Summary

  • A 31 August IETF list message called a perceived wave of AI-generated independent drafts a spam problem, floated an explicitly unserious US$100 bond and seriously proposed separating announcement traffic.
  • Replies located the scarce resource in human review, described recipient-side filters and warned that relying on established relationships could make worthwhile work from newcomers harder to see.
  • The thread provides no common count of AI-written drafts, review hours, false positives or newcomer outcomes. Its claims are contributions to discussion, not IETF policy or consensus.
  • Submission permission, reviewer attention and IETF standing are different states. A draft may remain open to submit, easy to filter and wholly unendorsed at the same time.
  • Daniel Kade proposes a neutral attention-routing receipt built from procedural facts, with no AI-likelihood score, author rank or automated quality verdict.

The hundred dollars was a joke. The cost was not.

Ross Finlayson’s opening message suggested a US$100 bond, refundable only if an independent Internet-Draft were later adopted by a working group. He immediately said he was not serious and that such a charge would contradict IETF ethos. The operative proposals were to treat the problem like spam and to let subscribers separate working-group announcements from independent ones.

The joke matters because it reveals where an overloaded system is tempted to put the meter. The author presses Submit. Reviewers bear the time needed to understand the proposal, compare prior work, test assumptions and decide whether to respond. A price at submission does not buy any of that labour. It selects for ability or willingness to pay.

Making adoption the refund condition would add a second distortion. Working-group adoption is a procedural decision about taking on work. It is not a retrospective quality certificate for every individual draft, and non-adoption does not prove that a contribution was useless. A financial consequence tied to that decision could place money around a judgment that should remain technical and procedural.

Phillip Hallam-Baker’s reply supplied another function of open individual submission. He uses drafts to place speculative work inside the IETF Note Well and rights framework before discussion. That is one participant’s account, not a legal ruling for every case. It shows why a posting rule can change more than inbox volume.

John Levine’s spam comparison also cautioned that forfeitable bonds are an old, complicated idea in email policy. The thread contains no adopted fee, cost model or evidence that a bond would identify the drafts that consume the most review.

Forty-eight announcements are not forty-eight AI verdicts

Brian E. Carpenter described a practical filter in his reply. He distinguished working-group drafts, author drafts aimed at a working group and other author drafts. His mail rules sent 48 announcements to a junk folder that morning, and he said scanning the subjects took about ten seconds.

That observation measures one inbox action on one morning. It does not say that 48 documents were generated by AI, that they were technically weak or that no one else should read them. It is useful precisely because it describes a low-authority control: a recipient decided what to inspect.

The public evidence set supplies no time series for independent-draft volume, no agreed method for attributing AI use, no count of review hours and no comparison of newcomer outcomes before and after the perceived surge. The debate can be urgent without those numbers. It cannot honestly convert the perception into a measured causal claim.

This boundary protects both sides. Reviewers do not have to pretend their capacity is unlimited. Authors do not have to disprove an unmeasured category before their technical claims can be considered.

Cheap generation makes reputation tempting

Lars Eggert’s intervention names the economic mechanism. Contributions can become cheap to generate while the cost of core review does not fall in the same way. Under that pressure, he expects to disregard unfamiliar contributors more quickly unless someone known engages or the subject is especially relevant.

That response is individually rational. It is also the route by which an open process can narrow without a formal rule. A known author arrives with accumulated trust. A newcomer must first acquire attention in order to build the history that would earn attention. The loop is hard to enter from outside.

Eggert explicitly identifies that downside: genuine new participants with worthwhile ideas may be harder to spot. The risk is not solved by declaring every submission equally valuable. It is solved by separating a person’s right to place work in the public draft space from any promise that a particular reviewer or group will spend time on it.

RFC 3935 makes that distinction possible. Its mission statement pairs an open process with technical competence, quality and relevance. Any interested person may participate and be heard. The IETF still has to choose where limited expertise and review time produce the greatest benefit to the Internet.

An AI score predicts the wrong object

The thread also discussed automation. Jay Daley wrote that his detector experiment had not been used to test Internet-Drafts and that certainty would require cooperation from people using AI. That statement blocks an easy but inaccurate headline: there was no disclosed IETF draft-screening system in this exchange.

Even a perfect account of production method would not settle the attention question. AI assistance can cover spelling, translation, critique, test creation, drafting or near-autonomous production. Rob Sayre’s reply distinguishes helpful use from asking for a broad draft and throwing the result over the fence. The human-review burden depends on the output and the claim, not only the tool label.

A detector also predicts authorship method, not relevance, novelty, protocol correctness or operating value. A human can write a useless draft. An assisted author can identify a real defect. Neither proposition means reviewers should ignore provenance. It means provenance cannot replace evaluation.

The current IETF explanation of Internet-Drafts supplies the status boundary. Anyone may author one; its opinions belong to its authors; it has no formal IETF status unless a later process changes that. Open submission is not endorsement.

Route the request for attention

The new-work guide already describes a landscape in which more than a thousand ideas arrive each year, many go nowhere, and authors seek a suitable working group, a DISPATCH venue or another route. That is a better starting point than deciding which authors look synthetic.

Every announcement could carry a small, neutral attention-routing receipt. It would state the document name and revision, submission time, author-declared topic, intended review venue and requested kind of review. It would distinguish “author says this targets WG X” from “WG X has acknowledged relevance” and from “WG X adopted the work.”

The receipt could then record structural checks without pretending they measure substance. It could note the first substantive human response, a routing change, replacement, expiry, adoption or publication. Recipients could subscribe by topic, venue, current state or requested expertise. Authors could correct a wrong route.

Three fields must remain absent. No AI-likelihood score should become a public stigma. No author-reputation rank should make prior access self-reinforcing. No automated quality grade should suggest that a parser has read the engineering.

This would not compel anyone to review a draft. It would make the request legible. A working group could decline it; a specialist could find it; an independent idea could remain public without borrowing institutional standing.

Sources

  1. Opening message on the perceived draft flood
  2. Brian E. Carpenter on announcement filters
  3. Phillip Hallam-Baker on pay-to-play and Note Well
  4. John Levine on the bond as a spam-filtering idea
  5. Lars Eggert on review cost and newcomer visibility
  6. Jay Daley on the detector experiment and drafts
  7. Rob Sayre on different forms of AI assistance
  8. IETF guide to Internet-Drafts
  9. IETF guide to bringing new work
  10. RFC 3935, A Mission Statement for the IETF
  11. Lu Heng on running-code primacy