Accenture accused of publishing AI paper littered with AI hallucinations

Accenture co-authored a report about artificial intelligence that was littered with false citations, incorrectly attributed research and case studies that appear not to exist, prompting accusations that the consultancy published information marred by AI hallucinations.

A paper about the potential of AI in African agriculture, published by the IT consulting firm and a Spanish university, contained several factual errors and 15 citations with errors such as incorrect titles or author names, according to an investigation by AI detection business GPTZero.

Edward Tian, co-founder of GPTZero, said the errors, which were checked by the FT, were “overwhelmingly consistent with the patterns of AI hallucinations rather than the occasional inaccuracies associated with human research”.

Accenture, which was named in September as an independent evaluator to assess the safety of Anthropic’s AI models, said there was “no evidence” that the errors were the result of AI hallucination but subsequently added that it would make corrections to the publication.

Accenture said: “Our review indicates that the errors identified in the report were the result of human error. There is no evidence to support the conclusion that they were generated through AI hallucination.”

The consultancy declined to say how it reached this conclusion or provide further details of how it concluded the errors were made.

It said the corrected report would be re-released, adding: “We will always embrace the opportunity to strengthen the quality of our reports.”

The November 2025 report — titled “Leapfrogging Development: Unlocking Africa’s Potential in Global Value Chains with Artificial Intelligence” — was published by Accenture and IE University. Two of its four authors and both of its editors are staff from Accenture, which has revamped its business to make heavy use of AI and advise clients on how to adopt the technology.

The paper, published on Accenture’s website, made two references to technology hosted by an organisation called SAATN — identified as both the “Southern Africa AgriTech Network” and the “Southern African Agriculture Transformation Network”. Neither network appears to exist.

It also claimed predictive AI had “significantly improved” the detection of swollen shoot disease among cocoa plants, reaching 99 per cent accuracy in Colombia and Ecuador. The source cited for this claim, however, was an article focused on west Africa without any mention of these South American countries, where there is minimal evidence to suggest the disease is present.

Other errors included incorrectly stating the rankings of Ghana and Uganda in a 2023 Oxford AI readiness index, providing a broken web link as citation.

Several citation-related errors were identified in the report. Tian said that “human citation errors tend to be small and isolated”, such as incorrect years, formatting or a broken link. “By contrast, LLMs generate references that sound entirely credible but do not exist. This results in fabricated paper titles, plausible-sounding but fake organisations, and invented case studies stitched to real actors,” he said.

The Accenture paper stated that “youth- and women-led digital platforms have achieved high uptake” in Colombia and Honduras “when embedded in [an] inclusive governance structure”. Neither of the citations provided for this claim appears to exist.

The 49-page report also contains web links that do not follow through to the source article cited. One citation misspelt the name of the author and listed two authors who were not named as writers on the source paper.

Accenture has heavily encouraged its staff to use AI — developing training courses and making the use of AI a criterion for deciding promotions — at a time when investors are watching to see if the technology will improve revenue per employee. A round of lay-offs last year was concentrated among staff who could not be retrained to make use of the technology.

AI detection tools, including GPTZero’s, are not foolproof and have been criticised for wrongly flagging human-written content as AI-generated in some cases.

GPTZero said it had manually checked each of the errors flagged by its tool in the Accenture paper and made additional checks with open-source intelligence. It has previously identified AI hallucinations in research by PwC, KPMG and EY, with the latter two retracting reports following its findings.

IE University said: “The authors of the report approach their work and analysis with the utmost rigour. Should any errors be confirmed, they will make the necessary corrections to the report and strengthen the processes in place to help prevent similar issues from occurring in the future.”

It added that the authors of the report were “reviewing the sources and data underlying the report to understand how these discrepancies may have arisen”.