Accenture "Upskills" Its Hallucinations Game

Earlier this month, Accenture officially brought its tokenmaxxing era to a close by imposing a $29-a-month cap on many employees’ Claude accounts. While Accenture rejects this characterization, claiming the policy implements a new “funding approach” rather than a “firm limit,” it’s clear that the days of employees spending up to $1000 a month on LLM usage are over.

Unfortunately, a GPTZero investigation of recent Accenture publications reveals that not all tokens paid for by Accenture were spent productively. Using an automated pipeline and our Hallucination Detector, we surfaced three Accenture products containing a combined 31 hallucinated sources, numerous fabricated claims, and other likely signs of AI generation. To avoid making the same mistake as Accenture, a real human verified each hallucination.

The first document is a policy paper co-produced by Accenture’s African arm and professors from the Madrid campus of IE University’s Global Policy Center. It examines the potential for African economies to capture more value from agricultural exports like coffee, cocoa, and sugarcane using AI. The second is an academic paper posted on arXiv by researchers from Accenture’s Center for Advanced AI. The third is another arXiv paper posted by members of an Accenture research lab in Bangalore, India.

While neither arXiv paper is an official Accenture publication, both were authored by Accenture researchers using their Accenture emails and bylines on behalf of Accenture. Also, code from the first arXiv paper is being stored on the official Accenture GitHub organization.

One specific citation from the third paper shows how the risks of AI-generated citations extend beyond hallucinations. On page 16, a reference intended to corroborate a claim about a malicious chatbot called “FraudGPT” included a link to a fake GitHub repository for the chatbot that contained real malware. Readers of the paper who attempted to access the linked repository prior to its deletion sometime in 2026 may have infected their computers. This type of mistake — failing to discern between a credible source and a fraudulent website proliferating malware — is indicative of the unsupervised use of an AI research or citation tool.

The Accenture Arc

Accenture is one of the largest consulting firms in the world, offering IT and other professional services from between 52 and 120 countries (we burned some of our own Claude tokens to determine the real number is closer to fifty). In 2025, the firm bet heavily on AI by reorganizing its operations, shrinking its workforce, and announcing major deals with Anthropic and OpenAI. Accenture CEO Julie Sweet warned employees that their jobs were dependent on AI upskilling, while the Anthropic deal made Accenture one of Anthropic’s three biggest enterprise clients. By February 2026, the firm was monitoring staff use of AI tools — making promotion to leadership positions dependent on “regular adoption.”

Yet this ruthless pursuit of LLM-fueled productivity slowed as token spending increased. During the summer, leaked meeting audio revealed that Accenture was struggling to prevent non-technical workers from burning tokens on converting PDFs to slides and other trivial tasks. Accenture even launched a new service (called Tokenomics) to help other companies track, manage, and reduce ballooning token spend across their workforces. Yet by early October, the firm decided its best option was a $29-a-month cap on per-employee spending.

Before the cap was implemented, an unknown quantity of tokens was burned on inaccurate reports, fake statistics, and hallucinated citations.

A List of Non-Existent or Mangled Sources

GPTZero’s Hallucination Detector surfaced 31 hallucinated citations from 245 total citations across three Accenture products. As always, GPTZero uses a strict definition of hallucination and manually confirms that each citation meets this definition, which is calibrated to give authors the benefit of the doubt.

Another eight citations are flawed in ways that seem indicative of LLM generation, but do not meet our definition of hallucination. These include citations that add or remove 50% or more of the authors, that change or summarize subtitles, or can’t be verified against a specific source.

The above citations are not an exhaustive list. The “ContextEcho” paper alone contains another nine citations that drop authors from the corresponding source (or add new authors). Only one citation in the “Authenticity Debt and the Synthetic Content Threat Landscape” paper actually includes the source authors (citation 26). Also, “Leapfrogging Development” cites an anonymous post on a defunct blog that now resolves to a Thai gambling site (citation 31). It’s hard to imagine that a human author would consider this blog a reliable source for a policy paper.

We should also disclose that these three are not the only Accenture publications with hallucinations or other signs of AI-generated citations. For example, this collaboration between Accenture and the World Economic Forum contains multiple citations with paraphrased titles as well as ten citations resolving to ChatGPT-generated URLs (each contains “utm_source=chatgpt.com”).

Fake Claims

In most cases, hallucinated citations indicate the presence of fake claims. We define fake claims in two ways: 1. claims that aren’t corroborated by the cited source, or 2. claims that are factually wrong. Because verifying claims manually is extremely labor-intensive, we limited our analysis to claims from the “Leapfrogging Development” policy paper. The following is a list of claims we believe are fake, the corresponding citation (if given), and an explanation of our reasoning.

Changing the Game

Accenture isn’t the only company struggling with the financial and reputational costs of tokenmaxxing while reducing traditional human checks. Over the past year, GPTZero has identified hallucinations in reports published by all of the “Big Four” consulting firms: Deloitte, Ernst & Young, KPMG, and PwC. We’ve also used our Hallucination Detector to surface fake citations in academic papers submitted to prestigious machine learning / artificial intelligence conferences like ICLR and NeurIPS.

However, the results of this investigation indicate that even firms (like Accenture) that have made major investments in AI literacy and employee upskilling are still publishing AI-generated text and hallucinations. In other words, despite significant investment in AI use and training, some Accenture employees still skipped basic precautions like human review and verification. Even worse, this investigation reveals that hallucinations aren’t just appearing in low-quality Accenture products or promotional material, but also in academic publications and research papers. “Leapfrogging Development” is a comprehensive policy paper written in collaboration with a globally recognized university. “ContextEcho” is an academic article published by researchers from Accenture’s Center for Advanced AI, which claims to help other companies navigate the challenges and limitations of AI integration (including hallucinations).

As always, GPTZero contends that hallucinated citations are a clear and present danger to researchers, academics, consultants, students, and anybody else who happens to search the internet for information. Our previous investigations prove that hallucinated citations and fake claims published by reputable organizations spread passively across the internet, surfacing in AI summaries, chatbot responses, and LLM-generated research reports. Over time, these recycled hallucinations can “poison the informational well,” even if the original publication is never read by humans.

Our Hallucination Detector is our solution to this threat, as well as a way to hold consulting firms accountable for the research they produce. Try GPTZero’s Hallucination Detector for yourself, or reach out to GPTZero’s team.