AI-Generated False Citations Found in Official Government Reports, Undermining Policy Integrity

The transition of AI hallucinations from an academic embarrassment to a systemic policy vulnerability is now complete. Between May 2025 and April 2026, three separate official government documents—spanning a European regulator, a national government, and the White House—were published containing fabricated academic citations generated by artificial intelligence.[1]

The discovery highlights a critical institutional blind spot. While governments worldwide have heavily invested in securing AI models against cyberattacks and catastrophic risks, they have largely overlooked the subtle corruption of the information supply chain. When policy is built on non-existent research, the foundation of regulatory and operational decisions begins to crack.[1][4]

The most prominent incident occurred within the United States government. A flagship "Make Our Children Healthy Assessment" released by the White House cited multiple scientific studies that simply did not exist, while misattributing findings to genuine authors who had never conducted the referenced research.[2]

The fabrication was first flagged by the nonprofit news publication NOTUS, which identified incorrect formatting, missing authors, and mismatched issue numbers in the report's bibliography. Following the inquiry, the report was quietly pulled, scrubbed of the fake citations, and reissued with authentic sources.

Further investigation by The Washington Post revealed the mechanical cause of the error. Journalists found dozens of footnotes in the original document containing an "oaicite" marker—a distinct digital fingerprint left behind by OpenAI's ChatGPT when it generates text and attempts to format references.[2]

Across the Atlantic, the European Union Agency for Cybersecurity (ENISA) faced a similar breach of research integrity. The agency's widely read threat landscape report was found by outside researchers to contain AI-generated or otherwise erroneous references that pointed to non-existent cybersecurity frameworks.[1]

According to an analysis detailed in TechPolicy.Press, one ENISA report contained 26 incorrect footnotes out of a total of 492. The cybersecurity agency was forced to issue a revised version of the document to correct the fabricated links, an embarrassing misstep for an organization dedicated to digital security and threat mitigation.[1]

In the Global South, the consequences of AI-assisted drafting were even more disruptive. The South African government was forced to completely withdraw its draft national AI policy after media reporting revealed that at least six of its sixty-seven academic citations were entirely fabricated.[1]

Rachel A. George, a Stanford lecturer and nonresident fellow at the Carnegie Endowment for International Peace, has cataloged these failures, arguing that research integrity must now be treated as a national security problem. Threat assessments and health policies directly dictate federal budgets, regulatory frameworks, and operational deployments.[1][4]

The underlying mechanism driving these fabrications is a well-documented flaw in large language models (LLMs) known as hallucination. When an AI model cannot find an exact match for a citation within its training data, it does not simply return a blank result. Instead, it relies on statistical inference to generate a plausible-sounding reference, complete with real author names, realistic journal titles, and standard formatting.[3]

Because these models are optimized for fluency and coherence rather than strict factual retrieval, they excel at mimicking the cadence of academic writing. The model cannot distinguish between a real citation it encountered in a legitimate academic paper and a fake citation it generated based on statistical probabilities.

A 2023 study published in Scientific Reports quantified the scale of this problem. Researchers asking ChatGPT to generate literature reviews found that 55 percent of citations produced by GPT-3.5, and 18 percent of those from the more advanced GPT-4, were entirely fabricated. More recent evaluations suggest the problem persists; a 2025 analysis of OpenAI's GPT-4o found that the model generated fabricated citations roughly 20 percent of the time.[1][3]

Crucially, the rate of AI hallucination increases significantly when models are queried about specialized, niche, or highly technical topics. This is precisely the domain where government policy analysis operates, making federal agencies uniquely vulnerable to AI-generated misinformation.[1]

The integration of web-browsing capabilities and Model Context Protocols (MCPs) into AI tools was supposed to anchor models to live data and reduce these hallucinations. However, experts note that models still struggle to distinguish between a legitimate academic paper and a fabricated citation encountered in a low-quality blog post.[1][3]

To combat this vulnerability, policy experts are advocating for a fundamental shift in how governments handle document publication. Researchers argue that automated citation verification must become standard infrastructure in the publishing pipeline, rather than an optional final check left to overburdened human staffers.[1]

Furthermore, there is a growing push for mandatory disclosure standards whenever AI is used in the drafting of government research. Just as financial conflicts of interest must be declared, the use of generative models in synthesizing policy data may soon require explicit attribution on the cover page of federal reports.[1][4]

Until these safeguards are institutionalized, the burden falls entirely on human reviewers. Agencies are being urged to invest in specialized training for staff to identify the subtle "tells" of AI-generated text, ensuring that the policies shaping the future are grounded in reality, not algorithmic imagination.[1]