A report can look highly professional: complete headings, polished paragraphs, neatly arranged citations, and even case studies from major organizations. You may assume the next step is simply to adjust the tone, add charts, and send it to a manager for approval.
But if the citations themselves are wrong, polished writing only makes the error easier to believe.
In June 2026, KPMG pulled its previously published AI adoption report, Redefining excellence in the age of agentic AI. Several media outlets cited GPTZero’s investigation, which found that many citations and case studies in the report could not be matched correctly to their sources. TechCrunch also reported that organizations including UBS, the UK’s NHS, Swiss Federal Railways, and Transport for London said the report’s descriptions of their AI use were incorrect or misleading. In response, KPMG said it had removed the report and opened an investigation, while emphasizing that responsible AI use requires human oversight, content validation, and confirmation from independent sources.
This is not merely a story about a consulting firm making a mistake. For everyday workers, the more useful reminder is this: whenever AI-assisted content becomes a report, slide deck, white paper, proposal, or public article, review should not focus only on whether the tone reads well. It must first confirm whether the citations, cases, and accountability can be traced back to original evidence.
The most dangerous mistake in an AI report may not be in the main text
When many people review AI-generated documents, they check three things first: typos, logic, and whether the tone sounds like the company. Those checks matter, but they are not where trust is most easily damaged.
The real trouble comes from mistakes that “look sourced.” For example:
- A cited article or study really exists, but the title, author, year, or conclusion has been rewritten until it is distorted.
- A company in a case study really exists, but it never did what the report claims.
- A number looks reasonable, but no one can find the original table, research method, or publication date.
- The footnotes are numerous, but nobody has clicked through to confirm whether they support the preceding sentence.
GPTZero described this kind of problem as something like “vibes-based citation”: the citation looks like a citation, but in practice it may mix real sources, fake titles, wrong authors, and overextended conclusions. For readers, this is harder to catch than having no citations at all, because the document already appears to have completed its fact-checking.
So once AI becomes part of the reporting process, do not only ask, “Does it sound right?” A better question is: “Which sentences in this document will people believe because they see a citation or case study attached to them?”
First split the report into three types of verifiable material
If you read the entire report from beginning to end, fluent writing can easily pull you along. A safer method is to extract the material that needs verification first.
| Material type | What to check | Common risk |
|---|---|---|
| Cited sources | Whether the article, study, report, regulation, or official document really exists, and whether it supports the previous sentence | The source exists, but AI has pushed the conclusion too far; the citation format looks real, but the details are wrong |
| Named cases | Whether the company, government agency, client, or product really did what the text describes | A pilot is described as full deployment, or one feature is described as an entire workflow |
| Numbers and timing | Whether percentages, amounts, user counts, publication dates, and version names can be traced to original material | Old figures are treated as the latest status, or data from different markets is mixed together |
The purpose of this table is not to turn everyone into a researcher. It is to make the report pause before it is sent out. If a paragraph may affect budget, procurement, contracts, customer trust, or public reputation, it cannot rely only on AI’s version of “seems reasonable.”
A simple approach is to ask the person responsible for the document to copy every citation, company name, and number into a verification sheet. Each row should have at least four fields: original sentence, source link, matching sentence in the source, and verification status. If there is no matching sentence, do not leave that claim in the final document.
Use three levels to decide: fix the wording, send it back, or do not publish
Not every mistake is equally serious. Risk needs to be layered so the team does not spend all its time on minor edits while letting genuinely dangerous passages through.
| Level | Example | Next step |
|---|---|---|
| Fixable | Citation format is incomplete, date format is wrong, or the source link needs to be added | Correct it, then click through and confirm again; the whole report does not need to be sent back |
| Must be rewritten | The original source for a case description cannot be found, or the source supports only a small part of the claim | Send it back to the author or AI generation process and rewrite it into a version the source can actually support |
| Cannot publish | A named organization denies the claim, a number affects business judgment, or healthcare, legal, financial, or cybersecurity conclusions cannot be proven | Pause publication and ask the accountable owner to decide whether to remove it, redo it, notify relevant parties, or keep a correction record |
The “accountable owner” here is not merely the nominal document owner. It is the person who can decide whether the document goes outside the company, whether customers need to be notified, and whether the team will absorb the cost of corrections. AI can help organize material, but it cannot carry the trust loss caused by bad citations on behalf of the company.
If you are responsible for finalizing reports in a team, add one sentence to the workflow: For any external document assisted by AI, citations and case studies without matching source text do not enter the final version. That sentence is far more actionable than “please watch out for AI hallucinations.”
When AI should not write directly into the final version
Some documents can safely use AI to organize a first draft before a person makes quick adjustments. Examples include internal meeting summaries, draft survey questions, and interview-note synthesis. As long as the original material remains available and readers know the output is a draft, the risk is relatively manageable.
But the following situations are not suitable for letting AI slide all the way from first draft to final version:
- External white papers, research reports, consulting proposals, or press releases.
- Documents that name customers, partners, government agencies, or competitors.
- Slide decks used to support procurement, investment, budgeting, layoffs, or compliance decisions.
- Numbers and conclusions that media, customers, managers, or legal teams may cite again.
AI can help with the first round of organization for these documents, but two checkpoints must sit in the middle: source confirmation and accountability confirmation. Source confirmation asks whether every key sentence can be traced back to original evidence. Accountability confirmation asks who has the authority to decide that the sentence can be said externally.
If time is short, it is better to narrow the report’s scope than to keep a pile of polished but unverified case studies. An honest report with fewer cases is usually safer than a complete-looking report full of incorrect ones.
Before delivering the next AI report, add one verification page
The most practical change is not to ask everyone to “be more careful.” It is to add one verification page to the document handoff package. The page does not need to be complicated. Each key claim only needs five signals:
- Key sentence: the sentence that may be cited, repeated, or used for decision-making.
- Original source: the official document, research, news article, contract, internal data, or interview record.
- Matching evidence: the paragraph, page number, or screenshot location in the source that truly supports the sentence.
- Verifier: the person who actually opened the source and confirmed it.
- Publication judgment: keep, rewrite, delete, send back for rewriting, or pause publication.
This verification page changes AI report review from “does it read smoothly?” to “can it stand up?” It also makes responsibility clearer: AI is not the author, reviewer, or publisher. It is only the assistant that helps organize material. The team is still the one that ultimately sends the content out.
KPMG’s case is a reminder that AI hallucinations do not only appear in chat windows. They can be packaged as reports, footnotes, and case studies, then enter real business decisions. The next time you see a polished document helped along by AI, do not rush to praise its efficiency. Ask the more important question first: can these citations and cases actually be verified?
Everyday four-panel comic
- A polished AI-generated report looks ready to send, with neat paragraphs, charts, and footnotes.
- Before sending, the reviewer pulls out every citation, company name, number, and case study into a check sheet.
- Some claims survive because the source supports them; others are rewritten or removed because the evidence does not match.
- The lesson is simple: an AI report is not finished when it reads well. It is finished only when its citations and responsibility can stand up.
AI handoff card
Ask AI to organize this article's specific situation
Copy this into your own AI chat tool to turn this mini class into a personal checklist. BMC will not see what you paste into your AI tool.
References
- TechCrunch: KPMG pulls report on AI usage due to apparent hallucinations — https://techcrunch.com/2026/06/13/kpmg-pulls-report-on-ai-usage-due-to-apparent-hallucinations/
- GPTZero: Chasing the Hallucinations: KPMG’s AI-Powered Attempt at “Redefining Excellence” — https://gptzero.me/news/investigations-kpmg/
- The Register: KPMG’s AI report becomes an accidental demo of AI hallucinations — https://www.theregister.com/ai-and-ml/2026/06/12/kpmgs-ai-report-turns-into-a-demo-of-ai-hallucinations/5255029
- Finance Monthly: KPMG Pulls AI Report After UBS, NHS and Transport for London Case Studies Found to Contain Hallucinations — https://www.finance-monthly.com/kpmg-ai-report-hallucinations-ubs-nhs-transport-for-london/
