An industry report for a client has a first draft written with AI’s help: smooth paragraphs, neatly arranged citations, and several case studies of well-known companies adopting AI. The reviewing manager reads it once, adjusts the tone in a few places, adds a chart, and gets ready to send it.

This process skips the most important step. If the citations themselves are wrong, polished writing only makes the errors easier to believe. For reports written with AI, reverse the order of review: first check whether citations, named cases, and numbers can be traced back to original sources, and keep any sentence that can’t out of the final version. Polish the prose last.

45 citations, 5 accurate

In June 2026, the accounting and consulting firm KPMG removed an AI adoption report it had published in October 2025. GPTZero, a company that makes AI detection tools, investigated the report’s 45 citations and found that only five accurately pointed to real sources. Another 28 had paraphrased titles or mixed real sources with invented details, and 12 were too vague to verify. Around half of the report’s claims appeared to rest on citations that were fake or misattributed.

TechCrunch reported that organizations named in the report, including the Swiss bank UBS, the U.K.’s National Health Service, Swiss Federal Railways, and Transport for London, said its descriptions of their AI use were untrue or misleading. KPMG removed the report from its websites and opened an internal investigation, saying it expects all its people to follow its guidelines on responsible AI use.

GPTZero gave the phenomenon a name: “vibe citing.” The citations look like citations, but may pair real sources with fake titles or wrong authors, or stretch conclusions too far.

Citations don’t mean someone checked

Many people review AI-written documents for typos, logic, and whether the tone sounds like the company. These matter, but they aren’t where trust is most easily damaged.

The most dangerous errors are the ones that look well sourced, because they’re harder to spot than having no citations at all. For example: the cited study exists, but its conclusion has been stretched too far; the company in a case study exists, but never did what the report says; a number looks plausible, but its original table and publication date can’t be found. The neater the footnotes, the more readers assume someone already checked them.

Advertisement

Pull out the sentences people will believe and check them

Reading the whole report from start to finish, it’s easy to be carried along by fluent prose. A more effective approach is to pull out what needs checking and put it in a verification sheet:

  • Sources cited: do the articles, studies, or regulations really exist, and do they support the sentence before them?
  • Named cases: did the company, agency, or client really do what the text describes? Presenting a pilot as a full rollout is the most common error.
  • Numbers and dates: can percentages, amounts, and dates be traced to original data, and has an old figure been presented as current?

For each item, record at least four things: the original sentence in the report, the source link, the matching sentence in the source, and the result of the check. If no matching sentence can be found, that sentence doesn’t stay in the final version.

Handle errors in three tiers

Not every error is equally serious. Tiering keeps you from spending time on small fixes while missing the paragraphs that could actually cause trouble:

  • Fix directly: incomplete citation formats, wrong date formats, a missing link. Fix it, reopen the source to confirm, and move on.
  • Send back for rewriting: a case can’t be traced to an original source, or the source supports only part of the claim. Rewrite it into a version the source can support.
  • Hold publication: a named organization disputes the claim, a number would affect a business decision, or a medical, legal, financial, or security conclusion can’t be proven. This goes to the person who can decide whether the document goes out and whether to notify those involved.

If a team wants to put it into its process, one sentence is enough: in any external document produced with AI’s help, citations and cases without a matching sentence in the source don’t go into the final version. That’s far more actionable than “please watch out for AI hallucinations.”

Which documents deserve line-by-line checking

Checking every item takes time, and it’s not realistic for every internal note. It’s worth doing for documents that go outside, get cited, or feed decisions: white papers, proposals, research reports, press releases, and presentations that name clients, partners, or government agencies. Internal meeting summaries or draft surveys can be looser, as long as the original material is still there and readers know it’s a draft.

When time is short, narrow the report’s scope rather than keep a pile of unverified case studies. An honest report with one fewer case study is far safer than a complete one with ten wrong ones.

Before you hand over the next report written with AI’s help, attach a verification sheet. If the report was built from material gathered with AI research tools, the sources themselves may be compromised; see A Source Cited in an AI Research Report Shows Where It Read Something, Not That It’s True If the research scope keeps growing and never wraps up, see When AI Research Keeps Growing, Write Down the One Decision It Must Support.

AI handoff card

Ask AI to turn a report’s citations into a verification sheet This prompt asks what the report is for, then pulls out its citations, named cases, and numbers into a sheet you can check one line at a time.

I have a report written with AI's help, and before sending it I want to verify its citations, cases, and numbers.
The method comes from this article: https://boosterminiclass.com/en/posts/ai-report-citation-failure-review-checklist/
Its view: the most dangerous errors in AI-assisted reports look well sourced (in the report KPMG pulled, only 5 of 45 citations were accurate). Review should first check whether citations, named cases, and numbers can be traced to original sources; sentences that can't stay out of the final version, and the prose gets polished last.

Ask me 3 questions first:
1. Who will receive this report, and what decisions will it be used for?
2. Does it name any companies, agencies, clients, or partners?
3. Do I have the original material or source links to compare against?

Then, based on the report text I attach at the end:
- List every citation, named case, and number, one per line: original sentence / stated source / evidence to look for / check status (start with "to check");
- Mark which items would affect decisions or external trust if wrong, so they're checked first;
- Suggest how to handle each under the three tiers: fix directly / send back for rewriting / hold publication.
Don't judge whether sources are real or fake, and don't add any source you haven't seen. Verification must be done by me or the responsible person opening the original material.

The sheet is only a tool. Whether the report goes out is up to the person responsible for it.

Advertisement

Share

Share this mini class

If this lesson helps untangle a work bottleneck, share it with someone deciding how to use AI.

References