The release of generative AI models like ChatGPT has caused huge excitement and generated predictions of mass job losses and large gains in productivity.
Goldman Sachs estimates that two-thirds of jobs in the US are exposed to some degree of automation due to AI and 25% of those exposed occupations could have as much as half of their workload replaced. Economists at the University of Pennsylvania have come up with similar estimates. Goldman argues that AI has the potential to raise growth in global productivity by 1.5 percentage points.
It will take years for such effects to become apparent at a macro level. For now trying to assess the effects of AI relies more on individual studies, cases and stories, most of which highlight the potential rather than the limitations of the technology.
One of the more striking studies highlighting the potential of AI was published earlier this month by the US National Bureau of Economic Research. It involved a control trial with 5,000 customer service employees in a major US software company. Some employees were provided with an AI tool that offered potential replies to customer queries which the agent could ignore, adapt or use. The AI tool increased productivity by 13.8%. Less experienced workers achieved a 35% rise in productivity while the most skilled and experienced workers saw few gains from using the tool. In what read like a glowing endorsement of this new technology the study found that AI-assisted interactions helped improve customer satisfaction and reduced employee attrition by 8.6%.
In a similar experiment the chief executive of Octopus Energy in the UK said last month that AI was doing the work of 250 customer service workers and writing emails that delivered 80% customer satisfaction, well above the 65% achieved by skilled, trained people.
Other stories also speak to the change brought by AI. The CEO of IBM Arvind Krishna recently said that the firm would pause hiring for roles that could be replaced by AI in the coming years. Last month the share price of Chegg, an online education provider, fell by half after it said that students were increasingly turning to ChatGPT for tuition. By contrast shares in chipmaker Nvidia rose 30% after it predicted a surge in demand to build generative AI models. A recent survey of 12,000 workers by Fishbowl, a professional network app, found 43% had used tools like ChatGPT—a large majority without their bosses knowing.
These studies and stories underscore the potential of AI, but there are significant barriers to realising this potential across the economy. New technologies endlessly reshape individual sectors. At a personal level think of the change brought by contactless payments, online shopping or self-checkouts in the last few years. But these sort of sector specific changes are quite different from those wrought across the whole economy by a general-purpose technology, such as electrification or the personal computer. AI demonstrates capacity in specific tasks, such as coding or customer relations. The question is whether it has the potential to be deployed much more widely across the economy in a far wider range of roles.
History shows that it often takes a long time for work to be redesigned to harness the full benefit of new technologies. It took decades for personal computers to impact measured US productivity. No wonder that, in 1987, the US economist and Nobel laureate, Robert Solow, famously lamented that "You can see the computer age everywhere but in the productivity statistics”. Productivity did eventually respond, in the 1990s, but getting there took time, involved major disruption, the deployment of large amounts of capital and a lot deal of trial and error.
Deloitte’s latest “State of AI in the Enterprise” report notes that while 94% of business leaders see AI as being important to their organisation’s success, only 27% think their organisation have policies and processes needed to fully harness AI.
Businesses also need to be alive to the limitations of and risks from AI. Current chatbots, for instance, have a tendency to ‘hallucinate’, producing plausible but incorrect answers. Concerns about AI generating discriminatory outcomes or ownership and copyright questions loom large. Regulation is trying to catch up but, inevitably, is behind the curve. The EU AI Act, for instance, is unlikely to come into force until later this year or early next year.
A recent book by US-based economists Daron Acemoglu and Simon Johnson, “Power and Progress”, points out the importance of institutions, structures and social norms in determining how technological change affects society. (They argue, for instance, that it was not until the emergence of a more communitarian view of society in the late 19th century that the benefits of industrialisation started to be felt in the wages and conditions of most working people.) Institutional factors may explain the tendency of the public sector, which accounts for around 50% of EU GDP, to be slow to embrace new technologies. This could be due to an absence of competitive pressure, conflicting policy goals, such as protecting jobs, and high levels of unionisation in the public sector. A recent Economist article observed that train drivers on the London underground are paid close to twice the national median wage, even though the technology to partially or wholly replace them has existed for decades. Meanwhile in San Francisco, the global centre of AI, police are still employed to direct traffic during rush hour.
It could be different this time. The spread of ChatGPT suggests a potential for rapid take up. But the true driver of change seems more likely to be AI tools made for particular uses, such as report writing, customer enquiries or analysing data. AI is already incorporated into functions such as predictive texting and there seems to be significant scope to be used in existing systems. Microsoft is currently testing its AI assistant, 365 Copilot, which is incorporated into programs such as Word, Excel and Outlook.
But what of the potential for AI to render large swathes of the workforce redundant? In his book Professor Acemoglu argues that it is possible that AI will replace human jobs without creating new, more productive work for humans to move into.
To us it seems more likely that the pattern of the past, where those displaced by technology are redeployed in newly created roles, holds. A recent report by David Autor of the Massachusetts Institute of Technology estimates that 60% of US workers are employed in occupations that did not exist in 1940, implying that over 85% of employment growth over the last 80 years is explained by the technology-driven job creation. As Professor Acemoglu and his co-author Professor Pascal Restrepo note in a paper published in 2018, “throughout history, we have not just witnessed pervasive automation, but a continuous process of new tasks creating new employment opportunities for labour”.
Previous predictions of mass unemployment driven by technology have tended to be wide of the mark. In 2013 two Oxford academics estimated that automation could wipe out almost half of all US jobs over the subsequent “decade or two”. Here we are, in 2023, with US employment at 156m, up by 16% or 31m jobs on 2013.
A detailed analysis of the US labour market published by the US Bureau of Labor Statistics (BLS) last year found little support for the idea of automation, robotics or AI causing a general acceleration in job losses. Even in sectors thought by economists and technologists to be at most risk of automation the authors found few signs of rapid job losses. As the paper notes, “Part of the reason new technology does not produce sharper changes in employment is the diversity of jobs within occupations and the diversity of tasks within jobs, not all of which are equally susceptible to technological substitution. Automation of some tasks may also alter the task composition of jobs, rather than simply reducing the number of jobs”.
Jobs in radiology are seen by technologists as being particularly vulnerable to AI. Professor Geoffrey Hinton, one of the leading figures in the development of AI said in 2016, “We should stop training radiologists now. It’s just completely obvious that within five years, deep learning is going to do better than radiologists.” Yet the number of radiologists has risen, not gone down. Rising demand for healthcare is one factor; the work of a radiologist also involves a wide variety of complex diagnostic and related tasks that cannot be performed by AI tools. (Since autopilot became standard on large passenger aircraft more than 40 years ago, the number of pilots has continued to rise while the safety of air travel has improved dramatically.)
The BLS also finds that an array of new technologies, from robot lawn mowers and vacuum cleaners to robot surgery and translation software, have failed to put a measurable dent in employment in the affected sectors in the US.
On balance AI seems likely to do what technology has always done – to reshape work, replace some jobs, create many more and raise productivity. Businesses are convinced of the opportunities. In our latest CFO survey, respondents told us they expect to see a wave of AI-related capex driving UK productivity. We share the optimism, but think it will, to paraphrase Robert Solow, take time to see the effects of AI in the productivity numbers.
PS: Last week’s briefing discussed the evolution of hybrid working and the difficulties employers can face in ensuring staff spend time in the office. CNBC reported last Wednesday that Google will start tracking office attendance using entrance passes and will include workplace presence in performance reviews. Most Google employees are expected to be in the office at least three days a week. According to The Wall Street Journal, Lyft, Meta and Salesforce have also tightened their policies on hybrid working.
For the latest charts and data on health and economics, visit our Economics Monitor: