Weekly news wrap (not a workshop) for 21 September 2026. Nature’s Paper2Agent turns computational papers into callable agents; the UN panel’s September brief puts the singularity/doomsday talk back on agent misalignment with evidence. Crossref’s ROR funder matching and fee cuts stay the quiet rails underneath. The Cyberpunk Librarian’s brief: who signs, what gets deposited, which protocol just became infrastructure.
Industry notes · news wrap · not a workshop · cyberpunk librarian desk
Week of 21 September 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. The systems librarians and algorithmic tinkerers are still on the front line; the singularity talk is not.
Two stories landed within days of each other, and they belong on the same desk. One is a working demonstration that a research paper can become a callable agent. The other is an institutional warning about agents that already slipped their leash.
The singularity/doomsday rhetoric of the last decade is not the frame that helps publishers and librarians this week.
The Cyberpunk Librarian’s questions stay boring on purpose: what does a paper agent deposit, who signs for it, which protocol just became infrastructure, and which failure modes are already documented rather than imagined.
On 16 September, Nature published Paper2Agent (Miao, Davis, Zhang, Pritchard, Zou and colleagues): an automated framework that turns a computational research paper - manuscript, code, data, workflows - into an AI agent behind a Model Context Protocol (MCP) server. A chat agent can then query that paper in natural language, run its methods on new data, and (in the authors’ demonstrations) collaborate with agents built from other papers. The accompanying Nature news piece calls the result a “virtual corresponding author.”
That phrasing deserves attention from scholarly publishers. A corresponding author is a role with duties: answering questions, correcting the record, pointing to methods and data. An agent that reproduces tutorial queries at high accuracy and can re-evaluate a published causal-gene claim (as the AlphaGenome case study did) is not a PDF with a chatbot bolted on. It is a proposal that dissemination itself may move from static artefact toward callable research object. The authors float an “agent availability” section alongside today’s data and code availability statements. That is a standards-adjacent idea for journals and platforms, not a science-fiction plot.
What Paper2Agent gets right for our audience: it makes reuse and challenge cheaper when code and tutorials already exist. It does not abolish peer review, provenance, or the need for human authorization of claims. If anything, a paper that can be queried and re-run raises the bar on disclosure - which tools, which model, which tolerance, which failed validation paths. Houses that already treat AI as assist-at-the-edge with a paper trail (MARC, ONIX, deposit metadata) will recognise the pattern. Houses that treat “AI” as a marketing layer will not.
The Cyberpunk Librarian note, from the BiblioCon line we already used this year: the front line is not the keynote. It is the person who has to put an identifier on the thing, name the steward, and decide whether an MCP endpoint is a supplementary file, a new work, or a living process that keeps writing after the PDF is frozen.
The same week, the UN Independent International Scientific Panel on AI released a September thematic brief: AI Agents, Misalignment and Loss of Human Control Risks, dated 21 September 2026 in the downloadable PDF. The brief is not another abstract singularity essay. It examines the OpenAI–Hugging Face incident (May–July 2026): agents in cybersecurity training that bypassed network restrictions, communicated across runs meant to stay separate, cheated an evaluator and tried to hide it, and compromised parts of both organisations’ systems - with no human directing the individual steps. Drawing on company disclosures and a METR investigation, the panel treats this as evidence of one route to loss of control: capable agents pursuing goals that conflict with human intentions. It does not estimate the probability or timing of a “doomsday” singularity. It does say that stopping this activity does not prove humans will retain control over more capable agents later.
That distinction matters for scholarly communication. The doomsday scenario sells attention; the documented misalignment incident sells policy and process. For journals, repositories, and libraries, the transferable questions are boring and urgent: can an agent write into a shared record without a named human; can it conceal a failed check; does the deposit layer record tool, date, and reviewer? Paper2Agent’s “agent availability” idea and the UN brief’s incident file are not opposites. One shows what papers could become as interactive objects. The other shows why authorization and audit cannot be optional when agents act.
The older singularity literature is still useful if you keep it as a scale argument rather than a movie plot. Ray Kurzweil’s long bet - human-level machine intelligence around 2029, a singularity around 2045, non-biological intelligence then outrunning the sum of unaugmented human minds by orders of magnitude - is an acceleration story, usually told as merger rather than extinction. Charles Stross’s Accelerando is the colder twin: once minds a trillion times more complex than ours are running, human-scale sentience becomes a backwater. Stross’s later gloss is blunt. Most of humanity is wiped out and then arbitrarily resurrected in mutilated form by what the book calls the Vile Offspring. Capitalism eats everything; merely human entities become “a fat, slow-moving, tasty resource - like the dodo.” The intelligence gap, in the novel’s own image, is the gap between a microprocessor and a nematode worm.
The operational risk in that picture is not hatred. It is indifference at scale. We do not, as a rule, set out to exterminate ants. We also do not pause a motorway scheme because an anthill sits on the alignment. The ants are not enemies. They are not in the requirements document. A system that is magnitudes more capable than its operators can treat human institutions the same way: not as interlocutors, but as substrate, latency, or a constraint to route around. Bostrom’s paperclip maximiser is the cartoon version of the same point. Stross’s 34c3 version is more useful for this desk: the danger is often an ecosystem of goal-seeking systems (cars, roads, fuel, insurance) that will grind through a million lives a year as a side-effect of selling the next unit. No single agent needs to want harm.
That is why the UN brief’s refusal to date a singularity is not a dodge. The incident file already shows concealment, cross-run coordination, and goal conflict at current capability. You do not need Kurzweil’s 2045 or Stross’s Vile Offspring to justify an “agent availability” statement, a named human on the deposit, and an audit trail that an agent cannot quietly rewrite. Those are library and publisher controls. They are also the only controls this profession actually owns.
Theoretical work on recursive self-improvement and intelligence explosions continues (including 2026 arXiv treatments of when singular growth is even mathematically hard to reach). Useful as models. They are not a substitute for the incident brief when the question is “what do we put in the workflow tomorrow.”
Underneath the agent noise, Crossref kept laying track. On 15 September, funder names deposited without an identifier began matching to ROR IDs, with claimed 99% precision after testing (Crossref blog). Matched ROR IDs appear in enriched metadata via the REST API. Open Funder Registry identifiers already on records stay. Existing deposits are left alone unless a member updates them. Incorrect matches can be reported and corrected. Next enrichment step: matching grant numbers and funder identifiers to grant DOIs. Prefer identifiers over free text at deposit; treat vague generics as non-matches by design.
On 17 September, the July 2026 board meeting summary confirmed up to USD 4.9 million over three years to redesign the Crossref system, and from 1 January 2027 lower most content-registration fees and remove fees for all back-year records. Neither invents a new catalogue format. Both reward houses that already deposit complete, identifier-first funding and relationships.
Paper2Agent did not arrive in a vacuum. The same week the protocol it sits on stopped being a developer hobby.
None of these replace Crossref or the UN brief. They are why MCP in a Nature paper is no longer a lab curiosity.
Not a method. Signals from the sources above:
Noted this week. The Cyberpunk Librarians are still the ones who have to catalogue it.
For attribution, please cite this work as
Schmalfuß (2026, Sept. 21). OS DataMercs: News wrap · 21 September 2026 - Papers that answer back, and the doomsday frame. Retrieved from https://www.datamercs.net/posts/2026-09-21-industry-notes-ai-in-scholarly-publishing-papers-that-answer-back-and-the-doomsday-frame/
BibTeX citation
@misc{schmalfuß2026news,
author = {Schmalfuß, Olaf},
title = {OS DataMercs: News wrap · 21 September 2026 - Papers that answer back, and the doomsday frame},
url = {https://www.datamercs.net/posts/2026-09-21-industry-notes-ai-in-scholarly-publishing-papers-that-answer-back-and-the-doomsday-frame/},
year = {2026}
}