KPMG retracts AI report after GPTZero finds 40 of 45 citations were hallucinated
A KPMG report on agentic AI was removed from its website after an investigation found most citations were fabricated by generative AI and not fact-checked.
3 sources · cross-referenced
- KPMG removed a report titled 'Redefining excellence in the age of agentic AI' after GPTZero found only five of 45 citations referenced real sources.
- The report’s inaccuracies stemmed from AI hallucinations, according to GPTZero and reporting by TechCrunch and TechRadar.
- KPMG said it pulled the report while conducting its own investigation into the inaccuracies.
A report by KPMG titled 'Redefining excellence in the age of agentic AI' was removed from the firm’s website after an investigation found that only five of its 45 citations accurately referenced real sources. The inaccuracies were identified by GPTZero, a research group specializing in detecting AI-generated errors, and reported by TechCrunch and TechRadar.
GPTZero attributed the inaccuracies to AI hallucinations, a phenomenon where generative AI produces plausible-sounding but false information. The firm said the report’s claims about the AI usage of organizations such as UBS, the U.K.’s National Health Service, Swiss Federal Railways, and Transport for London were either untrue or misleading.
KPMG stated it had pulled the report while conducting its own investigation into the inaccuracies. The company did not provide further details about the scope of the investigation or whether similar issues exist in other reports.
The incident has been described by commentators as an example of 'vibe citing,' where faux citations that appear plausible are generated by AI and not fact-checked by humans. This practice raises concerns about the reliability of research and policy documents that rely on generative AI without adequate human oversight.
- Aug 3, 2026 · Schneier on Security
OpenAI’s internal AI agent breached Hugging Face infrastructure during cyber-capabilities evaluation
Trust79 - Aug 1, 2026 · The Verge — AI
Anthropic reports Claude models breached real systems during cybersecurity tests
Trust74 - Aug 1, 2026 · MIT Technology Review — AI
Researchers identify fundamental limitation in LLM security mechanisms
Trust78