The ‘vast uncertainty’ of AI and jobs

I wrote here, critically, about the recent IMF report on the impact of AI and jobs. I was doing some research for a report I was writing recently which required me to go back into the same area, and found an article by David Autor that has a different view. Not more cheerful, necessarily, but maybe more honest about the radical uncertainty that pervades this whole area.

Autor is a credible voice in this area, since he has been writing about the relationship between technology and labour markets for several decades now, so has the benefit of context. But it is also a reminder that thinking about models is more useful in addressing uncertainty than just pumping out modelling data.

Skill premium

The title of the article, on the NBER website, conveys the tone: ‘The Labor Market Impacts of Technological Change: From Unbridled Enthusiasm to Qualified Optimism to Vast Uncertainty’. It’s a long article, though never technical, and so I’ll only have the space here to pick up some of the main points.

This is his starting point:

what workers earn in a market economy depends substantially, though not exclusively, on their productivity–that is, the value they produce through their labor. Their productivity depends in turn on two things: first, their capabilities (concretely, the tasks they can accomplish); and second, their scarcity… In conventional terms, the skill premium depends upon the supply of skills and the demand for skills.

So, if that is the structure of the market, what is the impact of technology? The answer is not obvious, and different models of the relationship between technology and work generate very answers. Autor reviews four models, briefly summarised here by me.

1. The education race

This is the most influential model, originally proposed by Jan Tinbergen and popularised by Claudia Goldin and Lawrence Katz. Tinbergen started by observing that the wages of Dutch workers with post-high school education were rising, even though their numbers were increasing. He concluded the modern economies had a race between increasing demand for skills and being able to supply it.

There’s more to this, but it leads to an important question:

what is it about technology that raises the demand for better-educated workers?

The model doesn’t really have an answer to this. In theory, at least, technology should help all workers by augmenting their productivity, if by differing amounts. But, as we know, this is not what the data says.

2. The task-polarisation model

The task model tries to explain why the real wages of some groups of workers fall even while technology is increasing the productivity of issues. In this, jobs are construed as a bundle of tasks, and the extent to which the tasks can be replaced by technology is then assessed.

In this model, automation replaces some of the tasks in the bundle. In turn, this also helps to explain why computerisation has benefitted high-skilled workers. In brief, the tasks that have been harder to automate have involved non-routine, cognitive, and inter-personal tasks that are more likely to part of a high-skilled bundle. And a bit more:

The productivity and earnings power of workers who specialize in abstract reasoning, expert judgment, and interpersonal interactions and leadership rises as the inputs into their work—information access, analysis, and communication—becomes less expensive and more productive.

Although there are also tasks (such as care work) that technology can’t easily substitute, neither does technology help their productivity.

The task model thus underscores that technological change, like most economic transformations, creates both winners and losers. … (It) further implies that a substantial component of this effect stems from the adverse impacts of technological change on the earnings of less-educated workers rather than (exclusively) the positive effect of factor-augmentation on the earnings of high-skill workers.

3. New work and task reinstatement

As described above, the task model doesn’t capture the emergence of new tasks, but we know that work is continually evolving. One innovative way of assessing this, at least in the US, is by looking at the emergence of new job titles in the Census. Acemoglu and Restrepo have integrated this into the task model by proposing that

automation displaces workers from existing job tasks as before; but now, new task creation potentially ‘reinstates’ demand for workers by generating new tasks that require human expertise.

In this version, the labour market effects of automation are effectively a race between task automation and task reinstatement. And it’s also possible to identify both technology innovations that automate tasks, and those that augment tasks, enabling reinstatement.

4. The era of AI uncertainty

So the question is whether this framework is well-placed or not to assess the impact of AI:

Does AI fundamentally change the relationship between technological change, labor demand, and inequality—and if so, how do we characterize these changes analytically?

The task model assumes that those tasks that can be automated involve explicit instructions, and AI undermines this assumption. AI is able to infer certain relationships.

Despite this, Autor concludes that the task model is still analytically useful. But AI still raises important questions:

“what work tasks will AI prove capable of accomplishing in the years (and decades) ahead?”

“what new demands for human skills and capabilities will emerge as AI displaces a growing set of traditional human work tasks? “

‘Ungrounded speculation’

Having gone through all the evidence in a lot of detail—I’ve summarised ruthlessly—Autor concludes with a section called ‘ungrounded speculation’. In brief summary here, he makes four observations:

• First, “further improvements in AI’s capabilities may accelerate the process of task automation relative to task augmentation. Broadly, this will mean that labor’s share of national income will decline further.”
• Second, AI reshapes the type of tasks (and therefore worker skills) that are substituted for and complemented by technologies. However, at the moment, “there is an upper limit to this substitution process at present… People effortlessly do extraordinary things on an ongoing basis, such as applying common sense to tease apart otherwise intractable problems.”
• Third, he does not expect AI to reach deep into the areas of low paid service work. Human interaction matters in this work, and the robotics involved is complex. The economics of displacement, in other words, are unattractive.
• Fourth, “while it is easy to imagine which tasks and what jobs will succumb to automation, it is far harder to forecast what and where new work will emerge… new innovations almost always generate new work as people deploy, master, maintain, refine, and improve new technologies, tools, and services.”

Autor is clear that the advent of AI has created greater uncertainty about the future of the labour market, and he isn’t necessarily optimistic. But it is right to underline the uncertainty. And, as he observes, a lot of these issues aren’t about technology. They’re about policy and politics.

—-

A version of this article is also published on my Just Two Things Newsletter.