KPMG pulls AI report: hallucination and data reliability risk

The timeline: what happened with the KPMG report

On June 13, 2026, TechCrunch has reported that KPMG has withdrawn a report dedicated to the use of artificial intelligence in organizations. The official reason: apparent hallucinations in the content generated by the models used in the research.

Therefore, data, statistics, and claims in the document turned out to be potentially unverifiable or outright wrong. KPMG chose to pull the report from circulation rather than publish a corrected version. This decision, as bold as it was for transparency's sake, raised deep questions.

In fact, this is not a company that is new to AI. KPMG is one of the Big Four of global consulting. Yet, even an organization with technical and human resources of that caliber put out a document compromised by generative errors. Consequently, the issue is not about competence, but process.

Hallucination: an operational definition for non-AI professionals

Language model hallucinations are statements generated with apparent confidence but lacking any real foundation. The model isn't lying on purpose. It simply churns out statistically plausible outputs that might not actually match the facts.

This phenomenon has been documented for years in technical literature. For example, MIT Technology Review had already taken a deep dive into how this works back in 2023. Even so, lots of organizations keep using AI results without proper check-up steps.

Specifically, the risk increases when models are asked about quantitative data, industry statistics, or specific sources. Therefore, the KPMG case is not unusual: it is the predictable consequence of a content production process that lacks a structured fact-checking layer.

The winners and losers of this episode

Every time something like this happens, it shakes up who is trusted in the market. So, it's worth looking at who gains ground and who loses out after the fact.

Who loses: KPMG suffers direct reputational damage, even if limited in time. Beyond this, the entire segment of research reports produced with generative AI loses credibility in the eyes of those who use them as a secondary source. Similarly, the involved AI model vendors—even if not explicitly named—see discussions about the reliability of their systems intensify.

Who wins: organizations that adopt hybrid approaches — AI assisted by expert human review — come out stronger from episodes like this. Furthermore, providers of AI governance and generative content auditing solutions find a powerful sales pitch in this case. In short, KPMG's transparency in withdrawing the report, while costly, sets a positive precedent for AI crisis management.

Why Italian SMEs cannot ignore this warning sign

A common mistake is thinking that slip-ups like this only happen to huge corporations with massive tech budgets. Actually, smaller businesses are often more vulnerable because they have fewer resources to double-check their AI-generated stuff.

Many mid-sized Italian companies today use AI tools to produce competitive analysis, industry reports, marketing content, and even commercial documents. However, these outputs are rarely subjected to systematic verification before distribution. As a result, the risk of publishing or sharing incorrect information is real and underestimated.

We at SHM Studio we regularly see this dynamic in our work with clients. Specifically, the problem pops up most often in sectors where market data changes fast: retail, manufacturing, and highly specialized B2B services. That's why AI content governance isn't just a theory topic: it's an immediate operational must-have.

SHM Studio's read: a process architecture problem

In our view, the KPMG case is not a failure of AI as a technology. It is an architectural failure in the process that led to the report's publication. So, the right question isn't "is AI reliable?" but rather "is our production and verification process up to scratch?".

There are at least three levels of control that every organization should integrate when using generative models to produce content with reputational or decision-making implications. First of all, verifying the primary sources cited by the model. Next, a human review layer with specific expertise in the subject area. Finally, a formal approval process before external distribution.

These three levels don't eliminate the risk, but they cut it down a lot. Plus, writing this process down keeps the organization safe if anyone argues about it later. The AI consulting by SHM Studio includes precisely the design of these operational workflows for Italian B2B contexts.

Implications for those using AI in marketing and communication

Marketing is one of the areas where the adoption of generative AI tools has grown the fastest over recent years. Because of this, the takeaways from the KPMG case directly impact communication and content marketing roles at small and medium-sized businesses.

For example, a company using AI to produce SEO copywriting , newsletters, or industry white papers must take into account that every automatically generated quantitative claim is potentially at risk. Similarly, competitive analyses produced with AI to support LinkedIn campaigns or google ads campaigns require data verification before use.

By the way, the issue of AI content credibility is also becoming relevant for organic positioning. Google's guidelines increasingly emphasize direct experience and the authority of sources. As a result, publishing content with incorrect AI-generated data can damage both reputation and organic visibility. To learn more, the section SEO of SHM Studio addresses these topics specifically for the Italian context.

What nobody is saying: the AI that talks about itself

There is one aspect of the KPMG case that deserves a separate thought. The withdrawn report was specifically about the use of AI in organizations. So, an AI model produced wrong information about itself—or rather, about its own tech category.

This is not a minor detail. It shows that language models struggle in particular when asked about up-to-date data, recent adoption stats, or fast-changing industry benchmarks. In fact, training data always has a cutoff date. Therefore, any quantitative claim about recent trends—like AI adoption in 2025 or 2026—is structurally at high risk for hallucinations.

Research like the one conducted by McKinsey on the Global AI Survey show how quickly adoption data changes. Therefore, using an AI model to cite statistics on these very trends is a high-risk exercise without external verification.

Next moves: what organizations using AI should do now

The KPMG incident gives us some practical tips for organizations that have already plugged AI tools into their workflows. Here are the main things we at SHM Studio we recommend to our B2B contacts.

• Audit of existing processes: map out where and how generative models are being used, keeping a close eye on stuff that gets shared with the outside world.
• Introduction of a verification layer: every number or stat whipped up by AI needs to be double-checked against original sources before it goes live.
• Internal training: teams using AI must understand the hallucination mechanism and know when the risk is highest.
• Process documentation: formalizing the production and review workflow protects the organization and improves quality over time.
• Review of tech partnerships: evaluate whether the AI vendors used offer tools for grounding, source citation, or integrated output verification.

Also, it's worth considering that European AI regulations — especially the AI Act — are introducing transparency and accountability requirements that will make these processes not just recommended but mandatory for certain use cases. For this reason, investing in AI governance today is also an investment in future compliance. The section Digital marketing and the one dedicated to web services by SHM Studio include consulting on these aspects for Italian SMEs. To learn more, you can contact our team or explore other articles on SHM Studio blog .