BRENNEN

Industries / Government, education and community / Nonprofits / Standards bodies

Writing standards

This work is how a standard is drafted, worded, reviewed, revised and withdrawn, from the first proposal to the day a committee confirms it is still current.

5

Milestones achieved, from 1969 to 2014.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1969

    Notes anyone could write

    In April 1969 Steve Crocker published "Host Software", the first in a series of notes on the protocols of the ARPANET, the US research network that grew into the internet; on Bill Duvall's suggestion he labeled them "Request for Comments" (RFC). Crocker later wrote that the "basic ground rules were that anyone could say anything and that nothing was official", and that the series stayed open in two ways: "anyone can write one for free and anyone can get them for free." Jon Postel became the first RFC Editor, a role formalized in RFC 902 in July 1984, and by December 2019 more than 8,500 RFCs had been published, from early notes to Internet standards.

    rfc-editor.org

    Credit

    Steve Crocker; Jon Postel; the RFC Editor

  2. 1983

    A model ahead of code

    In 1977 British computer industry representatives proposed an ISO committee for packet-switching networks, and ISO set one up under Charles Bachman, with Hubert Zimmermann among its active members. In May 1983 ISO published ISO 7498, the Basic Reference Model for Open Systems Interconnection (OSI), a seven-layer plan for linking any computers, and in 1988 the US Department of Commerce required OSI in all computers bought by US government agencies after August 1990. By the early 1990s the project had stalled against TCP/IP, the internet's protocols, which were already running, and OSI ended in the mid-1990s. The historian Andrew L. Russell points to two causes: formal rules that let any interested party take part in the design invited "structural tensions", and companies had to buy OSI standards on paper from ISO "one copy at a time", while internet protocols could be implemented free.

    LessonA standard designed in full before anyone had built and run it lost to one that already worked and cost nothing to read.

    spectrum.ieee.org

    Credit

    ISO's OSI committee under Charles Bachman; Andrew L. Russell for the history (IEEE Spectrum, 29 July 2013)

  3. 1996

    Standards shown working first

    RFC 2026, written by Scott Bradner in October 1996, set the IETF's process: a specification climbs three maturity levels, Proposed Standard, Draft Standard and Internet Standard, and can become a Draft Standard only once "at least two independent and interoperable implementations from different code bases have been developed" and have run successfully. In October 2011 RFC 6410, by Russ Housley, Dave Crocker and Eric Burger, cut the ladder to two levels, because "the vast majority of Standards Track documents are published as Proposed Standards and never advance" and "very few specifications have advanced on the maturity ladder in the last decade."

    LessonA ladder that asks for more proof at each step only works if people climb it; most specifications stopped at the first rung, so the IETF removed one.

    rfc-editor.orgrfc-editor.org

    Credit

    Scott Bradner; Russ Housley, Dave Crocker and Eric Burger

  4. 1997

    Words that bind

    A standard separates what a user must do from what is advised or allowed by fixed verbs. RFC 2119, by Scott Bradner in March 1997, defined the key words MUST, MUST NOT, REQUIRED, SHALL, SHOULD, RECOMMENDED and MAY for internet specifications. ISO and IEC set their own rules in the ISO/IEC Directives, Part 2: requirements use "shall" and recommendations "should", and the ninth edition, which replaced that of 2018, limits provisions to the verbal forms in its Tables 3 to 7 and no longer permits negative permissions. In May 2017 RFC 8174, by Barry Leiba, had to clarify RFC 2119, saying that "only UPPERCASE usage of the key words have the defined special meanings."

    LessonTwenty years after the key words were fixed, readers still disagreed about whether a lowercase "must" bound anyone, so the rule had to be restated.

    rfc-editor.orgrfc-editor.orgiso.org

    Credit

    Scott Bradner; Barry Leiba; the ISO Technical Management Board and IEC Standardization Management Board

  5. 2014

    A suspect algorithm withdrawn

    The US National Institute of Standards and Technology (NIST) included a random number generator called Dual_EC_DRBG in its guidance SP 800-90A; such generators make the secret keys that encryption depends on. After news reports in September 2013 raised public concern about it, NIST advised against using it and reopened the document for comment, and on 21 April 2014 it removed the algorithm from the revised draft, "in response to the lack of public confidence", and told users to switch "as quickly as possible." It also began a review of how it writes cryptographic standards, published as a draft process document (NISTIR 7977) and put to its Visiting Committee on Advanced Technology. Outside researchers led by Stephen Checkoway showed at USENIX Security in August 2014 that an attacker who had generated the algorithm's constants could recover encrypted web traffic from some real implementations, though how easily depended on how each library was built.

    LessonA standard is only as trusted as the process behind it; once confidence was lost, NIST had to withdraw the algorithm and open its own drafting process to review.

    nist.govusenix.org

    Credit

    NIST; Stephen Checkoway, Matthew Green, Tanja Lange, Daniel J. Bernstein, Hovav Shacham and co-authors

Watch for

  1. A ban enforced by detectors

    ESTA, a US entertainment technology association that runs a Technical Standards Program, approved a policy on 28 January 2026 that forbids participants to use AI to write any part of a draft standard, a public review comment or the reason for a ballot vote. It says it will use "local computer or web-based AI detection tools" to look for "AI fingerprints", and that a comment found to contain them "shall be irrevocably dismissed" and a vote dismissed, with a right of appeal.

    Achieved whenESTA reports how many comments or votes it has dismissed, which detector it used and that detector's tested error rate on human-written standards text, or an outside test shows that detector separates human from AI ballot text at a stated error rate.

    tsp.esta.org

    Claimed by

    ESTA's Technical Standards Council, January 2026

  2. An AI reviewer for standards

    Researchers built GB/T-Bench, a test of how well AI models review Chinese national standards against drafting rules, from 488 documents with 7,306 planted errors of 25 types, such as wrong normative wording, undefined terms and broken references. The best of 14 models scored 0.3280 on their exact-match measure against 0.6640 for human experts, and the authors say their multi-agent reviewer, GB/T-Reviewer, raised the best score to 0.5094.

    Achieved whena group other than the authors runs the reviewer on standards outside the benchmark, or a standards body uses it on live drafts, and publishes its scores against expert review.

    arxiv.org

    Claimed by

    Tao Wang, Qihao Yang and co-authors, August 2026

Open

  1. 01

    Detectors on ballot comments

    A public test of AI-text detectors on the writing that standards committees actually see: review comments, ballot reasons and mailing-list replies written by people before November 2022, when ChatGPT was released, set against AI-written comments on the same drafts, with each detector's false positive rate (human text flagged as AI) and false negative rate.

    The brief
    Credit · unclaimed
  2. 02

    Review drafts by ISO rules

    An English test set and an AI reviewer for drafts written to the ISO/IEC Directives, Part 2: drafts with planted errors (a requirement written with "must" or "will" instead of "shall", a negative permission, a term used but never defined, a reference to a document that does not exist) and real errors found by experienced editors, scored on whether the reviewer names the right place, rule and error type.

    The brief
    Credit · unclaimed

Open 01 · the brief

Detectors on ballot comments

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of AI-text detectors on the writing that standards committees actually see: review comments, ballot reasons and mailing-list replies written by people before November 2022, when ChatGPT was released, set against AI-written comments on the same drafts, with each detector's false positive rate (human text flagged as AI) and false negative rate.

Why it is still open

ESTA now dismisses votes and comments that its detectors flag, but published detector tests use other kinds of text. A 2023 study by Weixin Liang and co-authors found detectors misclassify writing by non-native English speakers, who are common on international committees; a 2025 test by Brian Jabarian and Alex Imas of the University of Chicago found that, of four detectors, only Pangram kept false positives at or below 0.5% on a corpus spanning many genres; and an August 2026 study of 135,389 professionally edited academic manuscripts found the false positive rates of 13 detectors ranged from 0% to 100%. None tested short, technical standards comments.

tsp.esta.orgarxiv.orgnber.orgarxiv.org

Built with what exists

The IETF's public mail archive, whose main list alone holds 146,794 messages, as a source of dated human comments on drafts; Pangram, GPTZero and Originality.ai, and an open RoBERTa-based detector; current AI models such as Claude, GPT and Gemini to write matched comments; Python for scoring.

mailarchive.ietf.org

  1. First step

    Pull 500 replies to Internet-Drafts from one IETF working group's list for 2015 to 2021 and run two detectors on them.

  2. Done looks like

    At least 2,000 human items from before December 2022 drawn from at least three standards bodies' public records and 2,000 AI-written items on the same drafts; false positive and false negative rates for each detector with 95% confidence intervals, reported separately for items under 50 words and for authors whose affiliations are outside English-speaking countries; item links and code published.

Open 02 · the brief

Review drafts by ISO rules

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An English test set and an AI reviewer for drafts written to the ISO/IEC Directives, Part 2: drafts with planted errors (a requirement written with "must" or "will" instead of "shall", a negative permission, a term used but never defined, a reference to a document that does not exist) and real errors found by experienced editors, scored on whether the reviewer names the right place, rule and error type.

Why it is still open

The only such benchmark found, GB/T-Bench (August 2026), covers Chinese national standards, and even there the best model reached about half the experts' score. The Directives set exact rules, limiting provisions to the verbal forms in their Tables 3 to 7. ISO's own drafting platform, OSD, already runs built-in checks based on the Directives and ISO house style: it flags broken cross-references, misused defined terms and words such as "shall" or "may" in parts of a document where they are not allowed. But its help pages give no figure for how many real errors these rules catch or miss, no English test of automatic checking against the Directives was found, and NIST describes an "integrated development environment" for standards that would highlight errors only as a project idea.

arxiv.orgiso.orghelpdesk-docs.iso.orgnist.gov

Built with what exists

The ISO/IEC Directives, Part 2, free to read on iso.org; drafts that national bodies post free for public comment, such as on BSI's standards development site (they stay under copyright, so the project publishes a script that plants the errors and the error locations, not the texts); an AI model such as Claude for review (see the Note "Seven more flagged" for a Claude-based reviewer of regulated documents); GB/T-Reviewer's multi-agent design as a baseline; Python.

standardsdevelopment.bsigroup.com

  1. First step

    Write the rule list from the Directives' clauses on verbal forms, terms and references, and plant 100 errors in five public-comment drafts.

  2. Done looks like

    60 drafts with at least 1,000 planted errors across 20 error types, plus the real errors two experienced standards editors mark independently; the reviewer's exact-match score, precision and recall for each error type published beside the editors' score and beside a simple rule-based checker of the kind OSD runs, with the reviewer within 0.05 of the editors.