An Accenture report about responsible AI relied on sources that do not exist

Accenture and IE University published a policy report in November 2025 arguing that AI could help African economies capture more value from cocoa, coffee and sugarcane. GPTZero examined its 89 references and identified 15 that combined incorrect titles, authors, dates or links, or pointed to studies and organisations that apparently do not exist. The company then manually checked each flagged reference rather than relying solely on an AI detector.

The errors propagate into substantive claims. The report said predictive AI identified cocoa swollen-shoot disease with 99.1% accuracy in Colombia and Ecuador. Its cited paper studied cultivation in Africa, and the disease is not a significant problem in those Latin American countries. It ranked Ghana 47th and Uganda 74th in a 2023 government-AI index; the actual positions were 141st and 132nd, reversing the comparison. A claimed Southern African agricultural network and its Mozambique pilot could not be found, while a purported systematic review was supported by a broken citation and a short vendor case study.

GPTZero found similar problems in two papers by Accenture researchers: 31 allegedly hallucinated citations out of 245 references across the three publications. One paper linked readers to a fake GitHub repository impersonating the malicious chatbot FraudGPT; the repository contained real malware before it disappeared. That example turns sloppy citation from an epistemic problem into a security risk.

The attribution remains contested. GPTZero sells tools for detecting AI-written text and has an incentive to describe the pattern as machine-generated. It says its detector assigned high AI probabilities and that the citation errors match common model behaviour, but neither establishes authorship. Accenture says its own review found human error and no evidence of hallucination; it has not explained how it determined the cause. Accenture and IE University say confirmed errors will be corrected and safeguards strengthened.

That dispute should not become an escape hatch. A reliable publication process does not need to identify whether a false reference originated in autocomplete, a search tool or a person’s memory. It needs a verification gate that resolves every DOI and URL, matches title and authors, and checks that the source supports the associated claim. The embarrassing fact is not merely that a consultancy advising clients on AI may have used AI badly. It is that prestigious institutional signals—named experts, editors, a university partnership and a polished bibliography—substituted for opening the sources.