Between mid-2025 and early 2026, multiple governments issued official documents underpinned by research citations later found to be fabricated or AI-generated, exposing weaknesses in institutional knowledge verification amid the rise of artificial intelligence.
- AI-fabricated citations found in official government policy documents worldwide
- Reliable AI text detection tools are improving but remain imperfect
- Research integrity now a critical national security and policy challenge
What happened
Between May 2025 and April 2026, several governments released official documents based in part on research that included fabricated or AI-generated citations. The European Union’s cybersecurity agency ENISA published threat analysis reports with dozens of incorrect references, later partially corrected. South Africa withdrew a draft national AI policy after media scrutiny revealed multiple fake academic citations. Similarly, a US government health report cited non-existent studies and misattributed sources, with many citations identified as likely generated by OpenAI tools.
These errors spanned different regions and policy areas yet shared a common cause: institutional reliance on AI-assisted research without sufficient verification. The incidents exposed a systemic vulnerability in the knowledge production processes governments have traditionally relied upon, raising alarms about the trustworthiness of publicly issued scientific and policy findings incorporating AI-generated content.
Why it matters
Governments have historically depended on rigorous processes to ensure the integrity of knowledge used for policymaking, including peer review and analytical validation. AI’s rapid integration into research workflows now challenges these assumptions by producing convincing but often fabricated data, especially in specialized or less visible topics where validation is difficult. This shifts the national security landscape beyond cyber threats to the integrity of information itself.
Studies have shown alarming rates of fabricated citations by advanced AI models like GPT-4, with up to a quarter or more of generated references either inaccurate or completely invented. The growing volume of AI-assisted academic content — estimated at up to 40 percent of recent biomedical abstracts in some fields — exacerbates this challenge, as undisclosed AI usage further complicates verification efforts. The risk is that flawed or false knowledge can misinform vital policy decisions or degrade public trust in official findings.
What to watch next
Efforts to detect AI-generated text and citations have made progress with tools including Pangram, OriginalityAI, GPTZero, and RoBERTa; however, these solutions remain imperfect and must be integrated into institutional workflows systematically. Stakeholders will need to develop standards for AI disclosure, enhance research review protocols, and invest in new methods for ensuring content authenticity to safeguard knowledge production.
Policymakers and security officials must elevate the integrity of research as a strategic priority, recognizing it as a national security concern on par with AI model safety and cyber defense. The ongoing expansion of AI’s footprint in science and policymaking will require adaptive responses to preserve the credibility and reliability of information critical to public trust and effective governance.