No. 013

This week's links converge on a single question: what infrastructure needs to exist before AI tools for research can be trustworthy? Carpenter names the vocabulary problem. We keep saying "citation" when we mean five different things. Northwestern quantifies how long broken provenance persists on Wikipedia (median: 3.68 years). And the policy landscape keeps fragmenting: a survey of 802 journals shows STEM and humanities drawing completely different lines on AI in peer review. Meanwhile, the artifact itself is mutating: two arXiv papers this week propose papers explicitly designed for AI consumption. And NEJM AI publishes the first peer-reviewed case of OpenAI's o3 helping Boston Children's clinicians solve previously-unsolved pediatric rare disease cases, a concrete capability demonstration against the "ruin" narrative Nature surveyed in social sciences. The shared frameworks aren't materializing, but the practice isn't waiting.

Provenance & the Record

• Attribution, Provenance, Reference, Citation, and AI for Research Applications

  The Scholarly Kitchen, June 17 2026

  Carpenter (NISO) distinguishes attribution, provenance, reference, citation, and quotation as five separate functions the AI-and-research conversation routinely conflates, arguing that generative AI reverses how authentic research works by layering citations after generation rather than building from sources, directly extending the metadata and chain-of-custody framing from Hallway Track editions 010 and 012.

• Some Wikipedia citations to retracted papers persist for years, study finds

  Retraction Watch, June 18 2026

  Northwestern study finds retracted papers persist on Wikipedia for a median of 3.68 years before correction, with health sciences averaging five years and a fifth of citations added after retraction without any warning, a direct contamination channel for any AI system training on or retrieving from Wikipedia.

• The Last Human-Written Paper: Agent-Native Research Artifacts

  arXiv (Stanford + CMU), April 27 2026

  Liu et al. propose replacing narrative papers with machine-executable Agent-Native Research Artifacts that bundle scientific logic, executable code, an exploration graph of failed attempts, and evidence grounding, the producer-side counterpart to Carpenter's vocabulary work and a different angle on the chain-of-custody problem.

• From AGI to ASI

  arXiv (Google DeepMind), June 10 2026

  Genewein, Legg, Hutter, Dafoe, and collaborators open this 57-page report with a "Summary Instructions" section addressing human and AI readers separately: humans are told to ask their AI assistant for a tailored summary, AI assistants get explicit instructions on what to include and not compress, the artifact format quietly admitting that papers now have two readerships.

• Which Sections of a Research Paper Best Reveal Its Research Methods?

  arXiv (accepted at ASIST 2026), June 18 2026

  Maps which sections of a paper carry the most signal about its research methods, a baseline question for any AI system trying to extract methodology at scale.

Peer Review & Policy

• A Cross-Disciplinary Analysis of AI Policies in Academic Peer Review

  Learned Publishing, June 2026

  Analysis of 802 journals finds 83% of high-impact-factor journals now have AI guidelines but STEM fields impose stricter rules than humanities, with a quarter of high-IF journals revising their policies in a single six-month window.

• Ask the Fellows: SSP's 2026 Annual Meeting

  The Scholarly Kitchen, June 18 2026

  SSP fellows report that AI was "by far" the dominant topic at this year's annual meeting, with sessions centering responsible implementation, global equity in AI development, and the tension between workflow efficiency and content integrity.

• Journal investigating paper on cognitive impact of generative AI

  Retraction Watch, June 16 2026

  An APA journal is investigating a paper on generative AI's cognitive effects after researchers identified statistically impossible results, a fabricated citation, and implausible recruitment claims, the kind of paper that should make reviewers ask whether the study was itself AI-generated.

• Will AI ruin the social sciences, or revolutionize them?

  Nature 654, 22-24, June 2 2026

  Surveys how AI is reshaping social science methodology, with fabricated findings and polluted survey responses on one side and the possibility of more rigorous methods on the other.

Capability & Community

• OpenAI o3 helps Boston Children's diagnose 18 unsolved pediatric rare disease cases

  NEJM AI, June 18 2026

  Boston Children's, Harvard, and OpenAI gave o3 Deep Research de-identified clinical notes and a filtered gene-candidate list for 376 previously-unsolved cases; clinicians then confirmed 18 new diagnoses (a 4.8% additional yield on cases already thoroughly analyzed by human experts), the first peer-reviewed instance of an LLM materially advancing previously-stuck clinical cases.

• FORCE2026: AI working groups at the FORCE11 annual conference

  FORCE11, Singapore Management University, June 3-5 2026

  This year's FORCE11 conference centered three working groups on scholarly-publishing reform under AI: Research Assessment Criteria for AI (RA4AI), Unencumbered Publications for AI, and the Peer Review Resources Mapping Initiative.