AI Slop Is Spurring Record Requests for Imaginary Journals

The ICRC warned that major AI models including OpenAI's ChatGPT, Google's Gemini and Microsoft's Copilot are generating incorrect or fabricated archival citations that direct users to nonexistent journals and records; archivists and researchers report these hallucinations are wasting staff time and creating challenges in proving unique records don't exist. The Library of Virginia estimates about 15% of emailed reference questions are now ChatGPT-generated and says some include fabricated citations for published and primary sources. The ICRC advises consulting catalogs and published references rather than trusting AI citations, and libraries plan to require researchers to vet and disclose AI-originated sources and limit time spent verifying, underscoring operational, reputational and research-quality risks for institutions using AI-assisted research.

Analysis

The International Committee of the Red Cross (ICRC) has publicly warned that major large language models — specifically OpenAI’s ChatGPT, Google’s Gemini and Microsoft’s Copilot — are producing incorrect or fabricated archival citations that point users to nonexistent journals and repositories. The article cites concrete examples of falsified references and notes that these hallucinations have confused students, researchers and archivists who rely on accurate archival metadata.

The Library of Virginia reports that roughly 15% of emailed reference questions are now ChatGPT-generated and that some include fabricated citations for both published works and unique primary-source documents; staff say it is increasingly difficult and time-consuming to prove a unique record does not exist. The ICRC recommends consulting authoritative online catalogs and published scholarly references and the Library of Virginia plans to require researchers to vet and disclose AI-originated sources while limiting staff time spent on verification.

These developments create operational and reputational risks for institutions using LLM outputs without robust verification and imply modest near-term market sensitivity: sentiment is moderately negative (sentiment_score -0.4) and the measured market impact is limited but nontrivial (0.25). Investors should monitor vendor responses, adoption of attribution/verification features, and any institutional policy changes that could affect enterprise demand for generative-AI services.

Key Decisions for Investors

• Reassess near-term exposure to consumer-facing LLM providers until they publish clear mitigation and source-attribution roadmaps, given the reputational and operational risks highlighted by the ICRC
• Prefer investments in vendors that demonstrate enterprise-grade verification, human-in-the-loop workflows and provenance controls and monitor product announcements and enterprise customer retention at libraries and archives
• Track leading indicators such as institutional contract renewals, published usage policies from major archives and reported hallucination rates as early signals of revenue or adoption risk
• Consider trimming or hedging thematic AI positions because sentiment is moderately negative and market-impact is measurable, and add back exposure once concrete reductions in fabrication rates and verified mitigation plans are evident