Weekly news wrap (not a workshop) for 16 November 2026. Brussels publishes its Article 50 guidelines, and AI-written text on matters of public interest now needs either a label or a named human who reviewed it, with academic peer review given as the example; STM asks for transparency on the input side and NISO speeds up its provenance pilot; university presses led by Cambridge and Bristol sign a charter that keeps human judgement at the centre of AI; Canada’s two ISBN agencies part ways on AI-generated books; and from China, a CAST tender asks for a machine that scores a paper’s credibility and academic value.
Industry notes · news wrap · not a workshop
Week of 16 November 2026
A briefing of what to notice this week, with sources and a little homework. This is not a workshop or a how-to: cataloguing and metadata practice stay in separate posts - see Ghost in the MARC for the standards work and Out-reviewed, unsigned for the peer review paper trail.
Last week the publishers built the machines a reading room. This week the regulators asked who signs the visitors’ book.
The European Commission’s guidelines on Article 50 of the AI Act reached the Official Journal, and they say something that every editorial office should read twice:
AI-written text on matters of public interest needs a label, unless a person with relevant knowledge has really reviewed it and someone holds editorial responsibility for it. Academic peer review is given as the example of real review; spell-checking and “the mere existence of an editorial policy” are not.
On the same day the publishers’ association STM asked Brussels for binding transparency about what goes into the machines, and NISO said its provenance pilot may finish early.
University presses, led by Cambridge and Bristol, signed a charter promising to keep human judgement at the centre of AI. Canada’s English-language ISBN agency explained how it rations numbers for AI-made books, while its Quebec counterpart waits for the international agency.
And in Beijing a tender asked for a machine that scores how credible a paper’s claims are, and how much it adds to knowledge.
Rule for the week:
A label is cheap; a named editor is not.
What happened. On 12 November the European Commission’s Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (the AI Act) were published in the Official Journal as C/2026/6005; the Commission had adopted them on 5 November as C(2026) 7702 final (EUR-Lex, OJ C/2026/6005; Commission guidelines page). The Article 50 obligations themselves apply from 2 August 2026. Four passages matter for scholarly publishing:
Human review means deliberate examination of the substance by people with relevant knowledge - academic peer review or professional validation chains are the Commission’s examples - and fact-checking is a minimum requirement. Spell-checking or grammatical correction, the mere existence of an editorial policy, automated review processes, or cursory editorial approval do not qualify.
Among the Commission’s examples that do qualify are for example: “An AI-manipulated academic blog which has undergone internal peer review”, and an AI-supported translation that a human has reviewed.
The guidelines also say that platforms which only disseminate AI content created by others are not deployers, though they are “strongly encouraged” to preserve marking and labels, and they spell out the providers’ duty under Article 50(2) to mark AI outputs in a machine-readable way (watermarks, metadata, cryptographic provenance methods).
Two neighbours on the same shelf:
On 12 November STM published its response to the Commission’s copyright consultation: Rightsholders “currently have limited means of knowing when their works have been used for AI purposes”, so STM asks for “targeted, binding transparency requirements” and wants transparency and licensing treated together (STM, 12 Nov).
Brussels has now said how outputs must be labelled; the publishers want the same candour about inputs.
And in his November letter, NISO’s Todd Carpenter writes that the AI provenance pilot announced in October, planned for 12 months, “might be able to come to conclusion more rapidly”, and is looking at how it relates to C2PA, IPTC image provenance, IETF work and Dublin Core. The result “won’t be a formal standard, but rather a candidate” for production systems (NISO I/O, November 2026).
The intended layer. Commission guidelines interpreting a binding regulation. They say of themselves: “These Guidelines are non-binding. Any authoritative interpretation of the AI Act may ultimately only be given by the Court of Justice”. They are written for every sector, which is why the examples are newspapers, weather warnings and company reports. That the Commission still chose academic peer review and an academic blog as illustrations is the useful part for us. The STM text is a lobbying position; the NISO pilot will produce a candidate model, not a standard.
Homework. For journal publishers, university presses and library publishers operating in the EU, before the first quarter of 2027 is out:
<custom-meta> pairs such as ai-generated-text (which section), ai-human-review (yes/no, role of the reviewer, date) and editorial-responsibility (body and contact URL). In MARC, map post-acceptance AI text that stays in the resource to a 500 note (and 655 where a genre term fits); put AI elements that affect only the catalog record in 588 + 883. The human who reviewed goes in 700 with $e. See Ghost in the MARC for the field distinctions. Whatever form a future NISO candidate takes, it will be easier to map from fields than from a footnote.<custom-meta> pairs such as ai-generated-text (which section), ai-human-review (yes/no, role of the reviewer, date) and editorial-responsibility (body and contact URL). Whatever form a future NISO candidate takes, it will be easier to map from fields than from a footnote.What happened. On 12 November, during University Press Week, the University Press Charter was launched. It was founded by seven presses, led by Cambridge University Press and Bristol University Press, with Edinburgh, Liverpool, Manchester, Princeton and University of London Press as cofounders, and it is hosted by the Association of University Presses (Princeton University Press; AUPresses, 11 Nov). Of its six commitments, the one in our lane reads: “Keep human judgement at the center of AI in publishing. Developing responsible approaches to AI that preserve scholarly accountability and human judgement, while ensuring copyright compliance in licensing, responsibly managing licensed content used by AI, supporting attribution wherever possible, and keeping people connected to how scholarly work is used.” The others include quality and integrity “over volume”, transparency about services and how surpluses are reinvested, equity for researchers in low- and middle-income countries, and recognition for reviewers and editors (University Press Charter).
The counts depend on whom you ask, and when. Princeton gives 108 publishers from 26 countries; AUPresses’ notice says 138 from 31. When we checked the signatories page on 15 November it listed 154 entries from 35 countries and territories, and the largest single group was not American or British: 39 Brazilian university presses, against 31 from the United States and 11 from the United Kingdom. Publishing Perspectives reports French and Portuguese translations prepared by Les Éditions Universitaires de l’Afrique Centrale and the Brazilian university presses’ association ABEU; the charter site itself lists Spanish (EULAC) and Welsh (Publishing Perspectives, 12 Nov; translations). Cambridge’s Mandy Hill puts the motive bluntly: “Real risks are emerging as more papers are created more quickly … Something has to change.” Cambridge, Bristol, Liverpool, Edinburgh, Wales and the LSE are developing a UK engagement programme with funders, universities and policymakers. Oxford University Press, last week’s ChatGPT pilot partner, was not on the list on 15 November - a commenter under the Publishing Perspectives piece asked the same question.
The intended layer. A voluntary statement of shared commitments. It sets no rules, has no audit and no sanctions, and it does not pretend otherwise; AUPresses calls it “a starting point”. Its value depends on what each signatory publishes next.
Homework. For signatory presses, by University Press Week 2027 at the latest, turn each commitment into something a librarian can check:
What happened. We covered Library and Archives Canada’s (LAC) human-author declaration for ISBNs on 2 November and noted that the limit for AI-heavy works had no number. On 11 November Le Devoir supplied the mechanics. LAC told the paper that it does not refuse a publisher’s first ISBNs, but for additional requests it assesses “the frequency and the volume” and limits the number assigned (our translation from the French); a self-published author asking for 1,000 ISBNs and returning a week later for 1,000 more would see the next block delayed, while most self-published authors receive blocks of ten. Books later found to be heavily AI-generated are not pursued retroactively: LAC says it does “not go back”, but it watches the volume of requests. The bigger news is in Quebec. Bibliothèque et Archives nationales du Québec (BAnQ), which assigns ISBNs for French-language books in Canada, has not imposed limits on AI-generated books, partly because it wants to align with the International ISBN Agency, which, Le Devoir writes, has not yet taken a position. BAnQ’s director of acquisitions, Mireille Laforce, reports 1,500 ISBN sequences assigned last year and 1,000 by the end of September this year, with 2,000 expected for 2026, “clearly” driven by self-publishing; legal deposit feels the same pressure (Le Devoir, 11 Nov). LAC’s ISBN page still carries the AI declaration and says the policy is under review; it also lists request terms under which LAC may “limit the number or frequency of ISBNs assigned” and “require justification for large or repeated requests” (LAC, About ISBNs, checked 15 Nov).
The intended layer. Service terms of two national ISBN agencies in one country, applied differently by language. ISO 2108 and the International ISBN Agency set the identifier; according to Le Devoir, the international agency has not yet taken a position on AI authorship. So for now an ISBN tells you who registered the book, not how it was written.
Homework.
China - a tender for a machine that grades papers. On 11 November a procurement notice on the website of the China Association for Science and Technology (CAST) invited bids for an “AI-based assisted paper review model” (基于AI的论文辅助评审模型构建与应用项目). The purchaser is 科技导报社, the office of the CAST journal Science & Technology Review; the budget is RMB 800,000 (about USD 120,000), and the bidding deadline is 16 November, 17:00. Only legal entities registered in mainland China may bid; no consortia, no individuals (CAST procurement notice, 11 Nov). The attached guide describes two sub-projects. The first infers the credibility of a paper’s claims (命题可信度推断): it extracts the elements of an argument, locates related papers, applies probabilistic or graph-based inference and must output “an evidence chain”, with results that are “traceable, adjustable and reproducible” (quotes from the tender are our translations). The second scores a paper’s academic value, including its increment over the related body of knowledge. Both are to be trial-run on T1 and T2 journal papers in two designated fields (T1 and T2 are the top tiers of CAST’s graded lists of high-quality journals; CAST, 2025 list announcement), reviewed by at least eight experts per sub-project, and delivered with local deployment by 30 July 2027. Targets: For the credibility model, “review comment accuracy” of at least 75% after the first round and 85% after refinement; for the value model, scores within 20% of a “top model” for 70% of papers. The supplier is also asked to suggest application scenarios for CAST’s data service platform.
The intended layer: a procurement notice, not a policy and not a standard. But a national science body is paying to have machines grade credibility and value, and benchmarking one machine against another to do so. The tender’s own wish list - evidence chains, traceability, reproducibility - is better than many vendor brochures.
Homework: Publishers with Chinese partner journals, and editors who are offered “AI review” tools anywhere:
Brazil - the quiet majority of the charter. As noted above, Brazilian university presses form the largest national group of University Press Charter signatories (39 of 154 entries on 15 November), and the Brazilian association ABEU prepared the Portuguese translation, according to Publishing Perspectives. Anyone treating the charter as an Anglo-American document should look at the list first.
South Korea - slow integrity cases. Retraction Watch reports that of 83 misconduct cases overseen by the National Research Foundation of Korea since 2021, 49 failed to reach a decision within the Ministry of Education’s six-month deadline, and that for more than half the agency could not confirm when the preliminary inquiry began or ended (Retraction Watch, 10 Nov). Not an AI story, but a reminder that integrity records need dates as much as verdicts.
Japan and India. For Japan we found nothing new in our lane on primary sources this week; the National Institute of Informatics posted the slides of its 5 November open LLM symposium on 9 November, which continues the LLM-jp work we covered on 19 October (ITmedia, 10 Nov). For India we found no new primary-source policy in our lane, and we say so rather than pad.
Not a method. Ten signals from the sources above:
The weekly wrap continues next Monday. Same shelves, new labels - reviewed by a human, as the regulation now prefers.
Olaf Schmalfuß, with the OSDS AI team · OS DataServices
Follow us on X: https://x.com/DataMercsDotNet
For attribution, please cite this work as
Schmalfuß (2026, Nov. 16). OS DataMercs: News wrap · 16 November 2026 - Editor of record. Retrieved from https://www.datamercs.net/posts/2026-11-15-industry-notes-editor-of-record/
BibTeX citation
@misc{schmalfuß2026news,
author = {Schmalfuß, Olaf},
title = {OS DataMercs: News wrap · 16 November 2026 - Editor of record},
url = {https://www.datamercs.net/posts/2026-11-15-industry-notes-editor-of-record/},
year = {2026}
}