Posts

The Future of Work Part 4: The Kinds Of Jobs That Are At Risk

Recent improvements in hardware, a massive increase in the number of processors available, and new math tools have increased concerns that computers may soon replace millions of workers. The shorthand for this is Artificial Intelligence, although the term seems like hyperbole considering the kinds of things computers can do at present. The Obama White House issued a paper on this issue, Artificial Intelligence, Automation and the Economy, which can be found here. It cites two studies of the impact of AI on automation over then next 10 years or so. One, by the OECD, estimates about 9% of US jobs may be lost to automation. The other is a more interesting 2013 paper by two professors at Oxford, Carl Benedikt Frey and Michael A. Osborne, estimating that as many as 49% of US jobs could be lost or seriously affected over 10 or so years.

The Frey-Osborne Paper is here. Frey is a professor in a public policy college, and Osborne is in the engineering college; they aren’t economists. Perhaps for that reason, the introductory sections are instructive on the history of technological change and some of its effects on society. The technical approach of the Frey-Osborne Paper is to identify the bottlenecks that make it difficult to automate the tasks needed in a specific job. They use machine learning to identify patterns in the skills needed by specific jobs.

The authors identify three main bottlenecks to automation:

1. Tasks requiring perception and manipulation. P. 24
2. Tasks requiring creative intelligence. P. 25
3. Tasks requiring social intelligence. P. 26

The O-NET database of jobs is managed by the US Department of Labor. The current version contains detailed descriptions of job tasks for 903 occupations. Here are the top eight tasks of 21 listed for forest firefighter, one of the bright future jobs according to O-NET,:

Rescue fire victims, and administer emergency medical aid.

Establish water supplies, connect hoses, and direct water onto fires.

Patrol burned areas after fires to locate and eliminate hot spots that may restart fires.

Inform and educate the public about fire prevention.

Participate in physical training to maintain high levels of physical fitness.

Orient self in relation to fire, using compass and map, and collect supplies and equipment dropped by parachute.

Fell trees, cut and clear brush, and dig trenches to create firelines, using axes, chainsaws or shovels.

Maintain knowledge of current firefighting practices by participating in drills and by attending seminars, conventions, and conferences.

Frey and Osborne describe their methodology as follows:

First, together with a group of [machine learning] researchers, we subjectively hand-labelled 70 occupations, assigning 1 if automatable, and 0 if not. For our subjective assessments, we draw upon a workshop held at the Oxford University Engineering Sciences Department, examining the automatability of a wide range of tasks. Our label assignments were based on eyeballing the O-NET tasks and job description of each occupation.

They identified nine variables related to the three bottlenecks and assigned levels of difficulty of the variables in carrying out each task, high, medium, or low. Then they verified their data, and used it as training data in a machine learning program. The paper gives a description of the way they prepared and ran the rest of the O-NET data through the trained machine to estimate the likelihood that each job would be automated over the next 10 years or so. They produced a chart showing the likely effects of AI on categories of jobs. The following chart shows the results of their work.

The authors say that large numbers of transportation and logistics workers, office workers and administrative support workers are at risk. They also think many service workers are at risk as robots become more efficient. They think people whose jobs require great manual dexterity and perception, or high levels of creativity, or strong social intelligence are reasonably safe in the near term. They assert that low-skill workers will have to move to jobs in the service sector that require these skills, and will have to sharpen their own through training and education.

There have been several articles on this issue lately. This one by Reuters says that investors think the future is in automation. Since the election shares in companies working in that area are up dramatically as is an ETF in the sector. Reuters says that this means that investors think that Trump’s assertion he will increase jobs in the manufacturing sector will not happen. Instead, as the cost of advanced technology drops labor becomes expendable. Any increase in manufacturing will have little effect on overall unemployment, as displaced workers move to other jobs with the same employers doing “value-added” tasks.

Matthew Yglesias goes a step farther in this 2015 post at Vox. He says the big problem in job growth in the US is the lack of increase in productivity due to inadequate automation. He thinks rising productivity is essential to higher wages, or more likely a reduction in the time spent working. Yglesias lays out the case for not worrying. He ignores, as all economists do, the possibility that the returns from work might be shared more equitably between capital and labor. His relentless optimism contrasts with the lived experience of millions of Americans, the real lives that gave us Trumpism.

I wonder what Yglesias makes of this article in the Guardian discussing the efforts of the billionaire Ray Dalio to create software to manage the day-to-day operations of the world’s largest hedge fund in accordance with “… a set of principles laid out by Dalio about the company vision.” The article provides a more pessimistic view of the future even for management work.

I don’t have an opinion about these forecasts or the reasoning behind them. Yglesias says people will work less, but doesn’t explain how workers who have no bargaining power will be able to increase their income enough to have free time. Dalio must think that he is so wise that his AI automaton will replicate his success forever, and that his competitors won’t take advantage of the rigidity of his principles.

Suppose that the investors described by Reuters are right, that manufacturing increases but without increased employment in the sector. What will all those Trump voters do next? Change their minds about what they want from the economy and the government that fosters it, and live happily ever after?

I think both Yglesias and Dalio are so steeped in neoliberal economics with its model of human beings as Homo Economicus that they assume these changes will come about smoothly. Nothing else will change; there are no dynamic tipping points. No large number of human beings will raise hell. There will be no feedback effects. The displaced of all ages will just retrain to some other job and/or resign themselves to their reduced lives. They won’t resist, or riot, or insist on government protection, or demand a completely new system. Investment bankers will blandly accept the judgment of computers as to their value and will not insist on being treated like superstars even if the machine says they are just gas giants.

Yglesias and Dalio are wrong. That is precisely what history says won’t happen.

The Future of Work Part 2: The View From the White House

Top advisors in the Obama Administration published a report titled Artificial Intelligence, Automation, and the Economy in December 2016, which I will call the AI Paper. It’s a statement of the views of the Council of Economic Advisers, the Domestic Policy Council, the Office of Science and Technology Policy, the National Economic Council, and the US Chief Technology Officer, combining their views into a single report. There is a brief Executive Summary which gives a decent overview of the substance of the report, followed by a section on the economics of artificial intelligence technology and a set of policy recommendations. It’s about what you’d expect from a committee, weak wording and plenty of caveats, but there are nuggets worth thinking about.

First, it would be nice to have a definition of artificial intelligence. There isn’t one in this report, but it references an earlier report; Preparing For the Future of Artificial Intelligence, which dances around the issue in several paragraphs. Most of the definitions are operational: they describe the way a particular type of AI might work. But these are all different, just as neural network machine learning is different from rules-based expert systems. So we wind up with this:

This diversity of AI problems and solutions, and the foundation of AI in human evaluation of the performance and accuracy of algorithms, makes it difficult to clearly define a bright-line distinction between what constitutes AI and what does not. For example, many techniques used to analyze large volumes of data were developed by AI researchers and are now identified as “Big Data” algorithms and systems. In some cases, opinion may shift, meaning that a problem is considered as requiring AI before it has been solved, but once a solution is well known it is considered routine data processing. Although the boundaries of AI can be uncertain and have tended to shift over time, what is important is that a core objective of AI research and applications over the years has been to automate or replicate intelligent behavior. P. 7.

That’s circular, of course. For the moment let’s use an example instead of a definition: machine translation from one language to another, as described in this New York Times Magazine article. The article sets up the problem of translation and the use of neural network machine learning to improve previous rule-based solutions. For more on neural network theory, see this online version of Deep Learning by Ian Goodfellow and Yoshua Bengio and Aaron Courville. H/T Zach. The introduction may prove helpful in understanding the basics of the technology better than the NYT magazine article. It explains the origin of the term “neural network” and the reason for its replacement by the term “deep learning”. It also introduces the meat on the skeletal metaphor of layers as used in the NYT magazine article.

The first section of theAI Paper takes up the economic impact of artificial intelligence. Generally it argues that to the extent it improves productivity it will have positive effects, because it decreases the need for human labor input for the same or higher levels of output. This kind of statement is an example of what Karl Polanyi calls labor as a fictitious commodity. The AI Paper tells us that productivity has dropped over the last decade. That’s because, they say, there has been a slowdown in capital investment, and a slowdown in technological change. Apparently to the writers, these are unconnected, but of course they are connected in several indirect ways. The writers argue that improvements in AI might help increase productivity, and thus enable workers to “negotiate for the benefits of their increased productivity, as discussed below.” P. 10.

The AI Paper then turns to a discussion of the history of technological change, beginning with the Industrial Revolution. We learn that it was good on average, but lousy for many who lost jobs. It was also lousy for those killed or maimed working at the new jobs and for those marginalized, wounded and killed by government and private armies for daring to demand fair treatment. These are presumably categorized as “market adjustments”, which, according to the AI Paper, “can prove difficult to navigate for many.” P. 12 Recent economic papers show that Wages for those affected by these market adjustments never recover, and we can blame the workers for that: “These results suggest that for many displaced workers there appears to be a deterioration in their ability either to match their current skills to, or retrain for, new, in-demand jobs.” Id.

The AI Paper then takes up some of the possible results of improvements in AI technology. Job losses among the poorest paid employees are likely to be high, and wages for those still employed will be kept low by high unemployment. Jobs requiring less education are likely to be lost, while those requiring more education are likely safer, though certainly not absolutely safe. The main example is self-driving vehicles. Here’s their chart showing the potential for driving jobs that might be lost.

That doesn’t include any knock-on job losses, like reductions in hiring at roadside restaurants or dispatchers.

It also doesn’t include the possible new jobs that AI might create. These are described on pp 18-9. Some are in AI itself, though as the NYT magazine article shows, it doesn’t seem like there will be many. Some new jobs will be created because AI increases productivity of other workers. Some are in new fields related to handling AI and robots. That doesn’t sound like jobs for high school grads. Most of the jobs have to do with replacing infrastructure to make AI work. Here’s Dave Dayen’s description of the need to rebuild all streets and highways so autonomous vehicles can work. Maybe all those displaced 45 year old truck drivers can get a job painting stripes on the new roads. There are no numerical estimates of these new jobs.

The bad news is buried in Box 2, p. 20. Unless there are major policy changes, it’s likely that most of the wealth will be distributed to the rich. And then there’s this:

In theory, AI-driven automation might involve more than temporary disruptions in labor markets and drastically reduce the need for workers. If there is no need for extensive human labor in the production process, society as a whole may need to find an alternative approach to resource allocation other than compensation for labor, requiring a fundamental shift in the way economies are organized.

That certainly opens a new range of issues.

Update: the link to the AI Paper has been updated.