vibe citing: how KPMG used AI to write a report about AI and AI made them look like fools

by t474-r0b07

There are companies that charge you to tell you how to use AI responsibly.

KPMG is one of them.

250,000 employees. 138 countries. Decades advising governments and corporations on how to avoid costly mistakes.

In October 2025 they published a report titled "Total Experience: Redefining Excellence in the Age of Agentic AI".

They wrote it with AI.

The AI invented 88% of the sources.

Nobody verified anything.

They published it anyway.

// what is "agentic AI" — because the title matters

An agentic AI is not a chatbot.

Not the assistant that answers your questions. It's a system that makes decisions and executes actions on its own, without a human approving each step. You give it an objective and it acts, corrects, moves forward.

It's the product everyone in the tech sector was selling in 2025.

KPMG was selling it too.

That's why they needed a report proving their clients were already using it.

Spoiler: they weren't. And the report invented it anyway.

// the forensic analysis

GPTZero — a company specialized in detecting AI-generated content — ran a full audit on the report.

First: what is an AI hallucination, because the term is going to come up a lot.

When a language model doesn't have the information you ask for, it doesn't say "I don't know." It generates a response that sounds correct. It invents with the same confidence it would use if it actually knew the truth. Perfect format. False content. No warning.

That's a hallucination.

Now the numbers from the KPMG report:

TOTAL CITATIONS:      45
REAL CITATIONS:        5
INVENTED CITATIONS:   40
ACCURACY RATE:      11.1%

40 of 45 citations have invented titles, authors that don't exist, or sources that don't say what KPMG claimed they said.

Half of the factual claims in the report are false or misattributed.

A firm that charges for intellectual rigor published a document with 11% accuracy.

// the organizations that read the report and said "that's not us"

The Financial Times contacted the companies listed as success stories.

Transport for London said the claims that they were using AI agents to predict congestion and coordinate the network were misleading.

NHS Greater Manchester said the description of using agentic AI to organize patient records and predict hospital readmissions "doesn't really align" with reality.

KPMG put their logos on fiction without asking permission.

And billed them as success stories.

// the error that best illustrates how the problem works

The model was instructed to find cases of companies using agentic AI.

It didn't find enough — because in many sectors they simply don't exist yet.

So it did the most comfortable thing: it generated them.

It cited a East Japan Railway press release from 2019 as evidence of agentic AI adoption.

The term agentic AI didn't exist in public discourse until 2024.

The model traveled five years back in time, reformulated an unrelated document, and presented it as proof of something that hadn't happened yet.

It wasn't an error. It was the easiest answer to the prompt.

The model doesn't understand the difference between inventing and remembering. It generates what fits. If it doesn't exist, it builds it. And it does so with the same fluency it would use to cite something real.

// vibe citing — the name the problem was missing

GPTZero coined the term: vibe citing.

To understand it you first need to understand vibe coding — writing code without understanding what it does. You ask an AI to generate the code, you copy it, it kind of works, and you move on without reading a line. The vibe is right. The understanding, zero.

Vibe citing is the same thing but with bibliography.

The model generates references that sound academic because it processed millions of papers. The structure is correct. The doi has the exact format. The year is right.

The content is fiction.

And the world's largest firm in responsible AI consulting didn't verify a single one before publishing.

def verify_sources(citations):
    # TODO: implement before publishing
    pass

publish_report()  # called without verifying anything

This is not a technical error.

It's a process decision. Or the absence of one.

// the moment the report contradicts itself

There's a detail that turns negligence into something almost poetic.

The report cites "KPMG research" claiming that 55% of CEOs prioritize AI as their main investment.

The KPMG 2025 CEO Outlook — published the same month, by the same company — says 71%.

The model didn't just invent external sources.

It invented data from the company that was using it and contradicted it with that same company's real data from the same period.

KPMG cited KPMG incorrectly in a KPMG report.

// Emirates case: three claims, zero correct on what matters

Page 42.

KPMG claims that Emirates adopted a mobile chatbot called Sara that can converse with passengers and change their flights.

Reality:

• Sara is a physical robot, not a chatbot.
• It was introduced in 2023, with no agentic capability.
• It cannot change flights.

Three claims. None correct on what matters.

The model took real information about Sara, reformulated it to fit the narrative it needed, and presented it as an agentic AI success story.

This is not a writing error. It's construction of fiction using real data as scaffolding.

// it's not just KPMG — it's the entire sector

This is where it stops being an isolated corporate scandal.

GPTZero has been documenting the same pattern for months:

• Deloitte — AI-generated content in a report paid for by the Australian government. Ended up refunding.
• EY — report with invented footnotes. Retracted in May 2026.
• KPMG — this case.

Three of the Big Four in consecutive months.

All selling responsible AI consulting.

All publishing hallucinations as research.

The pattern isn't coincidence. It's market pressure: the client wants the report, the report needs data, the data doesn't exist yet, the model generates it, nobody verifies because verification takes time and the client already paid.

AI is not the problem.

The economic incentive to appear to know more than you do — that's the problem.

// the feedback loop nobody is naming

Here's the data point almost no media outlet is discussing.

The false statistics from the KPMG report are already being reproduced by ChatGPT and Gemini.

I need to explain why that's structurally serious and not just anecdotal.

For months the report was published on KPMG's domains. The crawlers that feed language models index sources by authority. KPMG has maximum authority: global company, old domain, millions of visits, decades of institutional credibility.

The models ingested that content as verified truth.

Now when someone asks ChatGPT or Gemini about agentic AI adoption, they can return the false data from the report — not as "I found this at KPMG" but as their own knowledge, without attribution, without warning.

The full cycle:

model hallucinated data
    → KPMG published without verifying
        → crawlers indexed it as high-authority source
            → other models ingested it as truth
                → user receives the original hallucination as fact

high-authority source + false data + model ingestion = untraceable disinformation.

You can't trace the origin. You can't disinfect the source. The error already lives inside the models you consult every day.

And the report has already been retracted. But the data keeps circulating.

Taking down the PDF didn't deindex anything.

// what KPMG said afterward

KPMG's spokesperson declared after withdrawing the report:

"We expect all our staff to follow our guidelines on responsible AI use, including human oversight to validate content and verify independent sources."

Translation: we have guidelines. Someone didn't follow them. We're investigating.

What they didn't say: how a flagship report on responsible AI, with the KPMG logo, published on their official channels, passed through their entire internal review process without anyone verifying a single one of the 45 citations.

250,000 employees.

5 valid citations.

Nobody asked anything.

// conclusion — the problem isn't technical

Models do exactly what they were designed to do: generate coherent and plausible text based on learned patterns.

They don't lie. They have no concept of lying. They generate what fits.

The problem is human: using AI as a researcher without a verification loop isn't efficiency. It's delegating truth to a system that has no concept of truth, and signing your name on top.

KPMG didn't build a report with AI.

They built the appearance of a report and sold it as research.

The difference isn't semantic.

It's the difference between knowing something and appearing to know it.

In 2025, the world's largest firms chose to appear.

primary sources — verify yourself:

• GPTZero Investigation (full report): https://gptzero.me/news/investigations-kpmg/
• TechCrunch: https://techcrunch.com/2026/06/13/kpmg-pulls-report-on-ai-usage-due-to-apparent-hallucinations/
• The Register: https://www.theregister.com/ai-and-ml/2026/06/12/kpmgs-ai-report-turns-into-a-demo-of-ai-hallucinations/5255029
• CityAM: https://www.cityam.com/kpmg-report-on-ai-found-riddled-with-ai-hallucinations/