Is AI-led job displacement about to reshape the economy?
Not so long ago we were worried less about chatbots and more so about robots. In 2013 the Oxford researchers Carl Benedikt Frey and Michael Osborne released a paper claiming that 47 per cent of jobs would disappear to automation within one to two decades. The piece triggered a deluge of hype about automation.
Journalists scoured the economy, looking for examples of disappearing jobs. We heard stories of robots revolutionising package delivery, burger flipping, strawberry picking, T-shirt production and even journalism itself. The vast majority of the firms making these claims failed within months.
Ten years on we have seen little of the automation Frey and Osborne feared. Instead of mass unemployment, we have tight labour markets. Our economies continue to experience slow rates of productivity growth, which is the major cause of what economists call secular stagnation.
Will ChatGPT succeed where the previous wave of robot-based automation failed? Researchers from OpenAI, its creator, recently published a working paper claiming that ChatGPT is about to automate away vast numbers of jobs. The publication coincided with the release of OpenAI’s GPT-4.
Once again, there are reasons to doubt the hype. The paper proclaiming a new age of ChatGPT-inspired automation relies on the same faulty prediction techniques that Frey and Osborne used. Researchers asked human experts (as well as ChatGPT itself) to evaluate tasks in terms of their susceptibility to automation. The researchers then used O-NET, an online database of job requirements, to predict which jobs were likely to be automated.
Among the jobs most likely to be done by large language models, according to the OpenAI researchers? Central bank monetary authorities and clergymen – two jobs that definitely require a human touch – as well as scientists, engineers and sociologists.
A similar methodology led Frey and Osborne’s 2013 study to wildly overestimate susceptibility to automation. Their technique, which they adapted from efforts to predict the likelihood that jobs will be offshored – that is, taken over by other human beings, working overseas for lower wages – has never successfully been applied to digital technologies.
In part, that is because computer experts turn out to be bad at predicting computers’ capacities for autonomous operation. At the same time, the range of tasks involved in most jobs turns out to vary far more than O-NET suggests. Jobs such as school teacher or lathe operator look different across workplaces in the United States, and vary even more so across Germany, India and China. Legal frameworks, collective bargaining agreements, wage-levels, comparative advantages and business strategies all shape how jobs evolve, in terms of technologies used and tasks required.
[See also: How to prevent AI from taking over the world]
Volkswagen spent the 2010s investing in robots to improve productivity on car assembly lines. By contrast, Toyota, the most efficient car company in the world, removed robots from assembly lines over the same period, to improve the responsiveness of its factories to changes in sales trends.
Researchers at the OECD reran Frey and Osborne’s numbers in 2016 based on a more realistic account of the variety of forms that jobs take and found that less than 10 per cent of jobs in the US were likely to be automated, and even that figure has so far turned out to be an overestimate. The same is probably true of the figures cited by the OpenAI researchers’ working paper. They claim that, with complementary technologies, large language models threaten 49 per cent of jobs. We can expect to see that figure fall to 10 per cent of jobs or less with further research.
That rate of technical change is well within what a healthy economy can handle, which is not to say that our economy is healthy. For reference, something like 60 per cent of the job categories people worked in in the late 2010s had not yet been invented in 1940. Still, even if the vast majority of jobs are unlikely to disappear, and if many new jobs are likely to be created, the nature of work will change due to the implementation of technologies like ChatGPT. We need to shift our thinking about how that change occurs.
Over the course of the 20th century, few jobs were fully automated out of existence. Gone are the lift operators, film projectionists and travel agents of yesteryear, but it is difficult to cite more examples. Most jobs did not disappear with technological progress. Instead, their content changed. Longshoremen used to load cargo by hand; now they operate cranes.
As technologies diverge across countries, so too do job requirements. The same job on a film production crew might require different competencies in Hollywood than in Bollywood or Nollywood. Workers in Sweden report higher degrees of autonomy in deciding how they carry out their jobs than workers in the US or UK.
That jobs change in different ways, across countries, suggests that there is nothing inevitable in this process. Without holding back technologies, we can dig out new channels for them to flow into, as a way to ensure that new technologies improve rather than harm society.
How might jobs change with the adoption of ChatGPT and related applications? Even veritable revolutions, like those inaugurated by the advent of the steam engine in the 19th century and the internet in the 20th, unfolded gradually. So do not believe every company press release proclaiming a revolutionary advance. Do not rely on over-exuberant self-reported tests. Wait for actual, on-the-ground results.
ChatGPT will probably have its greatest impact on computer programming, technical writing and legal writing, fields which have seen woefully low rates of productivity growth. No one knows whether, on balance, software such as Copilot, which assists developers with writing code, will end up increasing or decreasing the demand for their services. Falling programming costs may unleash pent-up demand for programmers’ services.
The problem with using large language models like ChatGPT across the wider economy is their tendency to “hallucinate”. When they do not know the answer to a question, they lie. As the AI expert Gary Marcus has explained, these hallucinations are not a minor bug. They are a fundamental limit of the technology, which will persist no matter how large ChatGPT’s training data gets.
What the chatbots might eventually enable is a personal digital assistant much more capable than Apple’s Siri. A digital companion that could read emails, schedule meetings, draft communications, explain things we don’t understand, and provide helpful reminders would make a big difference in people’s lives.
Such tools would help to solve a problem that the internet and other digital technologies have themselves created: we are inundated with information. The advent of social media and smartphone notifications have made daily distractions much worse, and have probably had major, negative effects on our productivity, as well as our mental health.
Only time will tell whether ChatGPT will finally solve the problem of information overload – or will make it worse, by increasing the speed with which information and disinformation proliferate. In either case, the future of these technologies should not be left to Silicon Valley to divine.
[See also: Who’s really making money from HustleGPT?]
This article appears in the 19 Apr 2023 issue of the New Statesman, Axis of Autocrats