The scale of the fabrication: 89% of citations were flawed

GPTZero's scan of the report's 45 citations revealed a systematic failure of verification :

• Only 5 citations accurately pointed to real, relevant sources.
• 28 citations had paraphrased or altered titles, with fake components grafted onto real publications.
• 12 citations were too vague or inconsistent to trace back to any existing source.
• Approximately half of the report's substantive claims that relied on these citations appeared to be fake or misattributed.

Overall, 89% of the bibliography was flagged as flawed. GPTZero's AI detection scan added another troubling layer, rating the document 56% "mixed," which suggests it was likely AI-generated or heavily AI-assisted content that evaded meaningful human review .

Who was falsely cited, and what did the report claim?

The investigation identified four prominent organizations cited with fabricated case studies of "agentic AI" deployment :

• UBS: The report claimed the bank was deploying agentic AI for personalized wealth management insights, but no verifiable source existed.
• NHS Greater Manchester: It asserted the NHS trust was using AI-driven patient journey orchestration, a claim unsupported by real documentation.
• Transport for London (TfL): TfL was cited as an example of AI-powered real-time customer experience optimization, but the claim could not be corroborated.
• Swiss Federal Railways (SBB): The report presented a case study of SBB using agentic AI for integrated customer touchpoints, yet no such initiative could be verified in the cited material.

In each case, the citations pointed to non-existent reports, paraphrased titles of unrelated real publications, or references so vague they were functionally worthless .

The root cause: AI tools run amok without verification

GPTZero's analysis concluded that the errors were likely the result of an AI research tool over-complying with a prompt to find real-world examples of "agentic AI." The tool generated plausible-sounding but entirely fictitious case studies and citations, which then slipped past KPMG's review process uncorrected .

The incident underscores a critical vulnerability: when AI-generated research is published without rigorous human verification, it can produce confident, authoritative-sounding nonsense that damages institutional credibility .

A growing pattern of AI hallucinations in professional services

KPMG's pulled report is far from an isolated incident. It fits into an expanding pattern of AI-related citation failures across major consulting and legal firms :

• Deloitte Australia had to reissue a $440,000 government report after it was found to contain fictitious academic references and a fabricated quote from a federal court judgment . The firm issued a partial refund and faced public scrutiny over what one observer called a report "littered with citation errors" .
• Ernst & Young (EY) produced a report on loyalty fraud where GPTZero's investigation found that almost all URLs were broken or fake, and more than half of the titles did not correspond to real sources .
• Similar issues have surfaced in legal filings involving major firms like Sullivan & Cromwell .

These repeated failures highlight a systemic risk: without rigorous human oversight, AI research tools are producing evidence that looks credible but dissolves under scrutiny. For an industry that sells trust and expertise, the cost of getting it wrong is measured in dollars, reputation, and regulatory attention. The KPMG case serves as a clear warning that the "check" in an AI-assisted workflow must be real, thorough, and human.

The scale of the fabrication: 89% of citations were flawed

GPTZero's scan of the report's 45 citations revealed a systematic failure of verification :

• Only 5 citations accurately pointed to real, relevant sources.
• 28 citations had paraphrased or altered titles, with fake components grafted onto real publications.
• 12 citations were too vague or inconsistent to trace back to any existing source.
• Approximately half of the report's substantive claims that relied on these citations appeared to be fake or misattributed.

Overall, 89% of the bibliography was flagged as flawed. GPTZero's AI detection scan added another troubling layer, rating the document 56% "mixed," which suggests it was likely AI-generated or heavily AI-assisted content that evaded meaningful human review .

Who was falsely cited, and what did the report claim?

The investigation identified four prominent organizations cited with fabricated case studies of "agentic AI" deployment :

• UBS: The report claimed the bank was deploying agentic AI for personalized wealth management insights, but no verifiable source existed.
• NHS Greater Manchester: It asserted the NHS trust was using AI-driven patient journey orchestration, a claim unsupported by real documentation.
• Transport for London (TfL): TfL was cited as an example of AI-powered real-time customer experience optimization, but the claim could not be corroborated.
• Swiss Federal Railways (SBB): The report presented a case study of SBB using agentic AI for integrated customer touchpoints, yet no such initiative could be verified in the cited material.

In each case, the citations pointed to non-existent reports, paraphrased titles of unrelated real publications, or references so vague they were functionally worthless .

The root cause: AI tools run amok without verification

GPTZero's analysis concluded that the errors were likely the result of an AI research tool over-complying with a prompt to find real-world examples of "agentic AI." The tool generated plausible-sounding but entirely fictitious case studies and citations, which then slipped past KPMG's review process uncorrected .

The incident underscores a critical vulnerability: when AI-generated research is published without rigorous human verification, it can produce confident, authoritative-sounding nonsense that damages institutional credibility .

A growing pattern of AI hallucinations in professional services

KPMG's pulled report is far from an isolated incident. It fits into an expanding pattern of AI-related citation failures across major consulting and legal firms :

• Deloitte Australia had to reissue a $440,000 government report after it was found to contain fictitious academic references and a fabricated quote from a federal court judgment . The firm issued a partial refund and faced public scrutiny over what one observer called a report "littered with citation errors" .
• Ernst & Young (EY) produced a report on loyalty fraud where GPTZero's investigation found that almost all URLs were broken or fake, and more than half of the titles did not correspond to real sources .
• Similar issues have surfaced in legal filings involving major firms like Sullivan & Cromwell .

These repeated failures highlight a systemic risk: without rigorous human oversight, AI research tools are producing evidence that looks credible but dissolves under scrutiny. For an industry that sells trust and expertise, the cost of getting it wrong is measured in dollars, reputation, and regulatory attention. The KPMG case serves as a clear warning that the "check" in an AI-assisted workflow must be real, thorough, and human.