PwC joins three other Big Four accounting firms in publishing reports containing AI-generated content with fabricated sources and false claims, according to analysis by GPTZero, an AI detection platform.
GPTZero identified four PwC Middle East reports containing fabricated citations and unverified information. One governance report scored 84 percent AI-generated while promoting a PwC product with customer references that could not be verified. The findings follow similar discoveries at KPMG, Deloitte, and Ernst & Young, leaving all four major accounting firms exposed to the risks of AI hallucination.
The reports appear to have used large language models without adequate fact-checking or editorial oversight. AI language models regularly generate plausible-sounding but false citations, a problem known as hallucination. When these models invent sources or statistics, they do so with confidence that can fool readers unfamiliar with the technology's limitations.
The pattern across all Big Four firms reveals a systemic issue. These organizations deploy AI to produce client reports and research at scale, but their quality assurance processes have failed to catch fabricated sources before publication. For consulting firms, credibility depends entirely on accurate information. Publishing unverified claims damages client trust and exposes the firms to potential liability if clients rely on false data.
The issue extends beyond embarrassment. If a client makes business decisions based on a PwC report containing fabricated sources, the firm could face legal consequences. Professional advisory work carries implicit guarantees of accuracy. Consulting firms cannot delegate fact-checking to AI without maintaining rigorous human review.
PwC has not publicly responded to these findings. The pattern suggests these firms prioritized speed and cost reduction over accuracy when integrating generative AI into their operations. As more organizations deploy AI to produce client-facing content, the tension between efficiency and reliability becomes unavoidable. Without stricter verification protocols, similar hall