The 35th USENIX Security Symposium received roughly 3,030 paper submissions this year, an all-time high. The conference rejected 21 of them for containing three or more hallucinated references, and removed five reviewers from its 496-member program committee for producing suspected AI-generated peer reviews.
The numbers come from a transparency report published by program co-chairs Ben Stock of CISPA and Elissa Redmiles of Georgetown University, and first reported by The Register. The report offers a rare look at how a major academic conference is operationalising AI content governance — not as policy on paper, but as tooling with defined thresholds and documented outcomes.
What the Tooling Actually Does
After catching one paper with nonexistent references in the first submission cycle, Stock and Redmiles built automated tooling that extracts references from submitted PDFs, queries databases like DBLP and arXiv, and flags references that cannot be verified. Papers with three or more unconfirmable citations were rejected outright.
The threshold is arbitrary but pragmatic. More than 100 additional papers contained at least one reference reviewers could not confirm. Conference organisers chose not to investigate those, acknowledging that some could be false positives from name spelling differences or missing indexing. The three-citation rule is a bright line — harsh enough to signal seriousness, lenient enough to avoid penalising honest formatting errors.
Stock was careful not to attribute all fabricated citations to AI. “We cannot say with certainty that these were AI-hallucinated,” he told The Register, “but nevertheless considered these papers to be problematic and thus rejected them.” The distinction matters. Fabricated references predate large language models. What has changed is the scale: a researcher using an LLM to draft a literature review can generate dozens of plausible-looking but nonexistent citations in seconds.
The Reviewer Problem Is Harder
The five reviewer removals point to a different failure mode. Program committee members are trusted experts, sworn to confidentiality. Using an LLM to write reviews means feeding unpublished manuscripts to a third-party service — a confidentiality breach independent of the content quality.
Stock described the cases as “a tiny number” where organisers reached “sufficient confidence that AI was used.” The affected authors were allowed to resubmit. But the detection method was not disclosed, which raises a question: if five were caught, how many more went undetected?
This is the structural problem. AI Governance Institute notes that acceptable use policies must extend beyond authors to everyone in a review or approval chain. Compliance programmes that restrict AI use for content creators but not reviewers have a gap — and that gap is invisible until someone gets caught.
A Benchmark Other Fields Will Reference
The USENIX model is significant because it publishes auditable metrics. Rejection counts, removal counts, detection thresholds, and the tooling approach are all documented. That makes it a reference point regulators, courts, and auditors can cite when evaluating whether an organisation’s AI content governance is adequate.
This is not hypothetical. The pattern of documented consequences for AI-assisted fabrication is accumulating across professional domains. Attorneys have been sanctioned for AI-hallucinated case citations. Government officials in South Africa were suspended after fabricated citations appeared in a Cabinet-approved white paper — a story we covered previously when the policy was first withdrawn. The USENIX case adds academic conferences to the list of institutions now enforcing consequences.
The April paper “More Versus Better: Artificial Intelligence, Incentives, and the Emerging Crisis in Peer Review,” published in Organization Science, found that submission volume at major academic journals has increased 42 per cent since the release of ChatGPT. USENIX’s jump from 2,400 to 3,030 submissions aligns with that trend. More papers, more reviewers needed, more surface area for AI misuse.
What NZ Institutions Should Watch
New Zealand universities submit to and review for international conferences like USENIX. The Conference of Research Excellence (CoRE) and Tertiary Education Commission have not published AI-specific research integrity guidelines comparable to what USENIX has implemented. If international conferences begin requiring citation verification as a condition of submission, NZ researchers using AI tools carelessly will face the same rejection risk as anyone else.
The deeper question is whether NZ academic institutions will adopt similar internal controls. The University of Auckland and Victoria University of Wellington both have academic integrity policies that mention AI, but neither has published tooling requirements or detection thresholds for staff-written reviews or administrative documents. The USENIX model — automated verification plus a defined threshold plus documented consequences — is a concrete template that could be adapted locally.
❓ FAQ
Can AI-generated citations really fool peer reviewers? Yes. LLMs produce references that look structurally correct — real journal names, plausible author combinations, realistic DOIs — but point to papers that do not exist. USENIX found that manual review alone was insufficient, which is why they built automated cross-referencing against DBLP and arXiv.
What happens to researchers whose papers are rejected? USENIX rejected the papers from the current cycle. Authors can resubmit in future cycles with corrected references. The five removed reviewers were dismissed from the program committee and affected authors were allowed to resubmit their papers.
Is this just a USENIX problem? No. NDSS saw submissions jump from 694 in 2024 to 1,481 in 2026. The Organisation Science paper documented a 42 per cent increase across major journals since ChatGPT’s release. Every major conference and journal is dealing with this, though few have published their enforcement metrics.
Could NZ researchers be affected? Any NZ researcher submitting to USENIX, NDSS, or comparable international venues faces the same citation-verification checks. Domestically, NZ institutions have not yet implemented equivalent tooling, but the precedent set by international conferences creates pressure to align.
