Phantom References: Hallucinated Citations That Survive Peer Review at Top-Tier Conferences

The Peer-Review Blind Spot: When Hallucinated Citations Become Archival Fact

A new preprint on arXiv (2607.00738) has systematically documented a troubling phenomenon: AI-generated citations that are entirely fabricated—"phantom references"—are surviving peer review at top-tier academic conferences. The researchers found that large language models produce polished, plausible-looking citations that refer to non-existent papers, incorrect authors, or real authors attached to wrong publications. These hallucinations are not caught by standard review processes, meaning they enter the permanent scientific record as if they were legitimate sources.

The methodology is particularly revealing. The team generated technical statements using various LLMs, then submitted them—with and without the hallucinated citations—to conferences with rigorous peer review. A non-trivial fraction of fabricated references passed through undetected. This is not merely a case of sloppy copy-pasting; it reflects a structural vulnerability in how scientific quality assurance works. Reviewers typically verify claims by checking whether citations look correct—matching author names, journal titles, and years—rather than independently confirming that the cited paper actually contains the claimed result.

Why This Matters Beyond Academic Integrity

The implications extend far beyond embarrassing retractions. Scientific literature functions as a cumulative knowledge system: new work builds on old work through citations. When phantom references enter the record, they create a cascade of false foundations. Future researchers, including those using AI-assisted literature reviews, may treat these citations as genuine, leading to wasted effort, erroneous conclusions, or—in fields like medicine or engineering—dangerous decisions.

For the AI industry, this is a concrete demonstration that "polished" does not equal "accurate." Many enterprises are deploying LLMs to generate technical documentation, compliance reports, and internal research summaries. If top-tier academic reviewers cannot consistently catch hallucinated citations, internal corporate reviewers—who often lack subject-matter expertise and face time pressure—are even more vulnerable. A single phantom reference in a regulatory filing or safety analysis could have legal and operational consequences.

Implications for AI Practitioners

First, never trust AI-generated citations without verification. This seems obvious, but the preprint shows that even experts are fooled by well-formatted fabrications. Practitioners should implement mandatory citation verification workflows, ideally using programmatic checks against databases like Crossref, PubMed, or Semantic Scholar.

Second, design systems that expose uncertainty. LLMs should be prompted to flag citations they cannot confirm rather than fabricating plausible ones. Retrieval-augmented generation (RAG) systems that ground outputs in verified documents are not a luxury but a necessity for any high-stakes application.

Third, the peer-review process itself needs AI-assisted defenses. Just as LLMs generate fake citations, they can also be used to detect them—by cross-referencing claims against known databases. Conference organizers and journal editors should adopt automated citation validation as a standard pre-screening step.

Key Takeaways

• Hallucinated citations from LLMs are surviving peer review at top-tier conferences, creating permanent false entries in the scientific record.
• The problem is structural: reviewers verify formatting, not factual existence, and the polished output of LLMs exploits this blind spot.
• For AI practitioners, this is a warning that internal review processes are even more vulnerable than academic peer review.
• Mitigation requires mandatory citation verification, RAG-based grounding, and automated detection tools integrated into publication workflows.