# Policies for the future of work should be based on its past and present.

## Jason Furman    July 2024

I have no idea what the future of work will look like.

Will generative AI replace jobs? Complement jobs? Exacerbate inequality through skill-biased
technological change? Reduce inequality by competing with higher-skilled workers but not middle-skill
workers?

No economist, AI engineer, or policymaker knows the answers to these questions either. Economic
research holds out the hope, albeit often not realized, of coming up with a clear and widely agreed
understanding of the past, as EIG’s American Worker Project has aimed to do. But definitive guidance on the
future is too much to hope for.

Undaunted by this existential challenge, discussions about the “future of work” took off in 2009, as shown
in Figure 1—paradoxical timing given that the biggest problem of that moment was the unnecessary
absence of jobs for millions of workers who were unemployed for reasons having nothing whatsoever to do
with the future of work. The problem back then was instead the failure of policymakers to learn,
sufficiently, a critical lesson from the macroeconomic past: the need for adequate demand.

Since then, the unemployment rate has gone down, up, and down again—while the discussion of the future
of work has just gone up and up more.

In the fifteen years since this conversation took off, work has changed surprisingly little, except for a shift
to work from home which was largely ignored by most of the discussions prior to 2020. Yet the future is,
indeed, fifteen years closer. And even if the actual data does not lend any urgency, the newspaper
headlines about generative AI certainly should.


-----

**Figure 1: Frequency of 'Future of Work' mentions in printed sources**

_Source: Google Ngram_

This essay makes four arguments.

First, solutions to the future of work are not to be found in reshaping algorithms but in reshaping economic
policy.

Second, doing nothing is not a good solution. It was not how we enabled workers to cope with the massive
shifts from agriculture to manufacturing and from manufacturing to services; it should not be what we do
now.

Third, we should base our policies on what we know and can evaluate—the present and recent past of
work—rather than highly speculative guesses about the future of work.

Fourth, as much as the dignity of a job should be at the center of the economic agenda, there is only so
much we know about how to advance it. Redistribution is therefore likely to remain an important, perhaps
increasingly important, part of economic policy.

Let me now address these four points in turn.


-----

### Solutions should be found in economic policy, not in computer algorithms.

Gatherings of computer scientists and their affiliates spend their time, understandably, speculating about
[the ways that computers can solve various problems. Lately in AI circles, it is common to discuss how](https://partnershiponai.org/program/ai-labor-and-the-economy/)
computers can solve the challenges they themselves have created. AI models can reflect the biases of the
data sets on which they are trained, but other AIs can identify and offset these biases. AIs can be used to
create deepfakes, while other AIs can be used to authenticate real videos. And so on.

Perhaps. But when it comes to the potential challenges that automation, whether driven by AI or robotics,
will introduce into the workforce, an approach that mainly relies on regulations that attempt to limit jobdisplacing AI, or which try to rechannel innovation towards AI that is deliberately engineered to help
displaced workers, is unlikely to work.

It is tempting to say that AI should develop in a way that complements but does not replace work—that
rather than creating autonomous human-like bots that are modeled after different human professions, AI
should work to support the dignity and meaning of humans themselves.

This is entirely unrealistic. We already struggle to predict which innovations will be complements or
substitutes for different types of work, and sometimes we can’t even agree after the fact. Even if we did
attempt to guide innovation, we would soon confront the reality that innovation is done by many people
and businesses, and it is unlikely all of them could be dissuaded from seeking out and gaining substantial
rewards from technologies that might replace humans. And to the degree that the government tried to
redirect innovation, it is more likely to get merely less, rather than qualitatively better, technologies.

Instead, labor market policy should mostly take technology as given and instead focus on enabling people
to thrive as the nature of work changes.

### Doing nothing is not a good solution.

Why is labor market policy even needed? After all, we navigated the transition from agriculture to
manufacturing and then manufacturing to services without a permanent increase in the unemployment
rate. Surely we can do the same now?

There are two flaws with this laissez faire reasoning. The first is that there were substantial and painful
dislocations in both of those transitions. Moreover, the dislocations would have been even worse were it
not for the substantial efforts of the government, which included establishing universal high schools as the
economy shifted away from agriculture and dramatically expanding college as it shifted to services.


-----

The second flaw is that the United States already excels at flexible labor markets. Within the OECD, the U.S.
has one of the lowest ratios of its minimum wage relative to its median wage; the second lowest union
bargaining power; the lowest labor market regulation; and among the least generous disability and
unemployment insurance systems. And yet the United States still ranks toward the bottom of the OECD in
the employment rate for prime age workers, as shown in Figure 2.

**Figure 2: Prime-age (25-54) employment-population ratio, 2022 Q3**

_Source: International Labour Organization_


The issue is that while U.S. labor markets are very flexible, they are not supportive. They fail to provide the
connective tissue that both helps prepare people for jobs and then connects them to these jobs—which
may be one reason the United States has long had among the lowest employment rates for prime-age (25–
54) workers among advanced economies.

### We should base our policies on the recent past and present of work.

It is tempting to argue that policymakers should try to anticipate the future direction of work when crafting
policies that will support workers. Why invest a lot in training people for certain types of jobs that rapidly
become obsolete due to technology?

But orienting policy toward the future has two problems. The first is there are substantial errors in
predicting the future. What types of jobs will generative AI replace? How will robotics develop? Will


-----

machines develop the ability to be creative, form emotional connections, and be trusted as managers?
There is a massive likelihood that the answers given by policymakers now will prove to be wrong later.

The second problem is that even if we knew exactly what the future held in store, it would still be very
uncertain which policies would work. By definition, we have no empirical evidence of which policies work
in a world that does not yet exist. Randomized control trials and other such means of gathering evidence
can only be conducted on the circumstances of the past and present, not the future.

One part of the answer to the policy question is clear, or at least powerfully suggestive. For as long as we
have been collecting data, employment rates are higher and less volatile for people with greater degrees of
education, as shown in Figure 3. Although some of this effect is selection into education—people who are
already likely to succeed in the labor market are also more likely to pursue educational achievements—
there is a substantial amount of evidence for a big causal role. Maybe this educational effect will weaken
over the coming decades, but I would not bet against a long-term trend, especially when there is not much
else to go on.

**Figure 3: Prime age** **employment rate by education level**


**Bachelor’s degree**

**or higher**


**Some college or**

**associate degree**


**Less than a high**

**school diploma**


**High school graduates,**

**no college**


_Source: U.S. Bureau of Labor Statistics_


-----

What exactly does it mean to better educate future workers?

I am not an expert on education policy, but it should include both increasing the number of years of
education—starting earlier with universal preschool and going longer with college—and expanding federal
support for the programs of community colleges that are proven to work well.

A trickier question is whether education should focus on hard skills like coding or soft skills like
[management and interpersonal relations. Evidence from the recent past of work suggests the former are](https://www.nber.org/reporter/2017number4/value-soft-skills-labor-market)
more easily replaced by automation, so education should focus on the latter, but how exactly to do this is
not clear.

In addition, active labor market policies that help directly prepare people for jobs and also connect them
to jobs should play an important role. The United States currently is a substantial laggard in the area,
spending only 0.1 percent of GDP on such programs—about a twentieth of what Denmark spends, as
shown in Figure 4.

**Figure 4: Public expenditure on labor market policies as a percent of GDP**


_Source: Organization for economic co-operation and development_


-----

Carefully following the evidence is especially important in choosing which labor market policies to invest
[in. Programs for job training, apprenticeships, job search assistance and the like have a mixed track](https://www.annualreviews.org/content/journals/10.1146/annurev-economics-080614-115738)
[record—“mixed” in the sense that some of them really do work well and others really don’t. Policymakers](https://www.annualreviews.org/content/journals/10.1146/annurev-economics-080614-115738)
should scale up the former and eliminate the latter.

I said earlier that we cannot code our way out of the problem, but I want to qualify that statement just a
little. It is worth looking at which active labor market policies can indeed be done with the help of AI,
whether it is through digital tutors, job search assistance, or other ideas. The returns to these investments,
though, are unlikely to be large unless the government makes a clear commitment to sustaining them well
into the future.

None of these policies should take the form of “set it and forget it”. Evaluation and change should be
constant, ideally built into the way programs are designed and funded. It is possible that this will be even
more important if turnover and change is larger in the future—but regardless it is critical in just about any
scenario.

### Redistribution may have an increasingly important role to play.

The ideal set of policies would help connect people to good paying jobs with dignity, meaning, and robust
pay. We should be trying everything we can in this regard. Unfortunately, we simply do not understand well
enough how to both create such jobs and connect people to them—certainly not to the degree that we
could make this our sole response.

Moreover, given the large uncertainty in the future, there is substantial downside risk for many people in
terms of employment. And even if high employment levels are maintained, less advantaged workers might
suffer a fall in their wages—either relative to higher-income workers or even in absolute terms—as they are
forced to compete with machines and accept lower pay.

While less satisfying than employment policies, tax and transfer policies are much better understood. They
have less upside but also much less downside (a dollar is a dollar and you basically know whom it goes to).
And they have few economic constraints to their scale—although there are major political constraints,
particularly to taxation in the United States.

Some tax and transfer policies could be part of a pro-work agenda. Work subsidies could both improve the
distribution of income and simultaneously act as an incentive to work. Wage insurance helps workers who
lose their jobs and then find new employment at a lower wage, paying them enough to recover a fraction
[of their lost wages—helping them get back to work more quickly while providing insurance for a group that](https://www.nber.org/system/files/working_papers/w32464/w32464.pdf)
has been notoriously unable to benefit from other government policies (like job training).

But we should be prepared for the possibility that future automation will both accelerate productivity
growth while also concentrating the financial gains within a small group of workers or possibly even


-----

among owners of capital. This is not a reason to resist automation, as a larger pie is unambiguously good.
But it is a reason to ensure that policies are in place to share the pie better— through more progressive
taxes at the top and broader benefits for everyone else.

Overall, we should hope for the best but prepare for the worst. And the right way to prepare is to better
optimize our economic policies for the economy of today, ultimately ensuring their efficacy for the
economy of tomorrow.

_Jason Furman is the Aetna Professor of the Practice of Economic Policy jointly at Harvard Kennedy School and the_

_Department of Economics at Harvard University._

_Explore the Economic Innovation Group’s American Worker Project_ _[here.](http://eig.org/american-worker)_


-----

