PwC screwed up with AI; it’s happening with too many big firms

PwC screwed up with AI, giving the green light to projects littered with AI mistakes that somehow made it past all standard checks. The company is the latest in a long list of big names that have launched themselves into this entirely avoidable nightmare. And it’s all down to a lack of proper AI governance.

This time, PwC has been found to have issued multiple public-facing documents containing extensive AI hallucinations, all of which would have been caught – or even just not made in the first place – if there was a more thorough and enforced rulebook governing AI’s use across its network. 

It underscores the serious levels of risk that companies still seem content to make when it comes to AI. Corporate communications are the essential channel that firms need to do business, but if those communications are built on repeated errors, the level of trust and confidence will begin to tumble fast. 

PwC – one of the world’s largest professional services networks – was found to have published four reports containing AI hallucinations and unchecked material. The flaws were spotted by AI detection software developer GPTZero as part of a wider investigation into irresponsible AI use by big global firms. 

Crucially, one of the four pieces – published by the company’s Middle Eastern branch – contained references to a PwC framework called “Citizen Pulse”. The report claimed that “Citizen Pulse” was already being used by multiple governments, including Denmark, the US, Saudi Arabia and Australia. However, the framework has next to no public reference or acknowledgement outside the report, GPTZero said. 

“Together, these reports show a pattern of irresponsible AI usage resulting in hallucinated (vibe) citations, fabricated claims, and incomprehensible drafting and formatting decisions,” the company summarised.

It is, by no means, the first instance of high-profile companies making blunders with AI. While some of these blunders may meet the threshold for at least some understanding, the ones below echo the PwC in their blatant lack of governance oversight. 

• The time Air Canada deployed a customer service chatbot on its website to assist passengers, but responded to their queries with promises based on a poor (or simply absent) interpretation of the airline’s policies. 
• The time Deloitte Australia used unverified AI content in a report for the Australian government, valued at AU$440,000. The report had references to academic papers that didn’t exist, and a quote from a judge that was never made. 
• The time KPMG published a thought leadership report around deploying AI, containing several case studies that were complete hallucinations (this was also uncovered by a GPTZero investigation). Globally known names like Swiss Bank UBS, the UK’s NHS, and Transport for London, were mentioned.

It’s a huge deal. It might not seem like that to some corporate leaders, who will argue that “if everyone’s making these mistakes, we won’t stand out,” or “I don’t see our bottom line being impacted by AI errors”. 

Both of those viewpoints are seriously time-limited. If they work once, they’re not likely to work again. 

Ultimately, corporate leaders have to answer to their company’s own stakeholders, who will lose patience quickly if they see important documents issued with AI-generated drivel. 

This problem gets far worse if it emerges that multiple documents were bulk-produced by AI. Having one AI mistake in a human-written paper is bad enough; imagine the flatlining confidence among stakeholders if they find out that humans had next to no involvement beyond final design touches. 

Sadly, that is the state of play for many companies worldwide. 

The real risk is in the culture it creates: one of feeling at liberty to use unchecked AI for anything the company produces. That culture can easily creep into financial documents, contributing to serious errors that could ultimately mislead boards, shareholders, journalists and consumers. Eventually, the errors will come back to haunt you.

Beyond the obvious answer of checking everything that AI helps with, the bigger challenge is to change the “Laissez-faire” culture around AI, or stop it from developing in the first place. 

This has to come from boards and C-suite leaders. Their commitment to proper AI oversight must be visible from the get-go. There should be rulebooks; those rules should be enforced, and the appetite for proofing should be consistently high. 

Anything less than this kind of culture will encourage mistakes, and the real issue is, there’s no limit to how serious those mistakes can be. Even one wrong figure can change an entire revenue projection. One wrong case study can bring up conversations around defamation, snowballing faster than leaders can contain it. 

Ultimately, it’s all completely avoidable risk. That’s why it’s especially damming if a leader has to admit that a major blunder was the result of AI-generated content that simply wasn’t fact-checked.