KPMG 因幻覺疑慮撤回 AI 報告

🔑 增強重點摘要

•The retracted report, titled 'Total Experience: Redefining Excellence in the Age of Agentic AI' and published in October 2025, was found by AI detection software GPTZero to have only 5 accurate citations out of 45, with many others being fabricated or misattributed.
•GPTZero coined the term 'vibe citing' to describe how generative AI tools create fake references, mix real sources, or heavily paraphrase titles, leading to misinformation.
•The report contained specific false claims about major organizations, including UBS, Swiss Federal Railways (SBB), and Transport for London, exaggerating their adoption and capabilities of agentic AI, which these companies later confirmed as factually incorrect or misleading.
•This incident is part of a broader trend, as other 'Big Four' professional services firms like EY and Deloitte have also faced similar issues, with EY retracting a cybersecurity report in May 2026 due to fabricated footnotes and Deloitte refunding portions of a government contract for AI hallucinations.
•KPMG's own 2025 global study on trust in AI, conducted with the University of Melbourne, revealed that 56% of AI users reported making mistakes due to relying on unverified AI outputs, highlighting the known risks of hallucinations.

🛠️ 技術深入

AI hallucinations occur when large language models (LLMs) generate information that appears credible but is factually incorrect, nonsensical, or disconnected from reality.
These errors stem from the probabilistic nature of how AI models generate responses, predicting the most statistically probable next token rather than accessing verified facts or structured knowledge.
Causes of hallucinations include limitations in training data, where models 'fill in the blanks' incorrectly, and a lack of grounding, meaning models often lack direct access to external, real-time information to fact-check themselves.
The problem is often described as a 'context problem' rather than solely a 'model problem,' as LLMs can be highly capable when given the right information.
Retrieval-Augmented Generation (RAG) is identified as an effective method to prevent hallucinations by grounding model responses in actual, relevant documents and enterprise-specific information, rather than relying solely on the model's internal patterns.

🔮 前景展望基於引用來源的 AI 分析

There will be increased scrutiny and demand for robust human oversight in AI-generated content within professional services.

The reputational damage and need for retraction experienced by KPMG and other firms will compel the industry to implement stricter verification processes and human review for AI-assisted reports.

Enterprises will accelerate investment in AI governance frameworks and technologies like Retrieval-Augmented Generation (RAG) to mitigate hallucination risks.

The incident highlights the significant enterprise risks associated with unverified AI outputs, driving organizations to adopt solutions that ground AI in factual, company-specific data and establish clear accountability.

The incident will temper the hype around immediate, unverified AI adoption, leading to a more cautious and strategic approach to integrating LLMs for factual reporting.

The public exposure of 'vibe citing' and fabricated claims will serve as a reality check, emphasizing that while AI offers efficiency gains, its outputs require rigorous validation before being trusted for critical business decisions.

⏳ 時間線

2023-11

KPMG announces a $2 billion investment in generative AI over three years, partnering with Microsoft to embed GenAI across its business.

2025-10

KPMG publishes the 'Total Experience: Redefining Excellence in the Age of Agentic AI' report.

2026-03

KPMG fields a survey for its 'AI in Finance 2026' report, which would be published in May 2026.

2026-05

KPMG publishes 'AI in Finance 2026' report, building on earlier research and its Q1 2026 Global AI Pulse.

2026-06

GPTZero's investigation reveals that KPMG's October 2025 report contained widespread AI hallucinations, including fake citations and fabricated case studies.

2026-06

KPMG officially retracts the 'Total Experience: Redefining Excellence in the Age of Agentic AI' report due to hallucination concerns and begins reviewing the circumstances of its publication.