Weekly news wrap (not a workshop) for 19 October 2026. In Dublin, the European Commission revises its research guidance on generative AI, now with a chapter on AI agents and peer review, and collects first pledges from Elsevier, Frontiers and others; EDItEUR’s ONIX codelists gain a content audience for LLM training; a Scholarly Kitchen guest post asks for evidence before more adoption; and in Japan, NII documents how its open LLM-jp-4 models were built, with a rights level for every dataset, while the Humanitext project turns AI suggestions into expert-verified, sourced links.
Industry notes · news wrap · not a workshop
Week of 19 October 2026
A briefing of what to notice this week, with sources. How-to and cataloguing practice stay in separate posts — see Ghost in the MARC for the standards work.
Last week the trade asked whether an AI answer can say where it came from. This week the question moved one layer down, to the label on the package itself.
In Dublin, the European Commission revised its guidance on generative AI in research, now with explicit rules of thumb for AI agents and peer review, and collected the first private pledges for AI in European science.
EDItEUR’s new ONIX codelists give publishers a way to say, in metadata, “this text is meant for machines”.
And in Japan, the National Institute of Informatics published a training report in which every dataset carries a rights level.
The courier rule still applies: whatever is not written on the label will be guessed at the other end.
What happened. On 15 October, at the European AI in Science Summit in Dublin, the Commission launched the RAISE pledging initiative for AI in Science Enablers, which collects in-kind support for European researchers from industry and philanthropy. First commitments came from the Novo Nordisk Foundation, Nvidia, Elsevier, Sanofi, Frontiers, Mistral and Google DeepMind. Pledges are to show support “beyond their direct commercial interests” and will be reviewed annually (Commission press release IP/26/2175, 15 Oct; RAISE pledging page).
Two of the first pledgers are publishers:
At the same summit, the Commission presented the fourth version of the ERA Living Guidelines on the responsible use of generative AI in research, which it calls the most comprehensive update since the first version in 2024 (DG RTD news, 16 Oct; guidelines PDF).
The intended layer. This is a non-binding stakeholder document from the ERA Forum, addressed to researchers, research organisations and research funders. It says outright that it adds no new requirements; it translates existing research-integrity principles to a workflow that now includes AI. Publishers and libraries are named as part of the conversation, but the recommendations are not written for them. Read it as the expectations your authors, reviewers and funders will bring to you, not as a rulebook for your production desk.
What is new and useful.
The guidelines also point to the ISC, COPE, STM and Global Young Academy work on a global AI-disclosure standard, which will inform future versions. The next revision is planned for the third quarter of 2027.
Supplements for the metadata desk. The guidance stops at principles, as its layer should. Here is what we would add on the publishing side.
1. A disclosure record you can actually harvest. If authors are asked to describe substantive AI use and verification, give them fields rather than a free-text box. A minimal set:
| Field | Example value | Why |
|---|---|---|
ai_tool |
product name as shown to the user | identifies the system |
ai_version |
version or date string | outputs differ between versions |
task |
controlled list: literature review / data analysis / translation / code / other | matches the guidelines’ “substantial use” examples |
substantive |
yes / no | the threshold the guidelines use |
verified_by |
ORCID iD of the responsible person | names the human who checked |
verification_note |
one sentence | how output was checked |
records_kept |
none / on request / deposited (with DOI) | prompts, logs, parameters |
2. A hidden-prompt check at intake. Before any AI screener reads a submission, run a short audit: extract all text layers, including white or zero-size text and zero-width characters; read document properties and embedded XMP metadata; flag any imperative instructions aimed at a model. Log the result with the submission. The same check belongs in metadata ingestion: an abstract field is now an attack surface.
3. Four questions for any review-assistance vendor. Is our content used to train your models, and is that contractual? Which model version scored this manuscript, and is that logged per item? Do you strip hidden instructions before inference? Can editors see why a flag was raised?
The pledges themselves we welcome without irony. Free tool access and AI-ready datasets are acceleration, and Europe needs it. The useful check over the next year is the one the Commission has set itself: an annual review, with pledges and progress presented at future editions of the summit. Delivered access, not announced access, is what researchers will notice.
What changed. EDItEUR published Issue 75 of the ONIX for Books codelists in mid-October. The changes were ratified by the ONIX International Steering Committee at its meeting during the Frankfurt Book Fair, and the change summary is dated 7 October (Issue 75 changes, PDF; live codelist browser, List 154). The AI-relevant additions:
EDItEUR’s note is careful about scope. These codes are for “initial exploratory use” of data in a special format that the German ONIX IG (the national user group) is to define. The linked data file sits apart from text meant for people, and any constraints or opt-outs on the book’s own content “should not be applied” to this file, because it is supplied specifically for AI use. Usage constraints can still be set for each supporting resource separately.
The intended layer. This is a controlled vocabulary for exchanging book-trade metadata, not a licence and not a payload format. It gives a sender a way to point at machine-only material and tell it apart from marketing copy. What that material contains is left, for now, to the German group. That is not a gap to complain about; it is the next work item, and DACH publishers have a direct line into it.
Two quieter additions matter to scholarly presses: List 163, codes 41 and 42, for the manuscript submission date and manuscript accepted date, “typically for academic and scholarly publishing”. Book metadata can now carry the same history dates that journal articles have long carried in JATS. In a week when disclosure standards ask when a work was reviewed, that is useful.
Supplements: a working checklist for an ONIX sender.
An illustrative supporting-resource block (check element order against the ONIX 3.0 specification before use):
<SupportingResource>
<ResourceContentType>60</ResourceContentType> <!-- LLM training data -->
<ContentAudience>13</ContentAudience> <!-- LLM training -->
<ResourceMode>04</ResourceMode> <!-- Text -->
<ResourceVersion>
<ResourceForm>02</ResourceForm> <!-- Downloadable file -->
<ResourceVersionFeature>
<ResourceVersionFeatureType>01</ResourceVersionFeatureType>
<FeatureValue>D512</FeatureValue> <!-- JSON -->
</ResourceVersionFeature>
<ResourceVersionFeature>
<ResourceVersionFeatureType>08</ResourceVersionFeatureType>
<FeatureValue><!-- SHA-256 of the file --></FeatureValue>
</ResourceVersionFeature>
<ResourceLink>https://example.org/onix/9783000000000-llm.json</ResourceLink>
<ContentDate>
<ContentDateRole>17</ContentDateRole> <!-- Last updated -->
<Date>20261019</Date>
</ContentDate>
</ResourceVersion>
</SupportingResource>What we would like the German format to carry, as a suggestion from the cheap seats: the work’s persistent identifiers (ISBN, DOI), the rightsholder, a licence statement in machine-readable form, the version of record it summarises, and who wrote the summary - a person, a model, or a model checked by a person. That last field is the disclosure question again, arriving through the back door of the book trade.
ONIX homework: who acts, on what, by when.
The layer first. The ONIX codelists are a technical standard. They are a shared vocabulary for book-trade metadata, not a licence and not a deadline. Issue 75 is dated 12 October 2026 (Issue 75 codelists). Every new code is optional. EDItEUR marks the LLM codes for “initial exploratory use” and sets no date by which recipients must read them. So there is no deadline. The practical one is before your next feed-schema update.
<ContentAudience>13</ContentAudience> and <ResourceContentType>60</ResourceContentType> in a Block 2 <SupportingResource>?<ResourceVersionFeatureType>01</ResourceVersionFeatureType> with <FeatureValue>D512</FeatureValue>?<PublishingDate>?<EpubUsageConstraint>, using List 145 code 11 (or the narrower codes 13 and 14) with a List 146 status. EDItEUR says those book-level constraints should not be applied to the LLM-training file. Terms for that file go in its own <EpubUsageConstraint> inside the <SupportingResource>, and that placement needs ONIX 3.1.1 or later (ONIX 3.1.1 changes). If your feed is still on 3.0, it has nowhere to state terms for the file. Hold the file back until it does. Record both decisions in the same rights record.<PublishingDate>. These dates come from your submission and review system, not from ONIX. The work is mapping that field, not retyping it.On the journal side, two dates for the diary. Crossref’s annual meeting and board election is online on 22 October, with an in-depth update on the new Member Practices on the agenda (Crossref). On 3 November, a Crossref webinar with Patricia Feeney, Liz Allen and Mike Nason looks at bringing the NISO CRediT and JAV recommendations into Crossref metadata, and “what implementation actually involves” (CRediT, 17 Oct).
In a Scholarly Kitchen guest post on 14 October, Sarah Hands (Chief Operating Officer, PA EDitorial) argues that AI adoption in editorial workflows may be running ahead of the evidence. She is explicit that this is not resistance. Her concern is that success is measured too narrowly (speed and throughput), that wider costs show up elsewhere in the workflow, and that automation bias makes a system flag feel more authoritative than a colleague’s judgement. Her test: “An AI-generated flag isn’t necessarily a finding; it is a signal requiring context and interpretation.” She asks publishers to share null and negative results, not only successes (Scholarly Kitchen, 14 Oct).
The intended layer. An opinion piece from a service provider, not a policy. It sets out questions and leaves the measurement design to the reader, which is fair for the format.
Supplement: a one-page evaluation sheet for any AI step in an editorial or metadata workflow.
Publish the sheet, even when the result is “inconclusive”. We are in favour of moving fast; we would just like to know where we are going.
Japan - a training report with rights levels. On 14 October, the LLM-jp team, coordinated by the National Institute of Informatics (NII), posted the technical report for LLM-jp-4, an 8B dense and a 32B-A3B mixture-of-experts model trained from scratch on about 12 trillion tokens and released under Apache 2.0 (arXiv:2610.19331). What makes it relevant to publishers is not the benchmark table but section 2, the selection rules:
The NII’s media advisory of 15 October announces the second Open LLM Symposium for 5 November in Tokyo, with EuroLLM lead André Martins (Instituto Superior Técnico, Lisbon) among the keynotes. The symposium is part of a programme funded by Japan’s education and science ministry (MEXT) on transparency and reliability of generative AI models (NII media advisory, PDF).
For a licensing desk, the release-level table is a ready-made template. Separating training, retrieval and display and redistribution into three explicit permissions is exactly the distinction most content licences still blur. Asking an AI partner “which of the three do you need?” is a better opening than “may we discuss AI?”.
Japan - AI proposes, the expert verifies, the record remembers. The Humanitext project led by Naoya Iwata (Nagoya University and NII), with colleagues at J.F. Oberlin University, the University of Tokyo and NII, is one of 19 projects chosen from 501 proposals in 42 countries for Schmidt Sciences’ Humanities and AI Virtual Institute. The grant is US$816,667 over three years, October 2026 to September 2029 (Schmidt Sciences, 14 Oct; NII release, 15 Oct).
The method is the part to copy. A language model proposes candidate links between classical texts, such as quotation, allusion, paraphrase or critique. Experts and graduate students accept, correct or reject each one. Verified links are stored with their type, direction, confidence and verification history in Wikibase and Neo4j, and then used as grounding for GraphRAG answers, which will be tested against ordinary retrieval-augmented generation. The researchers first build 500 relations by hand as a benchmark, aim for more than 10,000 verified relations in three years, and plan to release the graph, annotation software and benchmarks under open licences. It builds on Humanitext Antiqua, a research platform, public since August 2024, that answers questions about more than 2,300 Greek and Latin works by pointing to the source passages.
For cataloguers this is a familiar workflow wearing new clothes: machine suggestion, human decision, recorded provenance. If you want to run the same pattern on subject indexing or authority links, the minimum record per decision is short: proposed_by (model and version), decided_by (person and role), decision (accept / modify / reject), basis (the passage or source consulted) and timestamp. Keep it, and the next audit takes an afternoon instead of a quarter.
China - integrity talks and a standard still in the queue. In her letter of 12 October, COPE Chair Nancy Chescheir reports that after a June delegation to Beijing at the invitation of the Society of China University Journals, COPE is setting up an ad hoc group to address research and publication integrity concerns together with Chinese colleagues. Co-Vice Chair Tim Kersjes now leads COPE’s work on AI and external liaison (COPE, 12 Oct). The national guidance document on AIGC usage boundaries in academic publishing that we noted last week still shows as at the approval stage (正在批准) in the national standards register when we checked on 18 October (project 20253114-Z-469).
India, Singapore and elsewhere. We found no new primary-source development this week for India: the NIRF retraction story we covered on 12 October has not moved on in any official source we could find. We also found nothing solid from Singapore or other BRICS members for this desk, and we say so rather than pad.
Not a method. Seven signals from the sources above:
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, Oct. 19). OS DataMercs: News wrap · 19 October 2026 - Labelled for machines. Retrieved from https://www.datamercs.net/posts/2026-10-19-industry-notes-labelled-for-machines/
BibTeX citation
@misc{schmalfuß2026news,
author = {Schmalfuß, Olaf},
title = {OS DataMercs: News wrap · 19 October 2026 - Labelled for machines},
url = {https://www.datamercs.net/posts/2026-10-19-industry-notes-labelled-for-machines/},
year = {2026}
}