Professional Documents
Culture Documents
ing the relative (and perhaps absolute) in- technology may disproportionately aug- POLICY FORUM
come of the previously highly-paid workers ment lower-skilled labor, reducing income
whose skills have been automated.
It is too early to determine with much cer-
tainty how this will play out for AI, whether
inequality. This, at the very least, calls into
question whether a change in the innova-
tor’s mindset is needed: Task automation
Generative
the impact on any particular job will be
positive or negative. Research is beginning
to emphasize which jobs are most likely to
may be a path to substantially improved
labor productivity.
This potential to reverse the recent
AI meets
be affected rather than lost (8). For exam-
ple, that classification tasks such as image
recognition can be done with AI will affect
trend toward skill-biased technical change
does not mean that AI is without risk. Other
concerns remain, including those related to
copyright
workers whose jobs involve classification privacy, liberty, democracy, and monopoly Ongoing lawsuits could
tasks, such as radiologists (7). Recent work power (3). Our emphasis is on understand- affect everyone who uses
examining differences between generative ing that one person’s automation is anoth-
AI (specifically, LLMs) and nongenerative er’s augmentation, and that it is difficult for generative AI
AI [as described in (7)] shows that millions engineers or policy-makers to pick which
of jobs have the potential to be affected by particular innovation will increase or re- By Pamela Samuelson
LLMs. Notably, these studies emphasize duce inequality overall. We believe that both
G
that “affect” does not mean “replace.” For regulators and engineers should be careful enerative artificial intelligence (AI)
many jobs, automating some aspects of the in shutting down a particular technology is a disruptive technology that is
workflow might increase productivity, the trajectory because it appears to automate widely adopted by members of the
wages of workers who have that job, and human work. In the process of automating general public as well as scientists
the number of workers hired to do that job. some work, other work can be augmented. and technologists who are enthusi-
Even when some jobs get automated, that Often, our analysis suggests that such astic about the potential to acceler-
might complement the tasks done by other augmentation from AI will increase the job ate research in a wide variety of fields. But
workers. Many empirical exercises [for ex- productivity of less-skilled workers who some professional artists, writers, and pro-
ample, (7–9)] emphasize the direct impact can now perform at levels achieved by their grammers fiercely object to the use of their
on jobs, but they do not explore the jobs skilled counterparts. This suggests that skill creations as training data for generative AI
that might be enhanced through comple- premia that have contributed to widening systems and to outputs that may compete
mentary production processes. For exam- inequality may be eroded. Thus, it is quite with or displace their works (1, 2). Lack of
ple, in January 2023, there were 186,417 job plausible that the use of AI to automate attribution and compensation for use of
postings in the United States that specified tasks will both increase productivity and their original creations are other sources
language skills (such as Spanish Language decrease income inequality. If so, then we of aggravation to critics of generative AI.
or American Sign Language), or about 5% of may want more automation, not less. j Copyright lawsuits that are now underway
the total job postings (see SM). Automating in the United States have substantial im-
REF ERENCES AND NOTES
language translation would directly affect plications for the future of generative AI
1. E. Brynjolfsson, Daedalus 151, 272 (2022).
many of these jobs. At the same time, many 2. J. S. Gans, A. Leigh, Innovation + Equality: Creating a systems. If the plaintiffs prevail, the only
other jobs that do not require language Future that is more Star Trek than Terminator (MIT Press, generative AI systems that may be lawful
skills would also be affected. For example, a 2019). in the United States would be those trained
3. D. Acemoglu, S. Johnson, Power and Progress: Our
recent study showed that small businesses Thousand-Year Struggle Over Technology and Prosperity on public domain works or under licenses,
that used a rudimentary automated lan- (PublicAffairs, 2023). which will affect everyone who deploys
guage translation tool on eBay experienced 4. D. Acemoglu, P. Restrepo, Am. Econ. Rev. 108, 1488 generative AI, integrates it into their prod-
(2018).
a 17.5% increase in exports to markets where 5. C. Goldin, L. Katz, The Race Between Education and ucts, and uses it for scientific research.
that language is used (15). Automation of Technology (Harvard Univ. Press, 2008). Conflicts between creators of copy-
some jobs could create opportunities for 6. D. H. Autor, Science 344, 843 (2014). righted works and developers of technolo-
7. E. Brynjolfsson, T. Mitchell, Science 358, 1530 (2017).
those whose work would appear to be un- 8. E. Felten, M. Raj, R. Seamans, Strateg. Manage. J. 42,
gies that enable the use of those creations
affected, as measured with the tasks and 2195 (2021). in unexpected and innovative ways is noth-
skills involved in current workflows. 9. T. Eloundou, S. Manning, P. Mishkin, D. Rock, ing new. In the early 20th century, the dis-
arXiv:2303.10130 [econ.GN] (2023). ruptive technology of the day was player
10. A. Agrawal, J. S. Gans, A. Goldfarb, The Economics of
CONCLUSION Artificial Intelligence: An Agenda (Univ. Chicago Press, pianos. Music copyright owners sued the
Many economists who have studied the im- 2019). makers of piano rolls, claiming that rolls of
pact of automation on labor markets have 11. K. Kanazawa, D. Kawaguchi, H. Shigeoka, Y. Watanabe, their musical compositions were infringe-
“AI, skill, and productivity: The case of taxi drivers,” no.
argued recently that the direction of AI w30612, National Bureau of Economic Research (2022). ments. Subsequent copyright-disruptive
research needs to be changed away from 12. E. Brynjolfsson, D. Li, L. Raymond, “Generative AI at technologies have included cable televi-
automating tasks to focusing on overall work,” National Bureau of Economic Research (NBER) sion, photocopiers, videotape recording
working paper 31161 (NBER, 2023).
job augmentation. The implicit argument 13. D. H. Autor, L. Katz, M. Kearney, Rev. Econ. Stat. 90, 300 machines, and MP3 players, each of which
is that a focus on augmentation will lead (2008). (except photocopiers) attracted copyright
to more complementarity with lower-wage 14. A. Agrawal, J. S. Gans, A. Goldfarb, Power and Prediction industry challenges (all of which failed in
(Harvard Business Review Press, 2022).
labor and more new tasks. However, many 15. E. Brynjolfsson, X. Hui, M. Liu, Manage. Sci. 65, 5449 the courts, although Congress sometimes
recent advances in AI that have been de- (2019). later extended protections in the after-
veloped with the explicit goal of task auto- math of failed lawsuits).
SUPPLEMENTARY MATERIALS
mation have appeared to increase worker
science.org/doi/10.1126/science.adh9429
productivity; that is, task automation has Berkeley Law School, University of California Berkeley,
been labor augmenting. Furthermore, AI 10.1126/science.adh9429 Berkeley, CA, USA.Email: psamuelson@berkeley.edu
A countervailing consideration is that Getty websites has harmed a licensing mar- programs. The Court thought this use was
the visual artists on whose works Stable ket. The class-action claim against Stability consistent with the constitutional objective
Diffusion trained made their works avail- AI is weaker because Stability AI could not of copyright to promote creative progress.
able on the open internet, as did Field in have gotten a license from the class of visual The public greatly benefited from Android’s
the Google, Inc., case. artists whose works were ingested to con- existence and the availability of large num-
Transformative purposes tend to have struct the Stable Diffusion model. bers of apps that ran on that platform.
spillover effects on other fair-use factors, es- The existence of a licensing market (or Stability AI will almost certainly channel