Sekėjai

Ieškoti šiame dienoraštyje

2022 m. lapkričio 2 d., trečiadienis

How Does Artificial Intelligence Learn and Why ‘Elites Are Making Choices That Are Not Good News’ for Most of Us

"Even as the economic pressures that drove millions of white working-class voters to the right are moderating, the hostility this key segment of the electorate feels toward the Democratic Party has deepened and is less and less amenable to change.

“You cannot really understand the working-class rightward shift without discussing what the Democratic Party is doing,” Daron Acemoglu, an economist at M.I.T., wrote by email:

Many of the trends that negatively impacted workers, especially non-college workers, including rapid automation and trade with China, were advocated and supported by Democratic politicians. Perhaps worse from a political point of view, when these politicians were advocating such policies, they were also viewed as adopting a tone of indifference to the plight of non-college workers.

Poll data suggest that Democratic struggles with the white working class are worsening. In “Elections and Demography: Democrats Lose Ground, Need Strong Turnout,” an Oct. 22 American Enterprise institute report by Ruy Teixeira, Karlyn Bowman and Nate Moore write:

The gap between non-college and college whites continues to grow. For the first time this cycle, the difference in margin between the two has surpassed an astounding 40 points, well above the 33-point gap in 2020’s presidential contest. Republicans trail with white college voters by 13.6 points but lead with non-college whites by more than 27 points. Democrats appear stuck in the low 30s with non-college whites — no poll this month has them above 34 percent — so a repeat of Biden’s 37 percent mark appears unlikely.

David Autor, an economist at M.I.T. who has written on the role of the trade shocks that have driven white working-class voters into the arms of the Republican Party, described his assessment of the current mood of these voters in an email:

The class and cultural resentments that were inflamed by the China trade shock (alongside other technological, cultural, and political forces) are now so burned-in that I strongly suspect that they are self-perpetuating. Like a forest fire, these resentments and frustrations create their own wind that carries them forward. While the economic forces that initially fanned those flames might have abated for now, there is plenty of fuel left to consume.

“The pandemic,” Autor noted, “has actually compressed earnings inequality sharply over the last two years. This potentially reduces some of the political pressure accompanying the decline of manufacturing and erosion of non-college wages.”

While this trend would seem to favor Democrats, Autor pointed out:

Inflation has risen so fast that the fall in inequality has not actually meant earnings growth for almost anyone; rather, middle- and upper-income workers have seen larger falls in earnings power than low-income workers. It’s unfortunately cold comfort to discover that your star is rising relative to the rich because their star is falling faster than yours.

In a 2020 study, “The Work of the Future: Building Better Jobs in an Age of Intelligent Machines,” Autor, David Mindell, professor of the history of engineering and manufacturing at M.I.T., and Elisabeth Reynolds, executive director of the M.I.T. Industrial Performance Center, contend that the United States is unique among developed countries in failing to counter the negative effects of technological change on workers:

What sets the United States apart are U.S.-specific institutional changes and policy choices that failed to blunt, and in some cases magnified, the consequences of these pressures on the U.S. labor market. The United States has allowed traditional channels of worker voice to atrophy without fostering new institutions or buttressing existing ones. It has permitted the federal minimum wage to recede to near-irrelevance, lowering the floor under the labor market for low-paid workers. It has embraced a policy-driven expansion of free trade with the developing world, Mexico and China in particular, yet failed to direct the gains toward redressing the employment losses and retraining needs of workers.

Acemoglu sounded a pessimistic note in his email: “Elites are making choices that are not good news for non-college workers. In fact, they are bad news for most workers.”

He also predicted that “robots and artificial intelligence — and especially A.I. — will continue to automate a broad range of jobs, and their main impact will be to destroy ‘good’ or ‘medium-quality’ jobs for non-college workers, as well as increasingly perhaps for workers with college degrees but without postgraduate degrees. They will tend to increase inequality.”

Robots will continue to spread throughout U.S. industry, Acemoglu continued, “but there are fewer and fewer non-college jobs in this sector, so perhaps robots will not be the main issue for non-college workers.” Instead, he argued,

Artificial intelligence and other digital technologies are likely to have a bigger impact. This is both via automation and worker surveillance. Digital technologies are being increasingly used to monitor workers closely and impose worse working arrangements on them.

In a September 2022 paper, “Tasks, Automation and The Rise In U.S. Wage Inequality,” Acemoglu and Pascual Restrepo, an economist at Boston University, found that automation “accounts for 50 percent of the changes in the wage structure” from 1980 and 2016, reducing “the real wage of high-school dropout men by 8.8 percent and high-school dropout women by 2.3 percent.”

Task displacement — the replacement of workers with machines — has wide-ranging adverse impacts, they write: “A 10 percentage point higher task displacement is associated with a 4.4 percentage point decline in employment between 1980 and 2016, and a similar 3.5 percentage point increase in nonparticipation (in the work force).”

Dani Rodrik, an economist at Harvard’s Kennedy School, emailed me to say that “it is extremely unlikely that we will create an employment miracle in manufacturing.” Even if the CHIPS and Science Act, which President Biden signed in August, is “successful in reshoring some manufacturing,” he argued,

I am afraid that will do very little to create good jobs for U.S. workers without college or advanced degrees. Semiconductors and advanced manufacturing are among the most capital- and skill-intensive sectors in the economy and ramping up investment in them — as worthwhile as it may be on geopolitical grounds — is one of the least effective ways of increasing demand for labor where it is most needed.

In addition, Rodrik wrote:

Many of America’s competitors have successfully increased the share of manufacturing in G.D.P., including Taiwan and South Korea. But in none of these cases has the employment share of manufacturing bounced back up. In fact, to my knowledge, there has never been a case of sustained reversal in the downward trend of the manufacturing employment among advanced economies.

There is, Rodrik observed,

broad and compelling evidence, from Europe as well the United States, that globalization-fueled shocks in labor markets have played an important role in driving up support for right-wing populist movements. This literature shows that these economic shocks often work through culture and identity. That is, voters who experience economic insecurity are prone to feel greater aversion to outsider groups, deepening cultural and identity divisions in society and enabling right-wing candidates to inflame (and appeal to) nativist sentiment.

In an April 2021 paper, “Why Does Globalization Fuel Populism? Economics, Culture, and the Rise of Right-Wing Populism,” Rodrik wrote that he studied

the characteristics of “switchers” in the 2016 presidential election — voters who switched to Trump in 2016 after having voted for Obama in 2012. While Republican voters were in general better off and associated themselves with higher social status, the switchers were different: they were worried about their economic circumstances and did not identify themselves with the upper social classes. Switchers viewed their economic and social status very differently from, and as much more precarious than, run-of-the-mill Republican voters for Trump.

In addition to expressing concern about economic insecurity, switchers were also hostile to all aspects of globalization — trade, immigration, finance.

I asked Gordon Hanson, a professor of urban policy at Harvard’s Kennedy School, whether there was any reason for these adverse economic trends to abate. “I see none,” he said, “at least in the medium run.”

The Democrats, he continued, “have come to be seen as the party of free trade, given President Clinton pushing through both NAFTA and China’s entry to the W.T.O. and President Obama championing the Trans-Pacific Partnership — they are seen as the engineers of manufacturing job loss.”

The strongest rightward push for the non-college educated, Hanson wrote,

came during the period of major manufacturing job loss of the early 2000s, which is when we document increasing support for the right wing of the G.O.P. The absence of recovery in the 2010s in regions hurt by this job loss means that forces luring non-college workers back to the Democrats were weak. We’ve not seen new shocks that would push more of the non-college educated to the G.O.P. But nor have we seen significant recovery in manufacturing that would help them make up for lost ground.

 Reshoring in the aggregate looks to have been quite small.

In 2024, Hanson predicted, “the G.O.P. will be in position to restate its 2016 message. And, at least in places hurt by globalization, Democrats will not have obvious arguments to make in their defense.”

In a July 2022 paper, “The Labor Market Impacts of Technological Change: From Unbridled Enthusiasm to Qualified Optimism to Vast Uncertainty,” Autor describes how artificial intelligence radically enlarges the potential of robotics and automation to replace workers not only performing routine tasks but more complex procedures: “What makes a task routine is that it follows an explicit, fully specified set of rules and procedures. Tasks fitting this description can in many cases be codified in computer software and executed by machines.”

Conversely, Autor goes on to say, tasks that rely on “tacit knowledge (e.g., riding a bicycle, telling a clever joke) have historically been challenging to program because the explicit steps for accomplishing these tasks are often not formally known.”

“Artificial intelligence,” Autor writes, “overturns the second piece of the task framework — specifically, the stipulation that computers can accomplish only explicitly understood (i.e., ‘routine’) tasks. A.I. tools surmount this longstanding constraint because they can be used to infer tacit relationships that are not fully specified by underlying software.”

Autor uses the manufacture of a chair to explain the power of A.I.:

It is extraordinarily challenging to explicitly define what makes a chair a chair: must it have legs, and if so, how many; must it have a back; what range of heights is acceptable; must it be comfortable; and what makes a chair comfortable, anyway? Writing the rules for this problem is maddening. If written too narrowly, they will exclude stools and rocking chairs. If written too broadly, they will include tables and countertops.

A.I. cuts through the problem of computerizing the manufacture of a chair, according to Autor, by learning

the solution inductively by training on examples. Given a suitable database of tagged images and sufficient processing power, A.I. can infer what image attributes are statistically associated with the label “chair” and can then use that information to classify untagged images of chairs with a high degree of accuracy What rules does A.I. use for this classification? In general, we do not know because the rules remain tacit. Nowhere in the learning process does A.I. formally codify or reveal the underlying features (i.e., rules) that constitute “chair-ness.” Rather, the classification decision emerges from layers of learned statistical associations with no human interpretable window into that decision-making process.

In comparison with the non-college workers hurt by earlier levels of automation, the impact of artificial intelligence will be on better-educated, more upscale employees, in Autor’s view:

A.I. will likely eat into a lot of management and decision-making jobs that formerly required college-educated workers or even workers with graduate credentials, such as lawyers. Hence, A.I. is not “more of the same.” While the last four decades of computerization have been very good for professional, managerial workers, and not at all good for blue-collar production and white-collar office/clerical/admin workers, the A.I. era may erode the college premium that has been either high or rising since 1980.

In addition, Autor writes,

A.I. will reduce the number of person-to-person jobs in sales, food service, general customer service and tech support. The jobs that are least likely to be adversely affected at present are the lowest-wage jobs in personal services (cleaning, home health aides, groundskeeping). These jobs are still cheap to accomplish with humans and still hard and expensive to accomplish with machines. On the positive side, A.I. will surely complement the most skilled and creative people in the labor market. The question is how narrow or broad that set will be. I’m worried that it may be narrow.

Autor joined Acemoglu in arguing that policymakers can influence the direction that artificial intelligence takes:

A.I. is a general-purpose technology and could be put to many invaluable purposes: improving the quality and accessibility of health care while reducing its cost; making education more accessible, engaging, and affordable; providing real-time guidance to workers who are engaged in construction, maintenance, repair, etc.; advancing medical innovation to eradicate the worldwide disease load; improving agriculture; finding efficiencies to reduce CO emissions.

There is, however, another side to the potential of A.I., Autor wrote:

It could also be used for counterproductive purposes, for example, building history’s greatest surveillance states — whether that surveillance is done by the government (e.g., China) or by the private sector (e.g., the U.S.).

None of these capabilities is intrinsic to A.I. But we will develop those A.I. capabilities if that’s where we put our money. At present, U.S. investments in A.I. seem heavily directed at (1) selling advertising; and (2) replacing workers. If that’s where we put our money, I’m confident we’ll achieve those ends. That’s worse than a missed opportunity.

In his May 2022 essay “The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence,” Erik Brynjolfsson, a professor at Stanford’s Institute for Human-Centered Artificial Intelligence, warns that “an excessive focus on developing and deploying Human-Like Artificial Intelligence can lead us into a trap. As machines become better substitutes for human labor, workers lose economic and political bargaining power and become increasingly dependent on those who control the technology.”

There is, Brynjolfsson argues, an alternative: “When A.I. is focused on augmenting humans rather than mimicking them, humans retain the power to insist on a share of the value created. What is more, augmentation creates new capabilities and new products and services, ultimately generating far more value than merely humanlike A.I.”

But, he adds, “While both types of A.I. can be enormously beneficial, there are currently excess incentives for automation rather than augmentation among technologists, business executives, and policymakers.”

The appeal to the technological elite “of a greater concentration of technological and economic power to beget a greater concentration of political power risks trapping a powerless majority into an unhappy equilibrium” and threatens a repeat of “the backlash against free trade” that blossomed with the election of Donald Trump.

“As the economic winners gained power,” Brynjolfsson writes, they left “many workers worse off than before,” fueling

a populist backlash that led to import tariffs and other barriers to free trade. Some of the same dynamics are already underway with A.I. More and more Americans, and indeed workers around the world, believe that while the technology may be creating a new billionaire class, it is not working for them. The more technology is used to replace rather than augment labor, the worse the disparity may become, and the greater the resentments that feed destructive political instincts and actions.

Brynjolfsson is not alone in the economic community — in fact, he has widespread support — for his argument that a “moral imperative of treating people as ends, and not merely as means, calls for everyone to share in the gains of automation.

The solution is not to slow down technology, but rather to eliminate or reverse the excess incentives for automation over augmentation.”

At the moment, calls for policies to institute a moral imperative like this are limited to the universe of artificial intelligence and automation technologies, with little or no momentum in the political community. Worse yet, the bitter divisions throughout our political system suggest that the development of this momentum will be a long time coming.”


Komentarų nėra: