Article contents:
- Introduction: a crisis of trust in consulting reports
- Details of the Canadian report and discovered errors
- AI “hallucinations” and fabricated references
- Reactions and expert concerns
- Deloitte’s defense and recurring incidents
- Implications of using artificial intelligence in official research
- Frequently asked questions
Introduction: a crisis of trust in consulting reports
In an incident highlighting the risks of unverified reliance on generative artificial intelligence, the Canadian branch of global consulting firm Deloitte is facing serious allegations. It is alleged that a healthcare consulting report, which cost a Canadian province nearly $1.6 million, contains errors potentially resulting from artificial intelligence, including citations of fake research.
This incident makes Canada the second country this year to allege fact-checking shortcomings at Deloitte, raising serious questions about the reliability of reports governments rely on to make critical decisions.
Details of the Canadian report and discovered errors
The discovered errors came to light in an investigation published by The Independent, a Canadian news outlet covering Newfoundland and Labrador. The report in question, spanning 526 pages and published by the provincial government in May, was intended to advise the Department of Health and Community Services on vital topics such as virtual care, staff retention incentives, and the impact of the COVID-19 pandemic on healthcare workers, at a time when the province’s health sector faces shortages of nurses and doctors.
According to disclosed information, the government spent just under $1.6 million on this report, paid in eight installments. To date, the report remains available on the Canadian government’s website.
AI “hallucinations” and fabricated references
The investigation found that Deloitte’s report contained false citations drawn from entirely fabricated academic papers to derive conclusions for cost-effectiveness analyses. The report also cited real researchers in papers they never worked on, and included bogus papers co-authored by researchers who confirmed they had never worked together.
For instance, the lengthy report cited an academic paper from the “Canadian Journal of Respiratory Therapy” that could not be found when searching the journal’s database. This phenomenon, known in the AI field as “hallucination,” occurs when generative models invent information that sounds convincing but is entirely incorrect.
Reactions and expert concerns
These discoveries have sparked concern among academic experts. Jill Tomblin-Murphy, an assistant professor at the Dalhousie University School of Nursing in Nova Scotia who was cited in the report for a “non-existent” academic paper, commented: “It appears that if you are coming up with things like this, they might be using AI heavily to generate the work.”
Tomblin-Murphy added that she had only worked with three of the other six authors listed in the erroneous citation, emphasizing the associated risks: “We have to be very careful to ensure that the evidence underpinning reports is the best evidence, and that it is verified evidence. Ultimately, these reports must be accurate, evidence-informed, and useful moving forward, not just because they cost governments and the public money.”
Deloitte’s defense and recurring incidents
Responding to these accusations, Deloitte Canada defended its report. A company spokesperson stated: “Deloitte Canada strongly stands behind the recommendations in our report. We are reviewing the report to make a small number of citation corrections, which do not impact the findings of the report.” The spokesperson added that “artificial intelligence was not used to write the report; rather, it was used selectively to support a small number of research citations.”
However, this incident is not Deloitte’s first of its kind this year. This revelation follows news last month that Deloitte used AI in a $290,000 report published in July to assist the Australian government. A researcher pointed out hallucinations in the 237-page study, which included references to non-existent academic research papers and a fabricated quote from a federal court ruling.
In the revised study quietly uploaded to the Australian government’s website, the firm admitted it used the Azure OpenAI generative AI system to assist in preparing the report. Deloitte in Australia was forced to issue a partial refund to the government for the report.
Implications of using artificial intelligence in official research
These incidents underscore the significant challenges facing the integration of artificial intelligence into professional research and consulting operations. While AI can accelerate information gathering and analysis, it still lacks the ability to reliably distinguish between fact and fiction. Relying on these tools without rigorous human oversight and thorough fact-checking can lead to disastrous results, especially when these technologies are used as a basis for public policy and government decisions affecting millions of lives.
Frequently asked questions
1. What are AI “hallucinations”?
It is a phenomenon where a generative AI model invents information that appears accurate and convincing, but is actually entirely false or baseless.
2. Did Deloitte admit to using AI in the Canadian report?
The firm acknowledged using it selectively to support certain research citations, but denied using it to write the entire report.
3. What consequences has Deloitte faced so far?
In a similar incident in Australia, the firm was forced to issue a partial refund to the government. No information has been announced yet regarding a potential refund for the Canadian report.
