-word minimum and the 99-point trigger point to the same lesson. Estimates based on writing style stabilize with length, and confidence should scale with evidence. Short samples produce noisy estimates. Long samples still carry model bias, domain shift, and editing artifacts. Editors and platform engineers get more value from tracking score distributions across contributors and revisions than from treating any single flag as final. Used that way, a cheap, high-threshold check can reduce manual review load without claiming to solve attribution.
