The emperor’s new AI

“I am the Vice President of Output Accountability at Microsoft. I report to nobody with that title because the title did not exist until I created it. The title is necessary. The outputs are not accountable to anyone.”

So begins a typical bulletin from Peter Girnus, who this year has become the most compelling chronicler of the changing workplace. The personas he adopts in his impeccably-sourced Twitter/X vignettes are often an executive cynically advancing his career by orchestrating a carnage of redundancies.

What makes it unusual, is that it marries the brevity and accuracy of journalism to satire. He skilfully blends two forms: the bullet-point braggadocio of the self-aggrandising LinkedIn influencer, and the corporate evasions of the professional HR manager. Faced with dissenting employees, for example, Girnus’ executives create a “career trajectory alignment session” leading to a “separation event”. 

The account is unusual in seeking to address a significant social phenomenon that our mainstream media has been largely unwilling to tackle. Since ChatGPT was opened to the public in late 2022, generative AI has created an avalanche of capital investment and a stock market bubble: all the growth in the value of equities since 2022 has come from AI-related stocks. But the AI revolution, talked about for so long, appears to be built on promises that cannot be fulfilled. AI creates wrong answers or provides false information, and so does not work reliably enough to automate common tasks. We now have an abundance of evidence that when given an honest accounting, it does not produce productivity benefits. 

As Girnus recounts: “MIT: 95% of AI pilots fail. IBM: only 25% deliver expected ROI. NBER: 90% of CEOs admit no productivity impact.”

That impact is often negative. There is also a fascinating sociological division opening up. 88 per cent of executives express confidence in the tools they have forced their staff to use. But only 21 per cent of workers agreed. One survey estimates AI is adding 51 work days a year. Only this week we learned that AI chatbots misdiagnose over 80 per cent of early medical cases. 

“Most tools don’t work. The metrics are circular. The fear is the product”, Girnus’ fictional executive explains.

This is not a gentle satire, it’s a brutal roughing up of the executive class, fully justified because the subject matter is brutal. Firms have begun depleting their human capital in a frenzy of redundancies. Rather like in the Emperor’s New Clothes, the courtiers can’t publicly tell the truth. It’s a peculiar form of carnage, with human sacrifices being made to reassure the markets that the firm is at the bleeding edge of a technological revolution. This is almost pagan. 

Girnus (@gothburz) is not a professional writer or consultant, but a threat researcher at Trend Micro, a large cybersecurity firm. As he explains in a rare blog post: “AI tools are making workers measurably slower. The approved vocabulary for saying so does not exist.”

Girnus is not the only one to find the inability to reflect reality troubling. It’s something that I have chronicled regularly in my weekly Telegraph column. Stock market pressure obliged executives to conform, one board level executive at a FTSE100 company told me off the record in 2023. We have to keep pretending.

This week one senior AI professional, Han Lee, whose job is to make AI work for major corporations, shrewdly compared the phenomenon of the Emperor’s New AI to Mao’s Great Leap Forward. Immense damage was being done, but nobody could dissent. 

“Officials staged photographs. Everyone knew the numbers were fake. Everyone reported them anyway, because the alternative was being labeled a saboteur. The central government, delighted by the bounty, increased grain requisitions based on the reported yields. Farmers starved eating the difference between the real number and the fantasy,” Lee writes. The same dynamic can also be seen in today’s workplace, Lee writes, as teams compete to outdo each with extravagant claims of productivity increases from using AI.

Do they have a choice but to comply? As Girnus writes:

“The most stable form of institutional dishonesty requires no conspiracy. No enforcement. No policy. Only enthusiasm. And enthusiasm has no mechanism for dissent.” 

A similar phenomenon became institutionalised in the Soviet Union. Factory managers who were obliged to report good news upwards to the central economic planning unit also had to pretend that everything was rosy

In a Substack post, Girnus say he’s motivated by a failure of a public debate to highlight the phenomenon. 

“Something quiet is happening across white-collar workplaces, and it shows up most clearly in the numbers that nobody is publishing,” he writes. “Workers at companies that adopted AI tools over the past year are reporting, in anonymous surveys and in the growing genre of unsigned workplace confessions, that the tools have made them slower. Not marginally. Measurably.”

“A recent study found that workers using an average of fourteen AI subscriptions at a cost of $278 per month report a net productivity loss of four hours per week. That number has the quality of a confession. It is specific enough to be believed and round enough to suggest that the real figure is worse,” writes Girnus.

The irony is devastating. Unable to fix the hallucinations, Microsoft recently reclassified its CoPilot AI as an “entertainment” product. Girnus reaction as the executive:

“Entertainment is a legal position. We do not sell accuracy. We sell a tool that may produce accuracy. If it does, that is a feature. If it doesn’t, that is also a feature. The Terms of Service make no distinction.

Entertainment means: we built it, we sold it, we marketed it as essential, we disclaim it as unreliable, and the $30 means you agreed.”

What we are seeing here is a new form of journalism, and this is long overdue. Since the New Journalism of the 1960s, which gave us Norman Mailer, Hunter S Thompson and Tom Wolfe, innovations have been dry and technocratic. “Explainer journalism”, “data journalism” and infographics owe more to a corporate consultancy report than anything else. They don’t seem apt for an era of one to one electronic publishing. Finally, perhaps journalism is catching up with the medium.