Eric Button

Notes on the White House AI Impact Report

O
n December 5, 2022 the White House released a 19,000-word joint US-EU report on the "impact of artificial intelligence on the future of workforces in the European Union and the United States of America."

Sadly, it is an uninspired report with a very short-term outlook and a narrow view of how AI may specifically impact the future of work. The authors expressly avoid discussing AGI (which betting markets estimate will arrive in a weak form by 2027). Additionally, the report contains scarce forward-looking speculation of how AI might transform the workplace and instead cites 100 external sources, many of which are backward-looking studies of AI's impact to-date.

For these reasons, <span id="hilite" class="hilite">this document is about as convincing as if the U.S. government issued a report in 1900 about the future of aviation, but restricted it to studies of aviation's impact to-date and classified heavier-than-air technology as "beyond the scope of study."</span>

Nevertheless, the report seems balanced, addressing both the positive and negative potential of AI. While this document won't be illuminating from a technical point of view, it can be useful to understand how America and Europe view the policy challenges ahead as they grapple with AI's impact in the workplace.

Here are my notes.

Introduction

Background

  • Focuses attention to
  • ~outcomes in
  • ~~employment
  • ~~wages
  • ~~the dispersion of labor market opportunities
  • Goal:
  • ~inform approaches to AI that are consistent with an “inclusive” policy that ensures benefits for workers across the wage scale

Scope

  • Not intended to be exhaustive, but rather offer highlights
  • Goal:
  • ~to “synthesize” perspectives between the US and EU
  • ~to strengthen collaboration between US and EU with a focus on policy

Executive Summary

  • AI has potential to:
  • ~make workers more productive
  • ~make firms more efficient
  • ~spur innovations in new products/services
  • But also has potential to:
  • ~automate jobs and exacerbate inequality
  • ~lead to discrimination
  • While previous technological leaps have tended to affect “routine” tasks, AI threatens to automate “nonroutine” tasks
  • ~this can expose large groups of workers to disruption
  • Challenge for policymakers:
  • ~to foster progress and innovation while shielding workers and consumers from any resulting harm

Part 1: Overview of AI

What is AI?

  • OECD definition of AI:
  • ~“an AI system is a machine-based system that is capable of influencing the environment by producing an output (predictions, recommendations, or decisions)for a given set of objectives. It uses machine and/or human-based data and inputs to
  • ~~perceive real and/or virtual environments;
  • ~~abstract these perceptions into models through analysis in an automated manner (e.g., with machine learning), or manually; and
  • ~~use model inference to formulate options for outcomes.
  • ~AI systems are designed to operate with varying levels of autonomy”

Recent Progress on AI

  • massive growth over the last decade
  • ~now commonplace: used by services such as Pandora (music recommendations) and Google (Translate) and Facebook (personalized ads)
  • The last 5 years have seen an uptick in research of neural networks
  • ~Neural networks can learn as they’re fed more data
  • ~“Deep” neural networks:
  • ~~networks with 3 or more layers of transformation between input and output
  • ~~~can therefore learn hierarchical abstractions
  • ~~~~can  therefore more efficiently characterize complex relationships
  • progress in AI has been spurred by:
  • ~powerful enough computers
  • ~availability of training data
  • ~~value of data demonstrated by Chinese firms’ access to superior data resulting in superior AI technology

Overall Progress and Future Directions

  • AI progress has followed cycles of “spring” and “winter”
  • ~2010’s: example of a “spring”
  • ~some suggest we’re currently in AI’s “Golden Age”
  • ~~<span id="hilite" class="hilite">others suggest a coming “winter” given that some goals (e.g. autonomous vehicles) remain elusive</span>
  • AGI
  • ~concept has existed since shortly after WWII
  • ~<span id="hilite" class="hilite">Not the focus of this study, but impact of AGI would be “extraordinary”</span>
  • <span id="hilite" class="hilite">Researchers argue that machine learning’s larger’s impact may be from being a new “general purpose technology (GPT) that is also an “invention in the method invention” (IMI)
  • ~IMI’s can “reshape the nature of the innovation process and the organization of R&D”
  • ~~example: AlphaFold predicting protein structures
  • ~~ML may be able to “automate discovery” across many domains and expand set of problems that can be feasibly addressed
  • ~Previous GPT’s that were IMI’s:
  • ~~Invention of optical lenses served as eyeglasses, but
  • ~~optical lenses in form of microscopes made massive advances in biology</span>

Economic Opportunities and Challenges Coming from AI

  • AI “may” have substantial impact on economy with respect to:
  • ~productivity
  • ~growth
  • ~inequality
  • ~market power
  • ~innovation
  • ~employment
  • Policymakers could also use AI to create more efficient and equitable policy
  • Quantifying future benefits of AI is difficult because of:
  • ~uncertainty around how AI develops
  • ~we haven’t yet been able to track AI’s current contributions well from a monetary standpoint, because most consumer AI products (search engines, etc) are free
  • ~~<span id="hilite" class="hilite">to account for this, a proposed metric: GDP-B
  • ~~~quantify the tech’s benefits, rather than its costs or revenue
  • ~~~evaluates consumers’ willingness-to-pay for free goods/services
  • ~~~~e.g. Facebook would have added 0.05-0.11% to US GDP-B growth</span>
  • Costs/dangers that AI poses to society:
  • ~privacy violations
  • ~creating “anti-competitive environments”
  • ~behavioral manipulation of consumers
  • ~displacement of workers
  • ~~no guarantee that “socially optimal” mix of automation and augmentation will be reached
  • ~AI has already exacerbated social problems:
  • ~~discrimination
  • ~~~accidental introduction of racial bias
  • ~~interference with democracy
  • ~~~e.g. echo chambers in social media
  • ~~these aren’t inherent to AI but a product of the development process
  • ~~~therefore, there’s “a central role for governments in the studying, monitoring and regulating of AI”
  • While the four decades post-WWII saw technological progress result in a strong labor market for all workers, this was not repeated in the tech wave of the 1980’s
  • ~the second wave was much less “inclusive” for low-paid workers
  • AI can increase monitor-ability of workers
  • ~while can be useful, can be excessive
  • In sum, “unfettered AI could result in:
  • ~less democratic labor markets
  • ~worse working conditions
  • ~erosion of labor market institutions that currently favor workers

Part II: The Current State of AI Adoption

Adoption of AI in the US

  • few firms have adopted AI, and the ones that have are usually owned/run by younger people
  • firms like information, professional services, management and finance are most likely to adopt AI, but workers in industries such as retail trade, transportation and utilities are more likely to be exposed to AI than average

Adoption of AI in the EU

  • Similar to US
  • Denmark leads in largest share of enterprises employing AI (24%)
  • Skills and financial constraints are leading reported barriers to AI adoption in enterprises
  • ~80% cite lack of skills in internal workforce and in external labor market, as well as high costs of buying and implementing the technology

Part III: The Impact of AI on Work

  • Routine tasks vs non-routine tasks:
  • ~Routine task: follows rules and procedures
  • ~~concentrated in middle-paid occupations (machine operators, office clerks)
  • ~Non-routine tasks: multi-step, often not formally established
  • ~~concentrated in low-paid occupations and high-paid (restaurant server, doctor)
  • ~tech advancements to-date have hollowed out the middle of routine tasks, resulting in “job polarization”
  • AI overturns assumption that computers can only perform routine tasks
  • We may see:
  • ~<span id="hilite" class="hilite">stronger relative employment growth in high-paid occupations (if AI automates non-routine tasks in low-paid occupations) or
  • ~stronger relative employment growth in low-paid occupations (if AI automates non-routine tasks in high-paid occupations)</span>

What Jobs and Worker Tasks Are at Risk from AI?

  • method of testing exposure of a job to AI:
  • ~use NLP to match job tasks with patent text to find matching verb-noun pairs
  • ~~[Eric’s note: this seems quite flawed, as it ignores the concept of IMI’s above]
  • ~Findings from this approach:
  • ~~in software, exposure to automation decreases with education, with individuals in middle-wage occupations most exposed
  • ~~men are much more exposed to software than women, who gravitate toward occupations other than software, and occupations requiring complex social interaction
  • ~~most-exposed occupations include:
  • ~~~clinical lab technicians
  • ~~~chem engineers
  • ~~~optometrists
  • ~~~power plant operators
  • ~~Most-exposed categories:
  • ~~~high-skilled
  • ~~~older workers most-exposed
  • ~~~some low-skilled occupations exposed
  • ~~~~e.g. inspection, quality control
  • ~note: so far, AI has no detectable effects on labor market at the aggregate
  • ~~this could lower our sense of urgency

What New Jobs and Tasks Will Emerge from AI?

  • History is full of examples of jobs that were predicted to be doomed by automation but then resulted in augmentation and flourishing
  • ~e.g. introduction of ATMs around 1970
  • ~~many predicted the end of bank tellers
  • It’s possible that even AGI will create many new jobs
  • <span id="hilite" class="hilite">Estimates show that 60% of US employment in 2018 comprised job titles that did not exist in 1940
  • ~examples:
  • ~~“fingernail technician” (2000)
  • ~~“solar photovoltaic electrician” (2018)</span>
  • “new work” job polarization started in the 1980’s
  • ~between 1940 and 1980, most non-college workers entered middle-skilled occupations
  • ~but after 1980:
  • ~~these new workers shifted down to lower-paid personal services and
  • ~~new college-educated workers became increasingly concentrated in higher-paid occupations
  • exposure to automation/augmentation is quite binary:
  • ~e.g. hotel clerks and clergy jobs require interpersonal skills and therefore very little automation exposure at present
  • examples of new potential jobs:
  • ~digital assistant engineer
  • ~warehouse robot engineer
  • ~content-tagger on social media
  • but are these the jobs that society wants AI to create?

What Will Be the Impact of AI on Workers?

  • competing forces:
  • ~automation and augmentation
  • policymakers shouldn’t just focus on this one dynamic, but also on job redesign
  • ~jobs that will require substantial redesign:
  • ~~concierges
  • ~~credit authorizers
  • ~~brokerage clerks
  • In 2019 study, firms that adopted AI reported that:
  • ~in 15% of firms, AI increased overall employment levels
  • ~in 6% of firms, AI decreased overall employment levels
  • ~41% increased their skill demand
  • ~<2% saw decrease in skill demand
  • so far, little evidence that AI has destroyed overall job growth

Worker Skills

  • Study in Germany:
  • ~little evidence that AI has affected the number of jobs
  • ~found that workers with vocational training benefit from AI adoption in the workplace more than workers with a college degree
  • ~~potential explanations:
  • ~~~AI augments vocational workers more than it does college-level jobs
  • ~~~Germany’s high levels of vocationally trained workers (76%) suits AI to augment these skills in specialized ways

Worker Mobility Across Jobs

  • Dutch study using administrative data:
  • ~Found that once a firm adopts AI, expected annual income loss across all workers accumulates to 9% of 1 year’s earnings after 5 years
  • ~Adverse affects more common among smaller firms and older workers

Labor Market Adjustment after the Automation of Telephone Operating

  • Between 1920-1940 mechanical telephone switching replaced manual switching
  • ~was one of women’s main occupations at the time
  • ~~many left the industry as the tech adoption occurred but those that stayed tended to move to lower-paying jobs
  • ~~those that left the industry tended to be older employees
  • ~this automation did not decrease overall demand for young women in their local labor markets

What Will Be the Impact of AI on the Workplace?

  • Algorithmic management of workplaces
  • ~data collection and surveillance of workers to manage workforces
  • ~increasingly used in settings such as warehouses, retail, manufacturing, marketing, consultancy, banking, hotels, call centers, journalism, lawyers, police
  • ~algorithms can easily mold behaviors by withholding information or payouts (like how rideshare apps do for drivers today)
  • algorithmic management can erode the need for traditional employment relationships
  • ~can exacerbate wage inequality and increase safety/health risks for workers
  • “unfettered AI” can become the “glue” to accelerate this trend
  • But can also allow companies to manage their labor supply chains better

Part IV: Case Studies

[two case studies: AI in hiring, and AI in warehousing]

Part V: Conclusions

  • “Use of AI undoubtedly presents many opportunities to positively transform the economy”
  • Likely to create new jobs that never would have existed without AI
  • At same time, AI poses challenges:
  • ~“huge swaths” of workforce likely exposed
  • ~Risk that “black box” AI will violate existing laws about bias, fraud or antitrust
  • Agenda Point A:
  • ~“Investing in training and job transition services so that those employees most disrupted by AI can transition effectively to new positions where their skills and experience are most applicable.”
  • ~~Likely will be a need for large investments of training for shifting job descriptions
  • ~~But average employment length is shortening, lowering incentives for firms to invest in employee training
  • ~~We need policies that promote/subsidize intermediaries that share cost/benefit of training to reduce skill gaps esp. for workers at risk of automation
  • ~~~These intermediaries can be public, private or hybrid
  • ~~~~examples:
  • ~~~~~Public Employment Services, which offer training
  • ~~~~~outplacement offices, funded by companies that lay off workers and help displaced workers find new jobs
  • ~~~~~temp help agencies
  • Agenda Point B:
  • ~“Encouragement of development and adoption of AI that is beneficial for labormarkets.”
  • ~~Firms are driven by goal of maximizing profits, so their efforts won’t necessarily be aligned with what’s optimal for labor markets
  • ~~One solution: use public funds to “encourage and stimulate” AI research that augments work instead of automating it
  • ~~do more publicly funded academic research:
  • ~~~explore impact on wages, etc
  • ~~~develop AI ethics
  • ~~Public procurement of AI that augments workers:
  • ~~~studies show that Chinese firms with access to data-rich gov contracts develop substantially more commercial AI software
  • ~~Incentivize private firms to adopt AI that augments workers
  • ~~~be cautious and aware of incentives that firms might have to cut costs:
  • ~~~~tax structures that unduly burden firms for their headcount
  • ~~~~“aspirations of researchers” at private firms who are excited and motivated to automate vs. augment
  • Agenda Point C:
  • ~“Investing in the capacity of regulatory agencies to ensure that AI systems are transparent and fair for workers.”
  • ~~In HR, AI has risk of bias, fraud, and automatic price-setting/collusion
  • ~~Opacity of algorithms is a risk
  • ~~We should:
  • ~~~create robust standards for algorithmic audits
  • ~~~assure regulatory agencies have the access to firms when they need to perform an audit
  • ~~~build technical expertise within agencies
  • ~~~revise/craft policies that are aware of the challenges of overseeing algo-driven workplaces