Accenture, the consultancy that sells artificial intelligence services to thousands of clients, has been accused of publishing research containing fabricated citations, in a case that has become an awkward test of how the industry checks AI-assisted work.

AI-detection firm GPTZero reported on 4 November 2026 that it found 31 citations it classifies as hallucinated out of 245 across three papers linked to Accenture. That is about one source in eight. The firm said a real human verified each flagged citation, but it also cautioned that its tools do not definitively prove that a language model wrote the papers or the citations, describing its findings as significant circumstantial evidence.

The largest count was in a November 2025 policy paper on artificial intelligence in African agriculture, titled Leapfrogging Development and co-produced by Accenture's African arm with researchers from IE University in Madrid. GPTZero found 15 flagged citations among its 89 sources, and said the paper also contains entirely fictitious case studies. The Financial Times reported the findings the same day under the headline Accenture accused of publishing AI paper littered with AI hallucinations.

The other two papers were research papers posted on arXiv in May 2026 by Accenture researchers. One, ContextEcho, had seven flagged citations among 102, and the other, Authenticity Debt, had nine among 54. GPTZero acknowledges that these two were not official Accenture publications, although they were written by Accenture researchers using company affiliations.

Some of the errors are striking. GPTZero said the agriculture paper cites references to organisations that do not appear to exist, including a Southern Africa AgriTech Network and a Southern African Agriculture Transformation Network. In one example examined by the Financial Times, the paper claimed predictive AI had improved detection of cocoa swollen shoot disease in Colombia and Ecuador to 99.1 percent accuracy. GPTZero said the cited study concerned African cocoa cultivation, not Latin America.

In another of the papers, a reference meant to support a claim about a malicious chatbot called FraudGPT linked to a fake GitHub repository that, GPTZero said, contained real malware. A further eight citations were flawed in ways GPTZero called indicative of LLM generation.

Accenture has pushed back on the cause, if not on the errors themselves. The company told the Financial Times that its review found the errors in the agriculture report stemmed from human error, and that there was no evidence to support the conclusion that they were caused by AI hallucinations. It said it would correct the report and republish it, but declined to explain how it reached that conclusion. IE University said its authors were reviewing the sources and data, and would correct confirmed errors and strengthen safeguards.

The dispute over what caused the errors matters because the label itself needs care. GPTZero's definition of a hallucinated citation covers references that do not match any real source, but also real publications with incorrect titles, dates, authors or other bibliographic details. The 31 should therefore be described as citations GPTZero flagged under its definition, not as 31 proven nonexistent studies. Independent experts have also cautioned that citation errors alone do not show that AI was involved, since ordinary citation tools and human mistakes can produce similar faults.

Even so, the episode is uncomfortable for a firm of Accenture's size and ambitions. Its chief executive has described AI as central to the company's future, and the firm advises clients on how to deploy the technology safely. GPTZero says its investigation team has documented the same fabricated-citation pattern at all four of the Big Four accounting firms over the past year, suggesting the problem reaches well beyond one consultancy.

What is not yet established is how the false citations got into the papers, or whether the two arXiv papers were investigated in the same way as the agriculture report. Accenture's documented response concerns the agriculture paper only, and the available reporting does not establish that it addressed the other two papers. Until the corrected paper is republished, the episode stands as a reminder that AI-generated text can look authoritative while resting on sources that do not exist, and that checking references remains a human job that cannot be skipped.