Mr Altman’s prediction taps into an established school of thought. As large language models first gained popularity in the early 2020s, economists and bosses were hopeful that they, and other AI tools, would level the playing field, with lower-skilled workers benefiting most. Software capable of handling tasks such as protein-folding and poetry-writing would surely democratise opportunity. Jensen Huang, chief executive of Nvidia, a chip designer, envisioned a future in which workers “are all going to be CEOs of AI agents”.
More recent findings have cast doubt on this vision, however. They instead suggest a future in which high-flyers fly still higher—and the rest are left behind. In complex tasks such as research and management, new evidence indicates that high performers are best positioned to work with AI (see table). Evaluating the output of models requires expertise and good judgment. Rather than narrowing disparities, AI is likely to widen workforce divides, much like past technological revolutions.
View Full Image
The case for AI as an equaliser was supported by research showing that the tech enhances output most for less experienced workers. A study in 2023 by Erik Brynjolfsson of Stanford University and Danielle Li and Lindsey Raymond of the Massachusetts Institute of Technology (MIT) found that generative-AI tools boosted productivity by 34% for novice customer-support workers, helping them resolve queries faster and more effectively. Experienced workers, by contrast, saw little benefit, as the AI reinforced approaches they were already using. This suggested the tech could narrow gaps by transferring best practices from talented to less talented employees.
A similar trend was observed in other knowledge-intensive tasks. Research by Shakked Noy and Whitney Zhang, both of MIT, found that weaker writers experienced the greatest improvements in the quality of their work when using OpenAI’s ChatGPT to draft materials such as press releases and reports. Many saw better quality simply by using the AI’s unedited output, underscoring its ability to elevate baseline performance. Similarly, Jonathan Choi of the University of Southern California and co-authors found a general-purpose AI tool improved the quality of legal work, such as drafting contracts, most notably for the least talented law students.
The problem is that this is swamped by another effect. A job can be considered as a bundle of tasks, which tech may either commoditise or assist with. For air-traffic controllers, tech is an augmentation: it processes flight data while leaving decisions to humans, keeping wages high. By contrast, self-check-out systems simplify cashiers’ roles, automating tasks such as calculating change. This lowers the skill requirement, causing wages to stagnate.
View Full Image
Thus despite the early optimism, customer-service agents and other low-skilled workers may face a future akin to cashiers. Their repetitive tasks are susceptible to automation. Amit Zavery of ServiceNow, a business-software company, estimates that more than 85% of customer-service cases for some clients no longer require human involvement. As AI advances, this figure will probably rise, leaving fewer agents to handle only the most complex cases. Although AI may at first boost productivity, its long-term impact will be to commoditise skills and automate tasks.
Unlike earlier automation, which replaced routine jobs such as assembly-line work and book-keeping, AI may extend its reach to non-routine and creative work. It can learn tacitly, recognise patterns and make predictions without explicit instruction; perhaps, in time, it will be able to write entertaining scripts and design useful products. For the moment it seems as though, in high-wage industries, it is junior staff who are the most vulnerable to automation. At A&O Shearman, a law firm, AI tools now handle much of the routine work once done by associates or paralegals. The company’s software can analyse contracts, compare them with past deals and suggest revisions in under 30 seconds. Top performers have been best at using the tech to make strategic decisions, says David Wakeling, the firm’s head of AI.
The shift in recent economic research supports his observation. Although early studies suggested that lower performers could benefit simply by copying AI outputs, newer studies look at more complex tasks, such as scientific research, running a business and investing money. In these contexts, high performers benefit far more than their lower-performing peers. In some cases, less productive workers see no improvement, or even lose ground.
Intelligent design
Aidan Toner-Rodgers of MIT, for instance, found that using an AI tool to assist with materials discovery nearly doubled the productivity of top researchers, while having no measurable impact on the bottom third. The software allowed researchers to specify desired features, then generate candidate materials predicted to possess these properties. Elite scientists, armed with plenty of subject expertise, could identify promising suggestions and discard poor ones. Less effective researchers, by contrast, struggled to filter useful outputs from irrelevant ones (see chart 2).
View Full Image
Similar results have emerged in other areas. Nicholas Otis of the University of California, Berkeley, and co-authors found that stronger Kenyan entrepreneurs raised their profits by over 15% with an AI assistant, and strugglers saw profits fall. The difference lay in how they applied AI recommendations. Low achievers followed generic advice such as doing more advertising; high achievers used AI to find tailored solutions, such as securing new power sources during blackouts (see chart 3).
View Full Image
In financial decision-making, Alex Kim of the University of Chicago and co-authors conducted an experiment where participants used AI to analyse earnings-call transcripts before allocating $1,000 in a simulated portfolio. Sophisticated investors achieved nearly 10% higher returns with AI; less sophisticated investors saw gains of 2%. Seasoned investors made better use of insights from earnings calls such as those concerning R&D spending, share repurchases and operating profit before depreciation and amortisation.
As AI reshapes work, new tasks are emerging. Rajeev Rajan of Atlassian, an office-software firm, says that AI tools free up a couple of hours a week for engineers, allowing them to focus on creative work. Junior lawyers spend less time on chores and more with clients. “Really smart people whomay bebored with analysingroutineearnings releaseswillbenefit the most,” says aboss at a large investment firm. “The skill that is going to be rewarded most in the short run is imaginationin finding creative ways to use AI.” The grunt work of these industries is being automated, allowing junior employees to take on advanced tasks earlier in their careers.
Labour markets have always been defined by the destruction of old roles and the creation of new ones. David Autor of MIT has estimated that 60% or so of work in America in 2018 did not exist in 1940. The job of “airplane designer” was added to the census in the 1950s; “conference planner” arrived in the 1990s. But who will take AI’s new jobs when they emerge? History suggests that technological upheavals favour the skilled. In the Industrial Revolution, engineers who mastered new machinery saw their wages soar as routine labourers lost out. The computer age rewarded software engineers and rendered typists obsolete. AI appears poised to follow a similar path, benefiting those with the judgment, agility and expertise to navigate complex, information-rich environments.
Moreover, today’s AI tools are just the beginning. As the technology grows more sophisticated, semi-autonomous agents capable of acting independently—of the sort envisioned by Mr Huang—may transform workplaces. That might make every worker a CEO of sorts, just as the Nvidia chief executive has predicted. But there will be no levelling-out: the most talented will still make the best CEOs.
© 2025, The Economist Newspaper Ltd. All rights reserved. From The Economist, published under licence. The original content can be found on www.economist.com