Accenture says human error, not AI, put bad citations in its research, but GPTZero, a company that sells AI-detection software, says three Accenture publications contain 31 hallucinated sources. A hallucination here means a reference an AI system invented.
The most prominent paper is Leapfrogging Development: Unlocking Africa's Potential in Global Value Chains with Artificial Intelligence, written with IE University in Madrid. GPTZero found 15 inaccuracies in it, plus four more flawed citations. Those are references that do not exist, exist under a different name, or do not support the claim they are attached to. Most of the flagged problems sit in that Africa paper, according to coverage of the findings.
Across all three documents, GPTZero said it found 31 hallucinated sources, numerous fabricated claims and other likely signs of AI generation. It said it used an automated pipeline, then had a person verify each hallucination.
An Accenture spokesman said the errors were human mistakes and that there is no evidence they came from AI. The company said the paper is being corrected ahead of a re-release.
The dispute is hard to settle from outside, because a citation to nothing looks the same whether a person or a model produced it. GPTZero also has a commercial interest in finding the problem. Still, the stakes it describes are concrete: fake claims in a document from a well-known organization can be recycled and repeated, and GPTZero warned they can poison the informational well even if nobody reads the original.
This is not the first such case. GPTZero made similar findings about reports from PwC in August 2026. Accenture's version has an awkward backdrop, since the firm has been pushing its employees to use AI more productively. The papers were collaborations produced outside its usual in-house process, so the failure may sit in that handoff rather than in any one tool.
The re-release may settle more than the denial does. Which citations get dropped or replaced will show how many were broken in the first place.