Artificial stupidity  Thousands of companies are using AI for the wrong things

Silicon Valley’s most pervasive product is hype – new versions of which are produced on a massive scale daily. Prediction and propaganda have become intimate bedfellows. This prompts an equal and opposite reaction from doom-mongers, grabbing attention (and consulting assignments), with the result that opinion polarises, and the curious and open-minded exit the discussion. What follows is an attempt at a disinterested, pragmatic analysis of where we are and where we may – or may not – be going.

Dearer than your mother’s WorldWideWeb

Most companies started using AI when ChatGPT broke through in November 2022. Yet few firms have an AI strategy, even as their younger workers toy with it daily.

There is little clarity about how AI works. Thus, managers frequently treat it as they would a garden-variety algorithm. Initial usage has largely focused on chatbots providing so-called customer service. This is a classic implementation of new tech as a cost-cutter. Yet the evidence is that chatbots make customer service worse. Most customers hate them. The risk of brand damage is high. Will they get better? Yes – and more expensive.

And there is the first conundrum. Because AI development requires vast amounts of capital expenditure, it may not (as was the case with the internet) get cheaper quickly. It’s possible that old-fashioned customer-service humans turn out to be both friendlier and a better investment – using customers as guinea pigs for new technology is risky. For a meaningful return on the cost of AI, using it to replace some higher-order work makes more sense than using it to replace front-of-house service teams. But this is a logic to which the people doing those more expensive jobs will be wilfully blind – because it is they who make decisions about where to experiment with AI first.

What makes AI so expensive? While it sounds ethereal, it depends on big buildings full of hardware whose constant consumption of energy is costly in money and carbon; one dollar’s worth of computing time produces roughly the same emissions as driving a car 20 miles. The result is that datacentres could account for up to 21% of global energy consumption by 2030 .

AI runs on chips that cost around $50,000 each. These are quickly rendered obsolete by more expensive upgrades, such that their depreciation affords little comfort to company accountants. Applying AI to activities that are either already very expensive, or where the return on investment is meaningful and/or quick may make more sense. Whether it will yield a decent ROI remains unproven. In this respect, economists Paul Krugman and short-seller Jim Chanos caution that drawing analogies between the dawn of the internet and that of AI are lazy.

So what’s so special?

Internet history can remind us that technology renders generic what once made firms unique. Your company won’t win loyalty for customer service when everyone uses the same annoying chatbot. If you need content to market your business, don’t fire your writers just yet. AI-generated articles will feel strangely like those of your competitors, unless you spend hours tweaking your prompts, in which case – you might as well use a writer.

AIs are trained on vast amounts of publicly available data. Thus, they can produce detailed reports, but absent of any distinctive conclusions or style. With no inevitable correlation between frequency and importance in public information, a gravitational pull to the banal may leave businesses struggling for insight. When companies can so easily copy competitors using similar tech, it is no surprise that they fail to stand out. I remember the then-head of Disney+ excitedly offering me a pre-launch preview of its homepage. The dramatic reveal? It looked just
like Netflix.

Genuine value lies where companies apply creative, highly specific questions to rich, proprietary datasets. Apple, in its search to maintain the elegant durability of its computers while honouring its sustainability strategy, identified a new alloy chemistry for making a new, endlessly recyclable aluminium that halved the laptop’s carbon footprint. That bold, competitive breakthrough prompted further innovation and partnerships that led to carbon-free smelting: the flywheel of invention every business aspires to.

AI development is already accelerating drug development. Such work historically required an average investment of $2.6 billion, over a period of 12–15 years, to glean a less than 10% chance of success. With more than 1060 molecules to work with, identifying the most promising shortens trial-and-error research times, predicts drug/protein interactions, estimates potential abreactions, and seeks additional therapeutic uses. The hard work
lies in managing the data with sufficient precision; integrating legacy data demands time and investment.

Winners and losers

The ownership of the data – and its prodigious offspring – will be contentious. When you combine datasets from myriad sources or trials, who is entitled to the rewards the resulting products deliver? How will companies or research institutions even know when their data has been used in ways unforeseen just a few years ago? If your biomedical data has been used to deliver a wonder drug, shouldn’t you partake in its upside?

This is not a new question. In the 1950s, cancer cells of an African American woman, Henrietta Lacks, subsequently became one of the most productive cell lines in medical research. But it was decades before her family learned that her cells had enabled numerous medical innovations – including polio and even Covid-19 vaccines. Lacks’ contribution was one which, prior to legal action, neither she nor her family had been notified
or compensated.

Similar concerns now afflict all owners of intellectual property. Not only do they fear that their online material might be training AI without any recognition or reward, but the capacity of any creator to earn income from their work is threatened by a technology that can generate limitless derivative works. Having found her TV script had been scraped, scriptwriter and MP Alison Hume now worries that AI will be used to generate further scripts for which she will receive neither credit nor remuneration but which, one day, could replace her. Any producer of IP will have similar concerns.

Calling creative minds

Predictions of how many jobs AI will destroy vary enormously and are notoriously unreliable; following these numbers is fruitless. Yet some job losses have been instantaneous. In music, writing jingles for commercials – the classic entry-level work for musicians – was automated almost overnight, while the number of jobs creating images fell by 35% between January 2022 and July 2023. Cautious companies replacing lower-level jobs with AI prompts a serious question: with the first rung of the career ladder sawn off, how will anyone’s career get started?

Instead of tracking job losses, companies would do better to consider what skills are required for navigating endemic uncertainty. Where companies once prized degrees, grades and credentials, they’re finding such specialisation too rigid for rapid change. Instead, the World Economic Forum’s The Future of Jobs Report 2025 prioritised analytical and creative thinking. By 2035, the capacity to communicate, collaborate, think creatively and navigate information will be essential. Mindsets as likely to be found in liberal-arts graduates as Stem specialists. This posits a radically different profile for hiring, as for any HR processes.

Use it or lose it

Asked to draft a speaker introduction recently, an unconfident colleague turned to AI. It duly produced a functional text for him – but it conveyed nothing of his own personality. Where he could have mastered a new skill, instead he
chose not to learn.

The skills we outsource to tech we lose, or don’t develop, in ourselves. I’ve outsourced remembering my children’s phone numbers to my phone. And I don’t mind because, should I lose my phone, it’s backed up. But writing is thinking. Do I want to outsource that? Shaping my own ideas, discovering contradictions and connections, developing my own voice and questions? Are these skills we are willing to lose?

The same applies to decision-making. The gold standard of a good decision is one which can be explained well enough that even those who don’t agree can understand it. Experience, insight and instinct are required to do this well. Outsource choices to technology and nobody learns. Leaders struggle to explain choices they didn’t make. Why should anyone trust or implement a decision no one contributed to? Where does accountability lie? And what happens when different AIs reach different conclusions; who decides then?

Existential threats

The greatest fear underlying the uncertainties of AI is that it will outsmart us – or go rogue. Stuart Russell, one of the sanest AI experts, describes the Midas effect, in which AI’s effective solution to a problem also proves deadly. Variants of this scenario abound.

All technology comes with side-effects and dangers, and AI is no exception. Consequently, Russell convenes legal scholars, ethicists, psychologists and philosophers to explore the ramifications together. The capacity to assemble the right people around the hard questions will become a key role for leadership. One model is to think of AI as a new drug, tested for potential perverse outcomes before being unleashed on the public. Do No Harm might be the appropriate oath: a mindset which characterises much emerging EU legislation.

Will it be worth it?

The impatient would rather forego safety in the search for economic growth. But there are at least two reasons to be cautious here. Unlike the birth of the internet, AI risks are exacerbated by dominant players who are already quasi – or actual – monopolies exerting largely unchecked market and political power. This endangers society, politics and the economy. Absent challengers or restraint, current inequalities will grow even greater, with all the social and political volatility that follows in their wake. That’s bad for business.

Will it all have been worth the job destruction, risks, threats to privacy, identity and society? The 2024 economics Nobel Prize winner Daron Acemoglu pegs productivity growth at “no more than a 0.66% increase in total factor productivity over ten years,” adding that 0.53% might be more realistic as the sheer difficulty of developing AI makes ROI slower. He sees no evidence that AI will reduce economic inequality, especially as bad actors may deliberately incur costly harms. It’s worth recognising that what upside might emerge is not achievable within an electoral cycle.

The propaganda of inevitability spewed out by Silicon Valley suggests we have no choice. But we have many: about how, when, where and why to use AI. About who gets hurt. We need leaders who are knowledgeable enough to convene the right people around the right questions; and sufficiently trusted and experienced to take independent decisions their people understand. These leaders must not take their steer from vested interests – but from the vast reservoir of knowledge and imagination that is human.