Relevance: People telling everyone how to manage the AI transition are proving criticism correct.
> PwC published reports on AI and electric vehicles riddled with fake footnotes, misattributed claims and unverifiable information, the latest example of a Big Four firm’s slapdash use of AI-generated content.
The AI hallucinations were contained in “thought leadership” reports designed to drum up consulting work for partners in the Middle East, according to an investigation by researchers at GPTZero verified by the FT. The discovery of AI-generated errors risks embarrassing consulting firms such as PwC that are marketing their services as advisers to companies adopting the technology, including on how to use it responsibly and implement policies to avoid errors. GPTZero’s earlier investigations led Big Four rivals EY and KPMG to retract reports that also appeared to include hallucinations. The research group identified four reports PwC Middle East produced over the past two years whose text appeared to have been written with particularly heavy help from AI and whose footnotes revealed shaky foundations. The reports included a playbook for corporate use of “agentic” autonomous AI bots, a guide for governments on how to improve public services and market predictions for electric and autonomous vehicles across the region.
> Many of the citations in the reports are problematic in some manner, GPTZero researchers found. Often they link to web pages that do not contain the evidence they purport to and some to pages that do not exist. An academic paper on air quality in Riyadh appears to have been hallucinated entirely by AI, because there is no trace of the study in the journal referenced or by the authors to whom it is attributed. Several more footnotes refer to unusual sources. For example, a teenage blogger with 280 followers on Medium is cited as PwC’s source for information about a JPMorgan initiative it calls a “real world success story” of agentic AI. The initiative — in which the bank automated the review of commercial loan agreements, saving hundreds of thousands of hours of human work — was first reported in 2017, five years before ChatGPT launched the era of generative AI. PwC Middle East told the FT it “takes the accuracy of our published research seriously and is updating a limited number of supporting citations” in the identified reports. “Consistent with our approach to responsible AI, we have quality control processes for research and content development we expect all our people to adhere to,” it added, without addressing how the errors were included in the reports. In the past few years, the Big Four have pumped out hundreds of thought-leadership pieces on AI to help attract clients, while also exhorting staff to use the technology themselves to speed up and improve their work.
> The chaotic signposting of source material in the four PwC Middle East reports is symptomatic of AI-generated research, said Paul Esau, researcher at GPTZero. In one tell-tale sign, a claim that human error is responsible for 90 per cent of traffic accidents is mentioned in three different places in one report, Esau said, the first with a footnote, the second without, and the third with two footnotes — each footnote citing a different source. “This claim isn’t fake, but no human is going to cite the same fact three times in two pages using three different sources,” he said. One report fails to properly cite other PwC work. It references a PwC survey of Middle Eastern chief executives, of whom 70 per cent said generative AI will significantly affect their business, but the accompanying footnote links to a media report that makes no mention of the survey. Another footnote betrays the use of AI. The URL for a media report on cyber security threats to the energy sector includes “utm_source=chatgpt.com”.
1 Comment
Relevance: People telling everyone how to manage the AI transition are proving criticism correct.
> PwC published reports on AI and electric vehicles riddled with fake footnotes, misattributed claims and unverifiable information, the latest example of a Big Four firm’s slapdash use of AI-generated content.
The AI hallucinations were contained in “thought leadership” reports designed to drum up consulting work for partners in the Middle East, according to an investigation by researchers at GPTZero verified by the FT. The discovery of AI-generated errors risks embarrassing consulting firms such as PwC that are marketing their services as advisers to companies adopting the technology, including on how to use it responsibly and implement policies to avoid errors. GPTZero’s earlier investigations led Big Four rivals EY and KPMG to retract reports that also appeared to include hallucinations. The research group identified four reports PwC Middle East produced over the past two years whose text appeared to have been written with particularly heavy help from AI and whose footnotes revealed shaky foundations. The reports included a playbook for corporate use of “agentic” autonomous AI bots, a guide for governments on how to improve public services and market predictions for electric and autonomous vehicles across the region.
> Many of the citations in the reports are problematic in some manner, GPTZero researchers found. Often they link to web pages that do not contain the evidence they purport to and some to pages that do not exist. An academic paper on air quality in Riyadh appears to have been hallucinated entirely by AI, because there is no trace of the study in the journal referenced or by the authors to whom it is attributed. Several more footnotes refer to unusual sources. For example, a teenage blogger with 280 followers on Medium is cited as PwC’s source for information about a JPMorgan initiative it calls a “real world success story” of agentic AI. The initiative — in which the bank automated the review of commercial loan agreements, saving hundreds of thousands of hours of human work — was first reported in 2017, five years before ChatGPT launched the era of generative AI. PwC Middle East told the FT it “takes the accuracy of our published research seriously and is updating a limited number of supporting citations” in the identified reports. “Consistent with our approach to responsible AI, we have quality control processes for research and content development we expect all our people to adhere to,” it added, without addressing how the errors were included in the reports. In the past few years, the Big Four have pumped out hundreds of thought-leadership pieces on AI to help attract clients, while also exhorting staff to use the technology themselves to speed up and improve their work.
> The chaotic signposting of source material in the four PwC Middle East reports is symptomatic of AI-generated research, said Paul Esau, researcher at GPTZero. In one tell-tale sign, a claim that human error is responsible for 90 per cent of traffic accidents is mentioned in three different places in one report, Esau said, the first with a footnote, the second without, and the third with two footnotes — each footnote citing a different source. “This claim isn’t fake, but no human is going to cite the same fact three times in two pages using three different sources,” he said. One report fails to properly cite other PwC work. It references a PwC survey of Middle Eastern chief executives, of whom 70 per cent said generative AI will significantly affect their business, but the accompanying footnote links to a media report that makes no mention of the survey. Another footnote betrays the use of AI. The URL for a media report on cyber security threats to the energy sector includes “utm_source=chatgpt.com”.