IETF grapples with flood of AI-generated Internet-Drafts

IETF participants are confronting a sharp rise in independent Internet-Drafts widely described as AI-generated slop, a volume that is consuming scarce volunteer review time while almost never advancing.

Ross Finlayson opened the exchange by likening the problem to spam. He floated a tongue-in-cheek bond that would be refunded only if a draft were adopted, then a more practical split of the i-d-announce list so readers could ignore non-working-group traffic. Figures compiled by S. Moonesamy show 3,429 independent drafts posted in the year to early September 2026; only 199 (5.8 percent) were later adopted by a working group. In one recent month, working-group announcements totaled 334 while a handful of authors accounted for dozens of independent filings each.

Reviewers report the same pattern: flawless grammar paired with invented terminology, thin logic, and ideas sliced across many long documents. The result feels like a tragedy of the commons. Proposals now on the table include hard rate limits on initial (-00) submissions, AI-detection tooling (commercial detectors such as Pangram already tested by executive director Jay Daley, plus experimental and watermark-based systems), and a web-of-trust model that would give newcomers one free draft then require a shepherd already active in IETF work.

Lars Eggert argued that AI-assisted contributions are permanent and can usefully accelerate standards work when used responsibly. The real risk, he said, is social: experienced participants will simply ignore unknown authors, making the IETF less open to genuine newcomers whose ideas drown in the noise. Carsten Bormann suggested private, per-reader attention scores inside the Datatracker, built from weighted commendations so each participant can surface what matters without a global ranking.

No consensus has formed. The underlying tension is how to defend limited human attention without erecting barriers that contradict the IETF’s open ethos.