Academic publishing is becoming AI-native before its governance is ready. This workshop convenes the PC chairs of ICLR, ICML, and CVPR to redesign authorship, peer review, citation, and conference governance under AI.
LLMs and AI agents are no longer just objects of study; they are now participants in the academic pipeline. Authors draft with them, reviewers summarize with them, program chairs deploy them for triage and meta-review. The institutions of science were not designed for this.
From AI for science to AI-native academia. Most AI-for-science venues ask how AI accelerates discovery. We ask a different question: how should the institutions of science themselves change when AI is embedded in authorship, review, citation, and scholarly memory?
Failure modes are appearing at every node: fabricated citations at submission, AI-generated reviews, prompt injection in manuscripts, recursive feedback into training corpora. Human-AI co-hallucination (claims no model or human would produce alone) is one among several. The workshop builds a shared taxonomy, security model, and governance framework.
Program chairs are already running scattered interventions, reviewer AI-use policies, self-ranking mechanisms, in-house detectors, adversarial-submission audits, AI-author pilots, citation audits, with no shared framework.
We invite technical, empirical, and policy submissions on nine topics organized around the AI-native academic pipeline: author → submission → reviewer → AC/SAC/PC → platform → dissemination.
| 1 | AI-Assisted Authorship and Submission Integrityauthor / submitter | LLM use in drafting, citation, and rebuttals; disclosure, AI-shaped novelty, fabricated claims, submission inflation, paper mills, accessibility for non-native writers, and the line between help and distortion. |
| 2 | Adversarial Manuscripts and Submission-Side Attackssubmission artifact | Hidden prompt injection, invisible text, adversarial figures, and jailbreaks against AI reviewers and citation checkers; document sanitization, submission-portal defenses, and review-tool security. |
| 3 | AI-Assisted Peer Review and Reviewer Accountabilityreviewer | AI-generated and AI-polished reviews, confidentiality risks, reviewer over-reliance, and review hallucination; measuring review quality and supporting reviewers without replacing human accountability. |
| 4 | Meta-Review, PCs, and Conference-Scale Decision SupportAC / SAC / PC / organizer | AI for reviewer assignment, desk-reject triage, score aggregation, and collusion detection; mechanism design under AI pressure, including author self-ranking, and audit trails for AI-influenced decisions. |
| 5 | Human-AI Co-Hallucination and Cross-Role Error Propagationcross-cutting failure mode | Taxonomies and measurements of false claims that arise through human-AI interaction and propagate through review, citation, and reuse, with intervention points before they become institutionalized. |
| 6 | Citation, Credit, and Scholarly Knowledge Integritycitation / indexing node | Citation hallucination, support-of-claim verification, citation manipulation, missing-credit detection, and knowledge-graph contamination; verifying not just that a citation exists but that it supports the claim. |
| 7 | Provenance, Plagiarism, Detection, and Due Processintegrity enforcement layer | AI-generated paper detection, watermarking, text and figure provenance, and memorization risks; detector reliability, false-positive harms to non-native authors, and appeal mechanisms. |
| 8 | Recursive Scholarly Feedback Loops and Corpus Contaminationpost-publication / future corpora | How AI-generated papers, reviews, and citations re-enter training corpora: model collapse in scientific text, self-reinforcing citation errors, benchmark leakage, and long-term homogenization. |
| 9 | Publication Infrastructure, Policy, and Venue Repositioningplatform / institution / governance | Enforceable AI-use policies, audit trails, confidentiality, and platform infrastructure (OpenReview, arXiv, ACM, IEEE); scaling review past 30,000+ submissions without collapsing into AI-only evaluation. |
Submission tracks (via OpenReview, non-archival): Short papers (4 pages) and long papers (9 pages), references excluded, double-blind reviewed. Work already accepted at NeurIPS main or other ML venues is excluded.
Submission template: Use the official NeurIPS 2026 LaTeX template on Overleaf.
We are considering Humanly as a provenance pilot: a configurable writing environment that records in-platform edits, clipboard activity, and AI use, then issues a signed, verifiable certificate with authorship statistics and anomaly signals. The signals support human review, not an automatic verdict.
Moderator: Atlas Wang (UT Austin / XTX Markets), framing the panel around how AI text re-entering the scientific corpus degrades future models.
Panel theme: "Redesigning AI Venues Under AI" (60-min All-PC-Chair Panel + 30-min Open Q&A with the PC chairs).
Panel question: "By NeurIPS 2027, what must AI venues do, operationally, about human-AI co-hallucination across authorship, review, citation, and governance, including AI-review identification, prompt-injection defense, citation grounding, disclosure, and review scaling under 30k+ submissions?"
