In mid-June 2026, KPMG withdrew one of its own reports on the use of artificial intelligence. The reason: the document contained data apparently generated by AI systems with obvious hallucinations. As a result, the published information was neither verifiable nor reliable.
This episode is not isolated. It is a systemic signal for every organization that uses AI tools in enterprise analysis. And the problem doesn't affect only large consulting firms. Italian SMEs that rely on unsupervised AI output also risk building strategic decisions on wrong data. As a result, governing AI-generated content becomes an operational priority, not a theoretical issue.
At SHM Studio, we follow these developments closely. We work with Italian SMEs to integrate AI into marketing and communication processes in a controlled, verifiable way. This KPMG case offers a concrete prompt to review internal workflows and strengthen fact-checking on every piece of AI-generated output.
Timeline of the KPMG case
On June 13, 2026, TechCrunch reported that KPMG had withdrawn a report on the use of AI. The document, aimed at an enterprise audience, contained data and claims that could not be verified. The identified cause: hallucinations produced by the artificial intelligence systems used to write it.
The withdrawal happened quietly, but the news spread quickly through the industry. KPMG is one of the Big Four global consulting firms, so a mistake of this magnitude carries real symbolic weight. This isn't a startup experimenting: it's an organization with structured internal processes.
Still, the story won't surprise anyone who closely follows how language models are evolving. Generative AI systems tend to produce plausible output even when there is no real data behind it. This phenomenon, known as hallucination, has been documented in the technical literature for years.
Winners and losers in this story
Who comes out of this episode damaged? First of all, KPMG itself. A consulting firm's reputation rests on the quality and accuracy of its analysis. A report withdrawn for AI hallucinations weakens the trust of institutional clients.
Second, the entire enterprise AI vendor ecosystem takes a hit. Many of these tools are sold as reliable solutions for producing reports and analysis. As a result, episodes like this fuel skepticism among decision makers. Some organizations may even slow down their adoption plans.
Who can benefit from this situation? Companies and agencies that have built human oversight into their AI output. Professionals who promote a hybrid approach, with AI as the tool and humans as the validators, also see their position strengthened. Likewise, AI governance frameworks now carry a concrete relevance that was more theoretical until yesterday.
Why hallucinations hit analytical reports in particular
Large language models don't reason: they generate text that is statistically consistent with the prompt they receive. So when they are asked about specific data, such as percentages, statistics or research, they tend to produce plausible numbers even when no real sources support them.
This mechanism is especially insidious in analytical reports. In a document like this, an invented but correctly formatted figure can slip past a superficial review. If the reviewer doesn't know the original source, checking becomes very difficult.
According to Gartner research on Large Language Models, managing hallucinations remains one of the main challenges for enterprise adoption of generative AI. So this is not a residual problem that will disappear with the next model updates.
On top of that, the problem grows when AI is used to analyze data about AI itself. In this case, the models draw on a training corpus that may contain contradictory or outdated information, or content generated by other AI systems. The result is a potentially self-referential cycle of errors.
The SHM Studio take: a governance problem, not a technology problem
At SHM Studio, we read the KPMG case as a process problem, not a tool problem. Generative AI is not inherently unreliable. But it becomes dangerous when it's inserted into workflows with no verification checkpoints.
The right question isn't
Related articles
Discover more articles exploring similar topics, selected to offer you a more complete and stimulating perspective. Each piece of content is carefully chosen to enrich your experience.
