KPMG Withdraws Report Over Alleged AI 'Hallucinations' and Misstated Usage Claims
Table of Contents
You might want to know
Could the use of AI in composing analyses produce false assertions about organizations' practices?
What safeguards are necessary to prevent AI-generated reports from presenting inaccurate or fabricated citations and claims?
Main Topic
Professional services firm KPMG removed a report titled "Redefining excellence in the age of agentic AI" from its public sites after multiple organizations disputed statements in the document describing their use of AI. The report, which had been published in October 2025, was found to contain several assertions that representatives of the named institutions characterized as incorrect or misleading. An independent research group, GPTZero, identified a number of inaccuracies and attributed them to what it described as AI "hallucinations."
The situation highlights a growing tension: firms are increasingly using AI tools to analyze and synthesize information, yet those same tools can produce confident-sounding but incorrect content. In this instance, major organizations such as UBS, the UK's National Health Service, Swiss Federal Railways, and Transport for London informed press outlets that the report's statements about their AI adoption or usage did not align with reality. Their pushback prompted KPMG to temporarily withdraw the document while it investigates the source of the errors.
KPMG issued a statement noting that the report had been taken down pending an internal review. The firm emphasized that employees are expected to adhere to established guidelines for responsible AI use, which include maintaining human oversight to validate content and verify independent sources. The comment underscores a broader responsibility: when AI is used as a drafting or research aid, organizations must ensure that final outputs are corroborated through human review and reliable references.
The episode is not unique to KPMG. The previous month, another professional services firm, EY, withdrew a separate study on loyalty rewards programs after observers found what appeared to be fabricated footnotes and other inconsistencies consistent with AI-generated hallucinations. Taken together, these incidents illustrate how generative AI can introduce factual errors in reports that mix analysis, quotation, and attribution.
From a methodological perspective, the risk arises when AI tools are given broad latitude to generate descriptive or inferential content about third parties without rigorous verification. Large language models are trained to produce plausible-sounding text but do not have an infallible fact-checking mechanism or direct knowledge of recent, institution-specific practices unless those are explicitly corroborated. Without deliberate cross-checking, claims about named organizations, their policies, or their technological deployments can be mistaken, exaggerated, or entirely fabricated.
Addressing this risk requires a combination of procedural controls and technical safeguards. Procedurally, authors and reviewers must treat AI outputs as drafts or research aids rather than finished work—verifying every factual assertion, attribution, and citation against primary sources or direct confirmation. Technically, organizations can use provenance tools, metadata tracking, and model settings that limit hallucination-prone generation. Equally important is training staff to recognize likely AI errors and to maintain a chain of accountability for published materials.
The reputational consequences of publishing unverified AI-generated content can be significant. When reports inaccurately describe the behavior of institutions or misattribute practices, they can harm relationships with stakeholders, misinform public debate, and require time-consuming corrections. For professional services firms whose credibility rests on careful analysis and trust, the imperative to avoid such errors is particularly strong.
In summary, the withdrawal of KPMG's report—and the prior EY incident—serves as a reminder that while AI can accelerate research and drafting, it cannot replace the need for thorough human validation. Ensuring factual accuracy demands explicit verification procedures and clear accountability for final outputs.
Key Insights Table
| Aspect | Description |
|---|---|
| Incident | KPMG removed an October 2025 report after several organizations disputed claims about their AI usage. |
| Cause Identified | Research group GPTZero found inaccuracies attributed to AI "hallucinations," suggesting generative models produced incorrect statements. |
| Organizations Affected | UBS, the UK National Health Service, Swiss Federal Railways, and Transport for London reported the claims were untrue or misleading. |
| Firm Response | KPMG withdrew the report for investigation and reiterated guidelines requiring human oversight and source verification when using AI. |
| Broader Trend | Similar withdrawals (e.g., EY) show a pattern where AI-assisted reports may include fabricated citations or errors if not thoroughly checked. |
Afterwards...
Looking ahead, organizations should invest in both governance frameworks and technical tools to reduce the risk of AI-generated inaccuracies. This includes developing robust editorial standards, mandatory verification steps for AI-derived claims, and training programs that help staff identify and correct hallucinations.
On the technical front, advancing research into model interpretability, provenance tracking, and hybrid workflows that combine retrieval-augmented generation with human validation will be valuable. Equally important is ongoing dialogue between AI developers, professional institutions, and regulators to define acceptable practices and accountability mechanisms.
Ultimately, the responsible deployment of generative AI in report-writing requires a culture that treats AI as a tool—not an authoritative source—and that preserves human responsibility for accuracy and integrity.