On Oct. 8, California became the first state to require Ethnic Studies courses for students in order to graduate from high school. All California high schools must offer Ethnic Studies beginning in the fall of 2025, and all students must complete one semester starting with the graduating class of 2030. In signing the bill, Gov. Gavin Newsom acknowledged that “America is shaped by our shared history, much of it painful and etched with woeful injustice.” Students “must understand our nation’s full history if we expect them to one day build a more just society.”

This is long overdue. Ethnic Studies will help Californians understand how our past informs our present, including the surge in anti-Asian violence and hate during the coronavirus pandemic.

As fears about the coronavirus increased early last year, Asian Americans—and especially Asian American women—began to sound the alarm about a rise in anti-Asian violence, harassment, and hate. Asian Americans have been stabbed, beaten, pushed, spit on, harassed, and vilified based on the false assumption that they were to blame for the origin and spread of COVID-19.

But it took the mass murder of eight people—including six Asian women—working in three massage parlors in Atlanta to catapult anti-Asian violence onto a national platform.

Working low-wage jobs that required not only their physical presence but also their physical touch during a global pandemic, their lives laid bare the vulnerabilities at the intersection of race, gender, class, nativity, and citizenship.

The lives of the six Asian women massacred may seem distant and unconnected to mine, but as a woman, an Asian American woman, an immigrant, and a daughter of immigrant entrepreneurs, I know that my fate—and that of many others—is intimately connected to theirs.

A national survey by SurveyMonkey and AAPI Data fielded immediately after the Atlanta massacre in March of this year revealed that since the onset of COVID-19, 1 in 8 Asian American adults has experienced a hate incident (Figure 1), and 1 in 7 Asian American women worry consistently about being victimized (Figure 2), revealing scars born from a legacy of anti-Asian bigotry, misogyny, and medical scapegoating that date back more than 150 years. This history is absent in most American high school textbooks and university curricula.

[Figure 1]

[Figure 2]

The 1882 Chinese Exclusion Act emerged from the lesser known 1875 Page Act, which prohibited the entry of unfree laborers, prostitutes, and women who were brought for “lewd and immoral purposes.” Proposed by Horace F. Page—a Republican congressman from California who exemplified his party’s stance toward Chinese women in the mid-1870s—the bill aimed to send “the brazen harlot who openly flaunts her wickedness in the faces of our wives and daughters back to her native country.” Page linked ethnicity and gender to disease and immorality in order to justify his argument for Chinese exclusion. Chinese women were characterized as “the most undesirable of population, who spread disease and moral death among our white population,” according to California Sen. Cornelius Cole.

Science and medicine were critical to the medical and moral scapegoating of Chinese women who were deemed invariably to be prostitutes. In his role as president of the American Medical Association in 1876, J. Marion Sims warned that syphilis had reached epidemic proportions in San Francisco because of the cheap sexual favors doled out by Chinese prostitutes to white men—and alarmingly, to white boys as young as eight.

Sims drew on the testimony of Dr. Hugh H. Toland who testified before the San Francisco legislature that Chinese prostitutes were spreading a unique strain of syphilis that failed to respond to therapy and proved deadly for his white patients. Toland claimed, “There is scarcely a single day that there are not a dozen young men come to my office with syphilis or gonorrhea… in nine cases out of ten it is the ruin of them,” adding that the “whole system becomes poisoned and debilitated.”

Apart from the medical threat, Toland underscored the moral threat posed by Chinese prostitutes to young, white boys:

“I have seen boys eight and ten years old with diseases they told me they contracted on Jackson Street. It is astonishing how soon they commence in indulging in that passion… I am satisfied, from my experience, that nearly all the boys in town, who have venereal disease, contracted it in Chinatown. They have no difficulty there, for the prices are so low that they can go whenever they please. The women there do not care how old the boys are, whether five years old or more, as long as they have money.”

White prostitutes far outnumbered Chinese prostitutes, yet they were not similarly vilified because according to the Chief of Police, “white women would not allow boys of ten, eleven, or fourteen years of age to enter their houses.” Moreover, “The prices are higher, and boys of that age will not take the liberties with white women that they do in Chinatown.”

While there was no evidence to support these statements, Mary P. Sawtelle published them nevertheless in The Medico-Literary Journal in 1878 in an essay titled, “The Foul Contagious Disease. A Phase of the Chinese Question. How the Chinese Women are Infusing a Poison into the Anglo-Saxon Blood.” These testimonies were reproduced for decades.

It is worth noting that Sims, Toland, and Sawtelle were noted pillars of white, liberal society: Sims is credited as “the father of gynecology”; Toland founded a Medical College in his name which he later transferred as a gift to the University of California, San Francisco; and Sawtelle was one of the first women on the West Coast to attend medical school, and also an early suffragist. Each used science and medicine to support claims about Chinese women as both medical and moral threats to White boys as young as ten, eight, and five.

Science and medicine are beginning to confront the shameful and painful aspects of its history, including the medical scapegoating of Chinese, the unethical surgeries performed without anesthesia on female slaves, and the promotion of eugenics in order to reduce “unfit” populations. Rather than ignoring this history, Ethnic Studies demands that we confront it—recognizing that we cannot understand the present nor guide a more just future if we do not reckon with the past. A new national survey from the American Historical Association reveals that the overwhelming majority of Americans support teaching history about the harms done to other groups even if it causes discomfort or feelings of guilt. California is leading the nation in this charge. It is time for the rest of the country to follow suit.

      

Over the past two decades, partnerships between agencies primarily conducting research and those primarily administering education have been transforming both research and practice in education. Even more, the partnerships themselves are also changing with time. Often cited as a potential mechanism to bridge the longstanding gap between education research and practice, research-practice partnerships (RPPs) also hold great promise for those interested in disrupting power asymmetries, centering equity, and building new pathways for knowledge to flow. Whether and how RPPs realize these important aims are questions we regularly contend with at the National Network of Education Research-Practice Partnerships (NNERPP), a professional learning community for RPPs of all types, models, and approaches in the education sector. Together with our 50-plus members, we have engaged in a number of critical conversations resulting in deep reflections on the various aspects of partnership work, lessons on defining RPPs, and new ideas about how to know whether partnership work is making a positive impact.

Despite these advancements, the field of RPPs is still nascent. As such, it is a dynamic space, with new partnerships emerging, existing partnerships adapting, and some partnerships ending. What we define as “RPPs” is itself a dynamic concept, shaped by those doing the work. Recently, some colleagues of mine put out a terrific new report on this very topic, which offers a comprehensive, field-sourced overview of where things stand with RPPs. They include an updated definition of RPPs based on a synthesis of partnership components important to those in the field. Rather than rehash what is in that report, here I share some reflections on how the definition has changed and what it means for those of us working in (rather than researching) RPPs.

RPPs: Then and Now

A formal definition of education RPPs was first introduced in a 2013 white paper supported by the William T. Grant Foundation. Revisiting that definition now leads me to a few reactions comparing where we were then versus now.

First, the authors in the 2013 paper characterized RPPs as partnerships involving only school districts on the “practice” side. Based on the types of partnerships emerging during the early 2010s (i.e., modeled after the success of the UChicago Consortium on School Research, which featured a university and local school district working in partnership on education issues), this was largely accurate in terms of what was visible. Today’s partnerships, in contrast, involve all types of organizations, including nonprofit agencies, universities, museums, research institutions, schools, districts, state education agencies, community groups, and so on.

Second, the early definition also introduced three types of partnerships:

1. Research alliances, similar to the UChicago Consortium;
2. Design-based implementation research partnerships, coming from a learning sciences orientation; and
3. Networked improvement communities, using the principles of improvement science to engage multiple practice-side organizations in improvement aims.

Although there are still examples of all three types in the field today, we have since moved away from rigidly categorizing RPPs in this way, given the variety of partnership models in existence that both mix and transcend these definitions.

Third, it’s interesting to see how the concept of “mutualism” in RPPs has evolved as well. The early RPP definition emphasized the idea that both research and practice partners would benefit from collaborative work. In our experience, we have seen many RPPs struggle with whether their work was truly “mutually beneficial” (e.g., since producing a peer-reviewed journal article based on partnership work is still mostly a benefit that accrues to research-side partners). As you’ll see in the updated definition below, “mutualism” has since been dropped, though keeping all partners meaningfully engaged is still a priority.

In July 2021, a current landscape scan of the RPP field and an updated definition was released by the William T. Grant Foundation, based on extensive field-sourced information from individuals working in RPPs. Here’s the definition of RPPs provided (from page 5):

“A long-term collaboration aimed at educational improvement or equitable transformation through engagement with research. These partnerships are intentionally organized to connect diverse forms of expertise and shift power relations in the research endeavor to ensure that all partners have a say in the joint work.”

This updated definition is well aligned with what we, at NNERPP, are seeing across our member RPPs, and we are excited to integrate the definition in our work. A few phrases from this new definition stand out for their immediate implications for people engaging in the “practice” of RPPs.

For those just starting to think about partnership work, the new definition might serve as a checklist of sorts—and a good starting point. For example, teams might ask themselves: Do we want to engage in a long-term collaboration? How might we go about connecting diverse forms of expertise? Are we creating the structures and opportunities needed to shift power relations among our partners?

More advanced partnerships might use this definition to assess how well they are attending to the foundational elements of their collaborations. For example, teams might question: What does it mean to ensure that all partners have a say in the joint work, and are we there? Whom do we consider a partner and, importantly, who is not a partner? Why? How do we make sense of a “long-term” collaboration with short-term projects within it?

Finally, I also see the new definition anchoring our conversations around RPP effectiveness. We have come a long way in understanding what makes RPPs “work” well, but as the field itself evolves, so too will our knowledge around evaluation of these efforts. In the case of the new definition, a set of critical evaluation-like questions may include: What role do “diverse forms of expertise” play in helping move the partnership toward its stated aims? What goals do partnerships identify within the broader aims of “educational improvement” or “equitable transformation”? How do RPPs organize themselves in service of those goals? What does a first-year partnership look like along the various principles articulated, versus a three-year or five-year partnership?

Based on this new definition, NNERPP has updated its membership application, inviting teams interested in joining NNERPP to reflect on how their own work realizes what we see as the key elements of the updated definition. Specifically:

• Partnerships must have an intention of being long-term;
• Partnerships must prioritize connecting diverse forms of expertise and shifting power relations in their research-related efforts to ensure all partners have a say in the joint work; and
• Partnerships must be organized to identify practice-centered candidates for improvement, to investigate these topics, and to work collaboratively to develop solutions for improving outcomes.

As we’ve seen with the shift in how the field has defined RPPs, definitions are not static, nor should we want them to be. They evolve based on the ideas and experiences of members from the field, who themselves change over time. What is helpful about definitions, though, is that they capture succinctly a sense of where we are in that moment, and they can help us understand the boundaries of the “thing” itself.

In our case, the ongoing effort to define RPPs has both guided the work of those inside the field and communicated to those outside of it. Given these observations, we are eager and curious to see how this definition continues to evolve as the people and priorities of RPPs continue to change.

      

This year has seen a flurry of activity in state legislatures as they enacted laws that either made it easier or harder to vote. Now that many legislatures have finished their work the question is—so what? The story is, as usual, more complicated than the headlines. In some states, election reforms won’t make much of a difference. As we will see, solidly red states tended to pass restrictive voting laws and solidly blue states tended to pass expansive voting laws. But in swing states, where the presidential race or the balance of power in Congress could be won, the attempts to expand or restrict the vote could matter greatly.

The attention to election law reform began in 2020 as states struggled to hold an election in the middle of a pandemic. The result was that states adopted an unprecedented degree of innovation in dealing with absentee ballots and early voting. Millions of Americans adjusted fairly easily to the new ways of casting votes. On election day, nearly half of all Americans voted remotely in one way or the other.

Well before Election Day, former President Trump was attacking these voting innovations—asserting, among other things, that foreign countries would be able to print ballots and mail them in. In the aftermath of the 2020 election, Trump’s relentless assault on the integrity of the election continued, and in that year, state legislatures considered an unprecedented amount of voting legislation. Most of the attention has been given to new laws which suppress the vote, but other legislation has, in fact, made permanent innovations that were adopted to deal with the challenges of the pandemic.

Thanks to the work of the Brennan Center and its compilation, “Voting Laws Roundup: October 2021,” we know that 19 states have passed laws making it harder to vote. Many of these changes targeted early voting and absentee ballots and are a direct result of the fact that Donald Trump started complaining that these systems would be vulnerable to corruption months before Election Day. Although there were no instances of widespread fraud in either absentee ballots or early voting, Republicans in these states have pursued reforms as if there were. Often these states garnered headlines for their sheer pettiness and racism. While the effort to make “souls to the polls”[1] illegal in Georgia failed, the state did make it illegal to pass out snacks or water to people waiting in line to vote. In Florida, supervisors are required to assign an employee to monitor drop boxes for early ballots and the law limits the use of drop boxes to early voting hours. Many states made it harder to get absentee ballots and imposed stricter voter identification requirements.

On the other hand, according to the Brennan Center, 25 states passed laws making it easier to vote. Most of the changes made voting by mail easier and expanded opportunities for early voting. In California, the law passed now requires that all voters receive a ballot by mail. Previously, this had been up to county officials. The law now extends this to all elections and all local officials. In Maryland, voters can now choose to get ballots in the mail forever without having to request an absentee ballot for every election.

But some states adopted a mix of laws; some make voting easier and some make it more difficult. Take the case of Indiana, a very Republican state. Its reforms dealt primarily with how ballots were counted. Indiana laws limited the availability of voter drop boxes while at the same time making it easier to vote by mail. In New York, a very Democratic state, the legislature required that voters get their absentee ballot applications submitted 15 days before an election instead of seven days. On the other hand, the state adopted “no excuse” absentee balloting—meaning the voter does not have to offer an excuse for requesting an absentee ballot.

So, what will be the political effect of these new laws on the upcoming elections? One way to look at this is to divide the changes into states where the laws passed are clearly restrictive, states where the laws passed are clearly expansive, and states where the laws are mixed. Table #1 provides a quick look at the likely impact of restrictive voting laws in 11 states. Some states on this list, like Alabama for instance, are so Republican that the restrictive laws are likely to have little impact on outcomes. But Arizona, Florida, and Georgia, where presidential election results were very close, will have Senate races and as many as seven House races that are likely to be high profile and somewhat competitive as of now. Who votes in 2022 is of critical importance to the balance of power in Congress.

Table #1: States with new restrictive voting laws

Table #2 charts results for states with mixed changes in their voting laws. In the presidential contest, all these states except for New Hampshire went for Trump, and in the Senate races, only New Hampshire has a competitive race so far. However, in the House races, there could be as many as seven seats in the “moderate to high risk” column.

Table #2: States with new mixed result voting laws  

Lastly, Table #3 lists states that expanded voting through their reforms. A quick glance shows that most of these states are Democratic. In 2020, only North Dakota went for Trump. It’s possible but unlikely that these laws will affect the Senate seats since so many of them are solidly Democratic. But there are seven House seats that are at “moderate to high risk” of changing parties and 29 seats at “slight risk” of changing parties.

Table #3: States with new expansive voting laws

What these charts make clear is that restrictive laws have largely passed in Republican states and expansive ones in Democratic states. But perhaps even more important is what is missing. Michigan, for instance, isn’t in this analysis because as of early October, the Republican-controlled legislature passed a very restrictive bill that the Democratic Governor vows to veto.

At the presidential level, the effect of these changes will be felt in the handful of states (of which Michigan is one) that are most competitive in presidential elections. Voter suppression laws can hurt the Democrats and help the Republicans in states like Arizona, Florida, and Georgia, where legislative political power is still in the hands of Republicans, but population growth—and thus new voters—tends to be in Democratic areas. Voter expansion laws may not be critical to Democratic victories in California or Massachusetts, but they may help keep Virginia Democratic in a presidential election. The effect in the Senate could be minimal given that so many of the Senate seats up this year are very safe. But in the House, where the Democratic majority is very close, changes in the voting laws could be crucial.

In states that haven’t changed their voting laws, time is running short. Some legislatures may still battle it out, perhaps making no changes at all or making changes that cancel each other out. Others will not. In the end, who votes is likely to have especially big consequences for all elections.

Footnote:

[1] “Souls to the polls” refers to the practice where African American ministers would lead their congregations to early voting sites after Sunday services.

      

Historically, the 2021 off-year elections have provided great insight into the mood of voters, especially as the 2022 midterms loom. Victories for Democrats in 2017 predicted the ‘blue wave’ of 2018, just as a clean sweep in 2009 for Republicans foretold success in 2010. Key voter issues will likely center around continued COVID safety policies, economic anxiety over inflation, systemic racism, and more. How will Democrats fare in a post-Trump election? Can Republicans benefit from sagging approval for Biden? Will the polls be accurate this year? 

On November 8, Governance Studies at Brookings will host a webinar examining the results of the 2021 off-year elections, and what those results mean for both the Democratic and the Republican Parties—and the Biden Administration’s agenda. Panelists will discuss voter turnout, election administration, and ramifications for the 2022 elections.  

Viewers can submit questions for speakers by emailing events@brookings.edu or by tweeting @BrookingsGov using the hashtag #Election2021. 

      

Over four billion people worldwide are estimated to use social media by 2025. Though a majority of people use social media to engage with family and friends, people also use platforms and apps to obtain news and engage with communities on a range of issues. The polarization and sharing of news content in an era of “alternative facts” and misinformation exacerbates potential conflicts online and can reinforce false rhetoric about specific social issues and racial groups. As a result, social media provides a forum for hate speech and cyberbullying to flourish with limited understanding about tools or tactics to counter these attacks. Consequently, about 70 percent of people report doing something abusive to someone online, a majority of whom report being cyberbullied themselves. Even more troubling, nearly 90 percent of teenagers report witnessing bullying online.

While false rhetoric, hate speech, and cyberbullying have many deleterious effects, there is a silver lining: over 80 percent of youth report seeing others stand up during cyberbullying incidents and engage in bystander intervention online. This high percentage shows the power of bystander intervention—a strategy that has proven effective in ensuring the dissemination of more fact-based public health information—and holds much promise for addressing and curbing online interactions that reinforce systemic racism. Even more promising, a majority of youth report wanting to identify effective strategies to intervene in cyberbullying situations.

While existing studies mostly focus on cyberbullying related to gender or LGBTQIA issues, research focused on hate speech and cyberbullying related to race and racism is less examined. Racism contin­ues to be one of the most polarizing topics in America. Social media polarization has helped re-open the Pandora’s box that allows white supremacy and racism to wreak havoc on people’s lives. As previous research has shown, reactions to the #BlackLivesMatter movement have created echo chambers on social media that enhance hate speech and cyberbullying related to race and racism. Social media offers the opportunity for people to mask their identity, similar to the KKK hoods of the past.

This report aims to identify effective strategies to combat hate speech and misinformation online. By examining how people respond to cyberbullying, our goal is to highlight bystander intervention strategies that are effective at constructing healthy communication, calming anger and frustration, and changing attitudes. This research has broader implications for leveraging strategies, tools, and tactics, many of which have already helped address the spread of public health misinformation, and for the development and implementation of positive coping strategies for better mental and emotional health outcomes among marginalized communities.

Accordingly, we conceptualize effective bystander strategies as those that:

• are perceived by other social media users as favorable;
• alter the discussion in more positive, objective, and less antagonistic ways; and
• change the online behavior of the bully or agitator.

Through this effort, the team aims to answer the following questions: How do people combat misinformation online, particularly related to systemic racism, and, more specifically, how do people engage in bystander intervention on social media? What strategies do they use and how effective are people at changing attitudes? How do people encourage healthy coping strategies for better mental and emotional health outcomes?

Analyzing over two million tweets and posts scraped from Twitter and Reddit from 2020, we examined the effectiveness of bystander strategies used online to combat racism. These social media platforms were specifically chosen because they have inherent ranking systems that allow us to examine which strategies are viewed as most effective. On Twitter, people like and retweet messages. On Reddit, people rank comments that move them up or down in the hierarchy queue of importance to be seen more by others. Both platforms are also open allowing most people to comment on most tweets or posts.

Methodologically, we conducted a quantitative analysis of tweets and posts and a content analysis of comments. The analysis focused on four domains related to addressing racism (systemic racism, police brutality, education inequality, and employment and wealth) by using synonyms to each term to search hashtags on Twitter and search posts on Reddit that use these terms.

We found four primary types of racist discourses: stereotyping, scapegoating, accusations of reverse racism, and echo chambers. We also found four types of bystander intervention strategies, which include: call-outs, insults or mocking, attempts to educate or provide evidence, and content moderation. However, only one in six Twitter discussions and slightly less than 40 percent of Reddit discussions featured bystander action. Our findings contribute to research identifying and disseminating results about patterns regarding online communication and effective strategies to combat hate speech and misinformation about systemic racism.

In this report, we provide an overview of academic research on cyberbullying and social support, a detailed methods section regarding our analytical approach, and quantitative and qualitative findings from our investigation into how discussions about systemic racism manifest on social media.

>> Download the full report <<

      

Executive Summary

International cooperation on artificial intelligence—why, what, and how

Since 2017, when Canada became the first country to adopt a national AI strategy, at least 60 countries have adopted some form of policy for artificial intelligence (AI). The prospect of an estimated boost of 16 percent, or US$13 trillion, to global output by 2030 has led to an unprecedented race to promote AI uptake across industry, consumer markets, and government services. Global corporate investment in AI has reportedly reached US$60 billion in 2020 and is projected to more than double by 2025.

At the same time, the work on developing global standards for AI has led to significant developments in various international bodies. These encompass both technical aspects of AI (in standards development organizations (SDOs) such as the International Organization for Standardization (ISO), the International Electrotechnical Commission (IEC), and the Institute of Electrical and Electronics Engineers (IEEE) among others) and the ethical and policy dimensions of responsible AI. In addition, in 2018 the G-7 agreed to establish the Global Partnership on AI, a multistakeholder initiative working on projects to explore regulatory issues and opportunities for AI development. The Organization for Economic Cooperation and Development (OECD) launched the AI Policy Observatory to support and inform AI policy development. Several other international organizations have become active in developing proposed frameworks for responsible AI development.

In addition, there has been a proliferation of declarations and frameworks from public and private organizations aimed at guiding the development of responsible AI. While many of these focus on general principles, the past two years have seen efforts to put principles into operation through fully-fledged policy frameworks. Canada’s directive on the use of AI in government, Singapore’s Model AI Governance Framework, Japan’s Social Principles of Human-Centric AI, and the U.K. guidance on understanding AI ethics and safety have been frontrunners in this sense; they were followed by the U.S. guidance to federal agencies on regulation of AI and an executive order on how these agencies should use AI. Most recently, the EU proposal for adoption of regulation on AI has marked the first attempt to introduce a comprehensive legislative scheme governing AI.

Global corporate investment in AI has reportedly reached US$60 billion in 2020 and is projected to more than double by 2025.

In exploring how to align these various policymaking efforts, we focus on the most compelling reasons for stepping up international cooperation (the “why”); the issues and policy domains that appear most ready for enhanced collaboration (the “what”); and the instruments and forums that could be leveraged to achieve meaningful results in advancing international AI standards, regulatory cooperation, and joint R&D projects to tackle global challenges (the “how”). At the end of this report, we list the topics that we propose to explore in our forthcoming group discussions.

Why international cooperation on AI is important

Even more than many domains of science and engineering in the 21st century, the international AI landscape is deeply collaborative, especially when it comes to research, innovation, and standardization. There are several reasons to sustain and enhance international cooperation.

1. AI research and development is an increasingly complex and resource-intensive endeavor, in which scale is an important advantage. Cooperation among governments and AI researchers and developers across national boundaries can maximize the advantage of scale and exploit comparative advantages for mutual benefit. An absence of international cooperation would lead to competitive and duplicative investments in AI capacity, creating unnecessary costs and leaving each government worse off in AI outcomes. Several essential inputs used in the development of AI, including access to high-quality data (especially for supervised machine learning) and large-scale computing capacity, knowledge, and talent, benefit from scale.
2. International cooperation based on commonly agreed democratic principles for responsible AI can help focus on responsible AI development and build trust. While much progress has been made aligning on responsible AI, there remain differences—even among Forum for Cooperation on AI (FCAI) participants. The next steps in AI governance involve translating AI principles into policy, regulatory frameworks, and standards. These will require deeper understanding of how AI works in practice and working through the operation of principles in specific contexts and in the face of inevitable tradeoffs, such as may arise when seeking AI that is both accurate and explainable. Effective cooperation will require concrete steps in specific areas, which the recommendations of this report aim to suggest.
3. When it comes to regulation, divergent approaches can create barriers to innovation and diffusion. Governments’ efforts to boost domestic AI development around concepts of digital sovereignty can have negative spillovers, such as restrictions on access to data, data localization, discriminatory investment, and other requirements. Likewise, diverging risk classification regimes and regulatory requirements can increase costs for businesses seeking to serve the global AI market. Varying governmental AI regulations may necessitate building variations of AI models that can increase the work necessary to build an AI system, leading to higher compliance costs that disproportionately affect smaller firms. Differing regulations may also force variation in how data sets are collected and stored, creating additional complexity in data systems and reducing the general downstream usefulness of the data for AI. Such additional costs may apply to AI as a service as well as hardware-software systems that embed AI solutions, such as autonomous vehicles, robots, or digital medical devices. Enhanced cooperation is key to create a larger market in which different countries can try to leverage their own competitive advantage. For example, the EU seeks to achieve a competitive advantage in “industrial AI”: EU enterprises could exploit that AI without the prospect of having to engage in substantial reengineering to meet requirements of another jurisdiction.
4. Aligning key aspects of AI regulation can enable specialized firms in AI development to thrive. Such companies generate business by developing expertise in a specialized AI system, then licensing these to other companies as one part of a broader tool. As AI becomes more ubiquitous, complex stacks of specialized AI systems may emerge in many sectors. A more open global market would allow a company to take advantage of digital supply chains, using a single product with a natural language model built in Canada, a video analysis algorithm trained in Japan, and network analysis developed in France. Enabling global competition by such specialized firms will encourage healthier markets and more AI innovation.
5. Enhanced cooperation in trade is essential to avoid unjustified restrictions to the flow of goods and data, which would substantially reduce the prospective benefits of AI diffusion. While the strategic importance of data and sovereignty has in many countries given rise to legitimate industrial policy initiatives aimed at mapping and reducing dependencies on the rest of the world, protectionist measures can jeopardize global cooperation, impinge on global value chains, and negatively affect consumer choice, thereby reducing market size and overall incentives to invest in meaningful AI solutions.
6. Enhanced cooperation is needed to tap the potential of AI solutions to address global challenges. No country can “go it alone” in AI, especially when it comes to sharing data and applying AI to tackle global challenges like climate change or pandemic preparedness. The governments involved in the FCAI share interests in deploying AI for global social, humanitarian, and environmental benefit. For example, the EU is proposing to employ AI to support its Green Deal, and the G-7 and GPAI have called for harnessing AI for U.N. Sustainable Development Goals. Collaborative “moonshots” can pool resources to leverage the potential of AI and related technologies to address key global problems in domains such as health care, climate science, or agriculture at the same time as they provide a way to test approaches to responsible AI together.
7. Cooperation among likeminded countries is important to reaffirm key principles of openness and protection of democracy, freedom of expression, and other human rights. The risks associated with the unconstrained use of AI solutions by techno-authoritarian regimes— such as China’s—expose citizens to potential violations of human rights and threaten to split cyberspace into incompatible technology stacks and fragment the global AI R&D process.

The fact that international cooperation is an element of most governments’ AI strategies indicates that governments appreciate the connection between AI development and collaboration across borders. This report is about concrete ways to realize this connection.

At the same time, international cooperation should not be interpreted as complete global harmonization: countries legitimately differ in national strategic priorities, legal traditions, economic structures, demography, and geography. International collaboration can nonetheless create the level playing field that would enable countries to engage in fruitful “co-opetition” in AI: agreeing on basic principles and when possible seeking joint outcomes, but also competing for the best solutions to be scaled up at the global level. Robust cooperation based on common principles and values is a foundation for successful national development of AI.

Rules, standards, and R&D projects: Key areas for collaboration

Our exploration of international AI governance through roundtables, other discussions, and research led us to identify three main areas where enhanced collaboration would provide fruitful: regulatory policies, standard-setting, and joint research and development (R&D) projects. Below, we summarize ways in which cooperation may unfold in each of these areas, as well as the extent of collaboration conceivable in the short term as well as in the longer term.

Cooperation on regulatory policy

AI policy development is in the relatively early stages in all countries, and so timely and focused international cooperation can help align AI policies and regulations.

International regulatory cooperation has the potential to reduce regulatory burdens and barriers to trade, incentivize AI development and use, and increase market competition at the global level. That said, countries differ in legal tradition, economic structure, comparative advantage in AI, weighing of civil and fundamental rights, and balance between ex ante regulation and ex post enforcement and litigation systems. Such differences will make it difficult to achieve complete regulatory convergence. Indeed, national AI strategies and policies reflect differences in countries’ willingness to move towards a comprehensive regulatory framework for AI. Despite these differences, AI policy development is in the relatively early stages in all countries, and so timely and focused international cooperation can help align AI policies and regulations.

Against this backdrop, it is reasonable to assume that AI policy development is less embedded in pre-existing legal tradition or frameworks at this stage, and thus that international cooperation in this field can achieve higher levels of integration. The following areas for cooperation emerged from the FCAI dialogues and our other explorations.

• Building international cooperation into AI policies. FCAI governments should give effect to their recognition of the need for international engagement on AI by committing to pursue coordination with each other and other international partners prior to adopting domestic AI initiatives.
• A common, technology-neutral definition of AI for regulatory purposes. Based on the definitions among FCAI participants and the work of the OECD expert group, converging on a common definition of AI and working together to gradually update the description of an AI system, and its possible configurations and techniques, appears feasible and already partly underway. A common definition is important to guide future cooperation in AI and determines the level of ambition that can be reached by such a process.
• Building on a risk-based approach to AI regulation. A variety of governments and other bodies have endorsed a risk-based approach to AI in national strategies and in bilateral or multilateral contexts. Most notably, a risk-based approach is central to the policy frameworks of the two most prominent exemplars of AI policy development—the U.S. and the EU. These recent, broadly parallel developments have opened the door to developing international cooperation on ways to address risks while maximizing benefits. However, there remain challenges to convergence on a risk-based approach. Dialogue on clear identification and classification of risks, approaches to benefit-risk analysis, possible convergence on cases in which the risks are too high to be mitigated, and the type of risk assessment to be performed and who should perform it, would greatly benefit cooperation on a risk-based approach.
• Sharing experiences and developing common criteria and standards for auditing AI systems. The field of accountability in AI and algorithms has been the subject of wide and valuable work by civil society organizations as well as governments. The exchange of good practices and—ultimately—a common, or at least a compatible, framework for AI auditing would eliminate significant barriers to the development of a truly international market for AI solutions. It also would facilitate the emergence of third-party auditing standards and an international market for AI auditing, with potential benefits in terms of quality, price, and access for auditing services for deployers of AI. Additionally, exchange of practices and international standards for AI auditing, monitoring, and oversight would significantly help the policy community keep up to speed in market monitoring.
• A joint platform for regulatory sandboxes. Even without convergence on risk assessments or regulatory measures, an international platform for regulatory learning involving all governments that participate in FCAI and possibly others is a promising avenue for deepening international cooperation on AI. Such a platform could host an international repository of ongoing experiments on AI-enabled innovations, including regulatory sandboxes. As use of sandboxes becomes a more common way for governments to test the viability and conformity of new AI solutions under legislative and regulatory requirements, updating information on ongoing government initiatives could save resources and inform AI developers and policymakers. Aligning the criteria and overall design of AI sandboxes in different administrations could also increase the prospective benefits and impact of these processes, as developers willing to enter the global market might be able to go through the sandbox process in a single participating country.
• Cooperation on AI use in government: procurement and accountability. A natural candidate for further exchange and cooperation in FCAI is the adoption of AI solutions in government, including both “back office” solutions and more public-facing applications. The sharing of good practices and overall lessons on what works when deploying AI in government would also be an important achievement. Important areas in this respect are procurement and effective oversight of deployment.
• Sectoral cooperation on AI use cases. A sector-specific approach can ensure higher levels of regulatory certainty. In sectors like finance, key criteria such as fairness, discrimination, and transparency have long been subject to extensive regulatory intervention, and sectoral regulation must ensure continuity while accounting for the increasing use of AI. In health and pharmaceuticals, the use of AI both as a stand-alone solution and embedded in medical devices has prompted a very specific, technical discussion regarding the risk-based approach to be adopted and has already enabled valuable sectoral initiatives. The adoption of different standards and criteria in sectoral regulation may increase regulatory costs for developers willing to serve more than one sector and country with their AI solutions. In such a cross-cutting framework, examples from mature areas of regulation such as finance and health can also become a form of regulatory sandbox to model regulation for other sectors in the future.

Cooperation on sharing data across borders

Data governance is a focal area for international cooperation on AI because of the importance of data as an input for AI R&D and because of the added complexity of regulatory regimes already in place that restrict certain information flows, including data protection and intellectual property laws. Effective international cooperation on AI needs a robust and coherent framework for data protection and data sharing. There are a variety of channels addressing these issues including the Asia-Pacific Economic Cooperation group, the working group on data governance of the Global Partnership on AI, and bilateral discussions between the EU and U.S. Nonetheless, the potential impact of such laws on data available for AI-driven medical and scientific research requires specific focus as the EU both reviews its General Data Protection Regulation and considers new legislation on private and public sector data sharing.

There are other significant data governance issues that may benefit from pooled efforts across borders that, by and large, are the subject of international cooperation. Key areas in this respect include opening government data including international data sharing, improving data interoperability, and promoting technologies for trustworthy data sharing.

Cooperation on international standards for AI

As countries move from developing frameworks and policies to more concrete efforts to regulate AI, demand for AI standards will grow. These include standards for risk management, data governance, and technical documentation that can establish compliance with emerging legal requirements. International AI standards will also be needed to develop commonly accepted labeling practices that can facilitate business-to-business (B2B) contracting and to demonstrate conformity with AI regulations; address the ethics of AI systems (transparency, neutrality/lack of bias, etc.); and maximize the harmonization and interoperability for AI systems globally. International standards from standards development organizations like the ISO/IEC and IEEE can help ensure that global AI systems are ethically sound, robust, and trustworthy, that opportunities from AI are widely distributed, and that standards are technically sound and research-driven regardless of sector or application.

International standards from standards development organizations like the ISO/IEC and IEEE can help ensure that global AI systems are ethically sound, robust, and trustworthy, that opportunities from AI are widely distributed, and that standards are technically sound and research-driven regardless of sector or application.

The governments participating in the FCAI recognize and support industry-led standards setting. While there are differences in how the FCAI participants engage with industry-led standards bodies, a common element is support for the central role of the private sector in driving standards. That said, there is a range of steps that FCAI participants can take to strengthen international cooperation in AI standards. The approach of FCAI participants that emphasizes an industry-led approach to developing international AI standards contrasts with the overall approach of other countries, such as China, where the state is at the center of standards making activities. The more direct involvement by the Chinese government in setting standards, driving the standards agenda, and aligning these with broader Chinese government priorities requires attention by all FCAI participants with the aim of encouraging Chinese engagement in international AI standard-setting consistent with outcomes that are technically robust and industry driven.

Sound AI standards can also support international trade and investment in AI, expanding AI opportunity globally and increasing returns to investment in AI R&D. The World Trade Organization (WTO) Technical Barriers to Trade (TBT) Agreement’s relevance to AI standards is limited by its application only to goods, whereas many AI standards will apply to services. Recent trade agreements have started to address AI issues, including support for AI standards, but more is needed. An effective international AI standards development process is also needed to avoid bifurcated AI standards—centered around China on the one hand and the West on the other. Which outcome prevails will to some extent depend on progress in effective international AI standards development.

R&D cooperation: Selecting international AI projects

Productive discussion of AI ethics, regulation, risks, and benefits requires use cases because the issues are highly contextual. As a result, AI policy development has tended to move from broad principles to specific sectors or use cases. Considering this need, we suggest that developing international cooperation on AI would benefit from putting cooperation into operation with specific use cases. To this end, we propose that FCAI participants expand efforts to deploy AI on important global problems collectively by working toward agreement on joint research aimed at a specific development project (or projects). Such an effort could stimulate development of AI for social benefit and also provide a forcing function for overcoming differences in approaches to AI policy and regulation.

Criteria for the kinds of goals or projects to consider include the following:

1. Global significance. The project should be aimed at important global issues that demand transnational solutions. The shared importance of the issues should give all participants a common stake and, if successful, could contribute toward global welfare.
2. Global scale. The problem and the scope of the project should require resources on a large enough scale that the pooled support of leading governments and institutions adds significant value.
3. A public good. Given its significance and scale, the project would amount to a public good. In turn, the output of the project should also be a public good and both the project and the output should be available to all participants and less developed countries.
4. A collaborative test bed. Governance of the project is likely to necessitate addressing regulatory, ethical, and risk questions in a context that is concrete and in which the participants have incentives to achieve results. It would amount to a very large and shared regulatory sandbox.
5. Assessable impact. The project will need to be monitored commensurately with its scale, public visibility, and experimental nature. Participants will need to assess progress toward both defined project goals and broader impact.
6. A multistakeholder effort. Considering its public importance and the resources it should marshal, the project will need to be government-initiated. But the architecture and governance should be open to nongovernmental participation on a shared basis.

This proposal could be modeled on several large-scale international scientific collaborations: CERN, the Human Genome Project, or the International Space Station. It would also build on numerous initiatives toward collaborative research and development on AI. Similar global collaboration will be more difficult in a world of increased geopolitical and economic competition, nationalism, nativism, and protectionism among governments that have been key players in these efforts.

Recommendations

Below, we present recommendations for developing international cooperation on AI based on our discussions and work to date.

R1. Commit to considering international cooperation in drafting and implementing national AI policies.

This recommendation could be implemented within a relatively short timeframe and initially would take the form of firm declarations by individual countries. Ultimately this could lead to a joint declaration with clear commitments on the part of the governments involved.

R2. Refine a common approach to responsible AI development.

This type of recommendation requires enhanced cooperation between FCAI governments, which can then provide a good basis for incremental forms of cooperation.

R3. Agree on a common, technology-neutral definition of AI systems.

FCAI governments should work on a common definition of AI that is technology-neutral and broad. This recommendation can be implemented in a relatively short term and requires joint action by FCAI governments. The time to act is short, as the rather broad definition given in the EU AI Act is still undergoing the legislative process in the EU and many other countries are still shaping their AI policy frameworks.

R4. Agree on the contours of a risk-based approach.

Alignment on this key element of AI policy would be an important step towards an interoperable system of responsible AI. It would also facilitate cooperation among FCAI governments, industry, and civil society working on AI standards in international SDOs. General agreement on a risk-based approach could be achieved in the short term; developing the contours of a risk-based classification system would probably take more time and require deeper cooperation among FCAI governments as well as stakeholders.

R5. Establish “redlines” in developing and deploying AI.

This may entail an iterative process. FCAI governments could agree on an initial, limited list of redlines such as certain AI uses for generalized social scoring by governments; and then gradually expand the list over time to include emerging AI uses on which there is substantial agreement on the need to prohibit use.

R6. Strengthen sectoral cooperation, starting with more developed policy domains.

Sectoral cooperation can be organized on relatively short timeframes starting from sectors that have well-developed regulatory systems and present higher risks, such as health care, transport and finance, in which sectoral regulation already exists, and its adaptation to AI could be achieved relatively swiftly.

R7. Create a joint platform for regulatory learning and experiments.

A joint repository could stimulate dialogue on how to design and implement sandboxes and secure sound governance, transparency, and reproducibility of results, and aid their transferability across jurisdictions and categories of users. This recommended action is independent of others and is feasible in the short term. It requires soft cooperation, in the form of a structured exchange of good practices. Over time, the repository should become richer in terms of content, and therefore more useful.

R8. Step up cooperation and exchange of practices on the use of AI in government.

FCAI governments could set up, either as a stand-alone initiative or in the context of a broader framework for cooperation, a structured exchange on government uses of AI. The dialogue may involve AI applications to improve the functioning of public administration such as the administration of public benefits or health care; AI-enabled regulation and regulatory governance practices; or other decision-making and standards and procedures for AI procurement. This recommended action could be implemented in the short term, although collecting all experiences and setting the stage for further cooperation would require more time.

R9. Step up cooperation on accountability.

FCAI governments could profit from enhanced cooperation on accountability, whether through market oversight and enforcement, auditing requirements, or otherwise. This could combine with sectoral cooperation and possibly also with standards development for auditing AI systems.

R10. Assess the impact of AI on international data governance.

There is a need for a common understanding of how data governance rules affect AI R&D in areas such as health research and other scientific research, and whether they inhibit the exploration that is an essential part of both scientific discovery and machine learning. There is also need for a critical look at R&D methods to develop a deeper understanding of appropriate boundaries on use of personal data or other protected information. In turn, there is also a need to expand R&D and understanding in privacy-protecting technologies that can enable exploration and discovery while protecting personal information.

R11. Adopt a stepwise, inclusive approach to international AI standardization.

A stepwise approach to standards development is needed to allow time for technology development and experimentation and to gather the data and use cases to support robust standards. It also would ensure that discussions at the international level happen once technology has reached a certain level of maturity or where a regulatory environment is adopted. To support such an approach, it would be helpful to establish a comprehensive database of AI standards under development at national and international levels.

R12. Develop a coordinated approach to AI standards development that encourages Chinese participation consistent with an industry-led, research-driven approach.

There is currently a risk of disconnect between growing concern among governments and national security officials alarmed by Chinese engagement in the standards process on the one hand, and industry participants’ perceptions of the impact of Chinese participation in SDOs on the other. To encourage constructive involvement and discourage self-serving standards, FCAI participants (and likeminded countries) should encourage Chinese engagement in international standards setting while also agreeing on costs for actions that use SDOs strategically to slow down or stall standards making. This can be accomplished through trade and other measures but will require cooperation among FCAI participants to be effective.

R13. Expand trade rules for AI standards.

The rules governing use of international standards in the WTO TBT Agreement and free trade agreements are limited to goods only, whereas AI standards will apply mainly to services. New trade rules are needed that extend rules on international standards to services. As a starting point, such rules should be developed in the context of bilateral free trade agreements or plurilateral agreements, with the aim to make them multilateral in the WTO. Trade rules are also needed to support data free flow with trust and to reduce barriers and costs to AI infrastructure. Consideration also should be given to linking participation in the development of AI standards in bodies such as ISO/IEC, with broader trade policy goals and compliance with core WTO commitments.

R14. Increase funding for participation in SDOs.

Funding should be earmarked for academics and industry participation in SDOs, as well as for SDO meetings in FCAI countries and more broadly in less developed countries. Broadened participation is important to democratize the standards making process and strengthen the legitimacy and adoption of the resulting standards. Hosting meetings of standards bodies in diverse countries can broaden exposure to standards-setting processes around AI and critical technology.

R15. Develop common criteria and governance arrangements for international large-scale R&D projects.

Joint research and development applying to large-scale global problems such as climate change or disease prevention and treatment can have two valuable effects: It can bring additional resources to the solution of pressing global challenges, and the collaboration can help to find common ground in addressing differences in approaches to AI. FCAI will seek to incubate a concrete roadmap on such R&D for adoption by FCAI participants as well as other governments and international organizations. Using collaboration on R&D as a mechanism to work through matters that affect international cooperation on AI policy means that this recommendation should play out in the near term.

Download the full report

      

The quest for national AI success has electrified the world—at last count, 44 countries have entered the race by creating their own national AI strategic plan. While the inclusion of countries like China, India, and the U.S. are expected, unexpected countries, including Uganda, Armenia, and Latvia, have also drafted national plans in hopes of realizing the promise. Our earlier posts, entitled “How different countries view artificial intelligence” and “Analyzing artificial intelligence plans in 34 countries” detailed how countries are approaching national AI plans, as well as how to interpret those plans. In this piece, we go a step further by examining indicators of future AI needs.

Evaluating fulfillment through factor analysis

Clearly, having a national AI plan is a necessary but not sufficient condition to achieve the goals of the various AI plans circulating around the world; 44 countries currently have such plans. In previous posts, we noted how AI plans were largely aspirational, and that moving from this aspiration to successful implementation required substantial public-private investments and efforts.

In order to analyze the implementation to-date of countries’ national AI objectives, we assembled a country-level dataset containing: the number and size of supercomputers in the country as a measure of technological infrastructure, the amount of public and private spending on AI initiatives, the number of AI startups in the country, the number of AI patents and conference papers the country’s scholars produced, and the number of people with STEM backgrounds in the country. Taken together, these elements provide valuable insights as to how far along a country is in implementing its plan.

As analyzing each of the data elements individually presented some data challenges, we conducted a factor analysis to determine if there was a logical grouping of the data elements. Factor analysis reveals the underlying structure of data; that is, the technique mathematically determines how many groups (or factors) of data exist by analyzing which data elements are most closely related to other elements.

Given that our data included five distinct dimensions (i.e., technology infrastructure, AI startups, spending, patents and conference papers, and people), we expected that five factors would emerge, particularly since the dimensions appear to be relatively separate and distinct. But the data showed otherwise. In all, this factor analysis revealed all of the data elements fall under two factors—people-related and technology-related.

The first factor is the set of AI hiring, STEM graduates, and technology skill penetration data points, which are all associated with the people side of AI. Without qualified people, AI implementations are unlikely to be effective.

The second factor is comprised of all the non-people data elements of AI, which include computing power, AI startups, investment, conference and journal papers, and AI patent submission data points. In looking at these data elements, we realized that all of the data elements in this factor were technology-related, either from a hardware or a thought-leadership standpoint.

Given these findings, we can treat the data as containing two distinct categories: people and technology. Figure 1 shows where a select set of countries sit along these dimensions.

Interpretation of Relative Positions

The countries that are in the upper right-hand corner we dub “Leaders;“ they have both the people (factor 1) and the technology (factor 2) to meet their goals. Countries in the lower right quadrant we dub “Technically Prepared,” and because they are higher on the technology dimensions (factor 2) but lower on the people dimensions (factor 1). Those countries in the upper left quadrant we dub the “People Prepared,” and largely due to their factors higher on the people dimension (factor 1), but lower on the technology dimension (factor 2). The final quadrant—the lower left quadrant—we dub the “Aspirational” quadrant since those countries have not yet substantially moved forward in either the people or technology dimension (factor 1 and 2 respectively) in achieving their national AI strategy.

China

China is unmistakably closer to achieving its national AI strategy goals. It is both a leader in the technical dimension and a leader in the people dimension. Of note is that, while China is strongly positioned in both dimensions, it is not highest in either dimension; the U.S. is higher in the technical dimension, and India, Singapore, and Germany are all higher on the people dimension. Given the population of China and its overall investment in AI-related spending, it is not surprising that China has an early and commanding lead over other countries.

United States

The U.S., while a leader in the technology dimension, particularly in the sub-dimensions of investments and patents, ranks a relatively dismal 15th place after such countries as Russia, Portugal, and Sweden in the people dimension. This is especially clear in the sub-dimension of STEM graduates, where it ranks near the bottom. While the vast U.S. spending advantage has given it an early lead in the technology dimensions, we suspect that the overall lack of STEM-qualified individuals is likely to significantly constrain the U.S. in achieving its strategic goals in the future.

India

By contrast, India holds a small but measurable lead over other countries in the people dimension, but is noticeably lagging in the technology dimension, particularly in the investment sub-dimension. This is not surprising, as India has long been known for its education prowess but has not invested equally with leaders in the technology dimension.

Our focus on China, the U.S., and India is not to suggest that these are the only countries that can achieve their national AI objectives. Other countries, notably South Korea, Germany, and the United Kingdom are just outside of top positions, and, by virtue of generally being well-balanced between the people and the technology dimensions, have an excellent chance to close the gap

Next Steps

At present, China, the U.S., and India are leading the way in implementing national AI plans. Yet China has already hit on a balanced strategy that has thus far eluded the U.S. and India. This suggests that China needs to merely continue its strategy. However, strategy refinement is necessary for the U.S. and India to keep pace. These leaders are closely followed by South Korea, Germany, and the United Kingdom.

In future posts, we will dive deeper into both the people and technology dimensions, and will dissect specific shortfalls for each country, as well as what can be done to address these shortfalls. Anything short of a substantial national commitment to AI achievement is likely to relegate the country to the status of a second-tier player in the space. If the U.S. wants to dominate this space, it needs to improve the people dimension of technology innovation and make sure it has the STEM graduates required to push its AI innovation to new heights.

      

With the delta variant fueling an explosion of positive COVID-19 cases across the country, it is becoming even more critical to address sources of vaccination hesitancy. Researchers have provided a range of explanations for vaccination hesitancy, including partisanship and political ideology, lack of trust in the federal government, and misinformation about the safety of the vaccines; however, there is little to no research examining the role of discrimination on vaccination status.

The African American Research Collaborative/Commonwealth Fund American COVID-19 Vaccine Poll is an extensive, diverse national survey with measures of discrimination experiences, COVID-19 vaccination status, and vaccine hesitancy. Drawing on this survey, we test the relationship between discrimination and vaccination uptake. We find that a significant segment of the national sample has experienced discrimination when interacting with the health system, and that these experiences influence their decision on whether to get vaccinated. It seems that this is another example of how structural racism harms both individual Americans and the wider population.

Discrimination and Health Outcomes

Discrimination experiences are a dominant condition for racial and ethnic populations in the United States. As such, scholars in public health and social sciences have documented the negative impact that discrimination has on health outcomes.[i] Although much of this work focuses on the implications of discrimination among African Americans, there are a growing number of researchers exploring the same relationship among other groups, including Latinos and Native Americans. For instance, researchers have found that Native people report experiencing discrimination at higher rates than whites when seeking health treatment; some Native American women have even been unknowingly sterilized.

The strength of our analysis is that we isolate specific discrimination experiences in the healthcare system. For example, one of the most common experiences among patients of color is not being offered the same medical treatments or options as white Americans; however, this is not just a recent phenomenon. Even before the COVID-19 pandemic, the Discrimination in America surveys document that nationally, people of color reported higher rates of discrimination when interacting with the healthcare system. Discriminatory behavior perpetrated against people of color is often attributed to implicit bias and, unfortunately, has been even more apparent during the pandemic. For example, President of Wayne State University, Dr. M. Roy Wilson, contributed the lack of testing of Black individuals to implicit bias: “There is some evidence that African Americans with symptoms have not been tested as frequently.”   There is also evidence that Latino adults lacked access to Coronavirus testing early in the pandemic.

Similarly, racial discrimination has also increased during the pandemic. The most evident and dangerous example of discrimination was the racialization of COVID-19 blame directly targeting Asian Americans. These communities faced anti-Asian violence and other forms of hostility stroked by the rhetoric of President Trump, which has ultimately accelerated anti-Asian Xenophobia. Communities of color beyond Asian Americans have also been subject to discriminatory treatment during the pandemic, including being blamed for high infection rates. This, understandably, may impact their views toward the vaccination process.

Data and Methods

As formerly mentioned, the 2021 African American Research Collaborative/Commonwealth Fund American COVID-19 Vaccine Poll surveyed more than 12,000 individuals nationwide, with large samples of African Americans, Latinos, Asian American/Pacific Islanders, and Native Americans. It is designed to understand individual hesitancy and barriers to the COVID-19 vaccine; by drawing on these measures, as well as measures of discrimination experiences, we can directly explore how healthcare discrimination impacts vaccination uptake. This is particularly relevant as the nation takes aggressive steps to increase vaccination coverage. A full discussion of the methodology is available here.

Discussion of Descriptive Results

The survey identifies that healthcare discrimination is, unfortunately, a common experience for racial and ethnic minorities. According to the survey, more than 40% of Latino, African American, and Native American adults have experienced discrimination within the healthcare system; for instance, not being offered the best available treatment is the most identified experience for respondents. Members of the African American and Native American communities are the most likely to report negative medical experiences; however, Latinos are the most likely to report having language barriers block access to medical care. The Asian American/Pacific Islander population is the least likely of the major racial and ethnic minority groups to report medical discrimination. Coincidentally, they are also the population with the highest vaccination rate. white Americans are the group least likely to report unfair discrimination experiences in the health care system across all measures in the vaccination survey.

Next, we explore whether there is a relationship between these discrimination experiences and vaccine hesitancy. We do this by recording whether respondents answered “yes” to at least one of the discrimination questions and pairing it with their reported vaccine status. Then, we focus solely on the respondents identified as being vaccine-hesitant, which are those who reported not yet getting the vaccine. The graph below isolates respondents who have expressed vaccination hesitancy in the survey and have also experienced at least one form of discrimination experience within the health care system. The graph suggests that the influence of discrimination experiences on vaccination hesitancy is more pronounced for African American, Native American and Latino Americans.

Unvaccinated respondents were asked for their reaction to the statement that “members of their racial group face discrimination from medical professionals which makes it hard to trust the COVID-19 vaccines are safe and effective for themselves and others from their community.” This is a more direct assessment of the implications of discrimination on vaccination uptake. Among all racial and ethnic minority respondents, 39% reported that they had heard about this issue, with 16% reporting that it has made them less likely to want to get a vaccine. Consistent with the differences in discrimination experiences, African American and Native American respondents were more likely to report that racial discrimination directed towards their community makes trust in the vaccine more difficult. More specifically, 27% of Black and 22% of Native American respondents reported that racial discrimination has made them less likely to get a vaccine. This is significantly higher than Latinos (14%) or Asian American/Pacific Islanders (16%).

Conclusions

Discrimination remains a crucial factor in shaping the lives of people of color, especially in healthcare access and treatment. These historical and contemporary experiences shape willingness, or lack thereof, to receive the COVID-19 vaccine. It is not widely recognized that for many Americans of color, deciding to get the vaccine requires trust in a medical system that has not been shown to treat them fairly or prioritize their communities. Consequently, these experiences directly impact vaccination uptake through vaccination hesitancy, particularly among African Americans and Native Americans.

When discussing vaccine rates by race, it is vital to include the role of healthcare discrimination as a source of vaccine hesitancy. As the nation becomes more familiar with structural racism and systemic discrimination, this analysis provides a clear example of how powerful these concepts are in shaping the behavior of Americans who face unfair treatment in healthcare due to their race/ethnicity.

Footnotes

[i] See the following review article for a summary of this literature: https://www.annualreviews.org/doi/abs/10.1146/annurev-publhealth-040218-043750/.

About the Authors

Gabriel R. Sanchez, Ph.D., is a David M. Rubenstein Fellow in Governance Studies, a Professor of Political Science, and Founding Robert Wood Johnson Foundation Chair in Health Policy at the University of New Mexico. Sanchez is also the Director of Research at BSP Research. 

Matt Barreto, Ph.D. is Matt A. Barreto is a Professor of Political Science and Chicana/o & Central American Studies at UCLA and the co-founder of the research and polling firm BSP Research.

 Ray Block, Ph.D. is the Brown-McCourtney Career Development Professor in the McCourtney Institute and Associate Professor of Political Science and African American Studies at Penn State University.

Henry Fernandez is CEO of the African American Research Collaborative and the CEO of Fernandez Advisors, a consulting firm that counsels clients in management, planning, project development, and political strategy.  He serves as a Senior Fellow at the Center for American Progress.

Ray Foxworth, Ph.D. is Vice President of First Nations Development Institute, a Native American-led community and economic development organization, and is a visiting scholar in the Department of Political Science at the University of New Mexico.

      

While Facebook profiles may not explicitly state users’ race or ethnicity, my research demonstrates that Facebook’s current advertising algorithms can discriminate by these factors. Based on research conducted in 2020 and 2021, I used Facebook’s advertising tools to test how advertisers can use their targeting options like “multicultural affinity” groups, Lookalike Audiences, and Special Ad Audiences to ensure their ads are reaching white, African American, Asian, or Hispanic users. What I found is that discrimination by race and ethnicity on Facebook’s platforms is a significant threat to the public interest for two reasons. First, it is a violation of the existing civil rights laws that protect marginalized consumers against advertising harms and discrimination by race and ethnicity, especially in the areas of housing, employment, and credit. Second, these same Facebook advertising tools can be used to disseminate targeted misinformation and controversial political messages to vulnerable demographic groups. To solve these concerns, regulators, advocacy groups, and industry must directly address these issues with Facebook and other advertising platforms to ensure that online advertising is transparent and fair to all Americans.

Racial Bias and Discrimination in Facebook Ad Targeting

Facebook’s long-standing complaints in advertising

Over the last 5 years, Facebook has faced repeated criticism, lawsuits, and controversies over the potential for discrimination on its ad platform. Journalists have demonstrated how easy it is to exclude users whom Facebook has algorithmically classified into certain racial or ethnic affinity groups from being targeted by housing or employment ads. Researchers have also demonstrated racial and ethnic biases in Facebook’s Lookalike Audience and Special Ad Audience algorithms, which identify new Facebook users similar to an advertiser’s existing customers. In response to these allegations, Facebook has been sued by the National Fair Housing Alliance, the ACLU, the Communications Workers of America, the U.S. Department of Housing and Urban Development (HUD), and others over issues of discrimination on its advertising platform and violations of civil rights laws such as the Fair Housing Act and the Civil Rights Act of 1964, which expresses intolerance of any types of racial discrimination.

There are also ongoing controversies over how Facebook’s platform can be used by political actors, both foreign and domestic, to spread misinformation and target racial and ethnic minorities in the 2016 and 2020 election cycles. In July 2020, a high-profile boycott of Facebook’s advertising platform to “Stop Hate for Profit” was organized by civil rights and advocacy groups including the NAACP, the Anti-Defamation League, Color of Change, and other organizations over misinformation and civil rights violations. These groups called on major corporations to stop advertising on Facebook for the entire month of July. More than 1,000 large companies including Microsoft, Starbucks, Target, and others participated in the boycott.

On July 8, 2020, Facebook released its own civil rights audit conducted by Laura Murphy, former Director of the ACLU Legislative Office, and attorneys at the law firm Relman Colfax. The audit criticized Facebook for having “placed greater emphasis on free expression” over the “value of non-discrimination.” Seeking to make concrete steps towards reducing discrimination, on August 11, 2020, Facebook announced that it would retire its controversial “multicultural affinity” groups that allowed advertisers to target users whom Facebook has categorized as “African American (US),” “Asian American (US),” or “Hispanic (US – All).”

has Facebook sufficiently addressed discriminatory ad harms?

No.

Not every case of advertising discrimination may be illegal. Keeping this in mind, I devised a research study and tested the degree to which the different tools on Facebook’s ad platform can—intentionally or not—carry out racial and ethnic discrimination, following the Fair Housing Act’s criteria which makes it illegal to publish an ad that indicates “any preference, limitation, or discrimination” based on race or ethnicity.

To date, Facebook provides advertisers with four major ways to target advertisements:

1. “Detailed Targeting” options are prepackaged groups of Facebook users who share common attributes based on Facebook’s data analysis of their behaviors online,
2. “Custom Audiences” allow an advertiser to upload their own list of customers or individuals for Facebook to target,
3. “Lookalike Audiences” allow an advertiser to reach people who Facebook determines to be similar to a designated Custom Audience, and
4. “Special Ad Audiences” allow an advertiser to create a Lookalike Audience that finds people similar to their designated Custom Audience in online behavior without considering sensitive attributes like age, gender, or ZIP code, specifically for housing, employment, and credit ads which are regulated by anti-discrimination laws.

I conducted my study in January 2020 and again in January 2021, before and after the Stop Hate for Profit boycott in July 2020. As Facebook no longer offers multicultural affinity groups – “African American (US),” “Asian American (US),” and “Hispanic (US – All)” – as targeting options for advertisers in 2021, I instead examined the similar-sounding cultural interest groups that Facebook still offered as targeting options, such as “African-American Culture,” “Asian American Culture,” and “Hispanic American Culture.” In both years, I tested how many minority users could be targeted by these race- and ethnicity-related advertising options by Facebook.

Figure 1: Example of the difference in Facebook’s ad-targeting options, 2020 to 2021

I also tested if Facebook’s other advertising tools, such as Lookalike Audiences and Special Ad Audiences, can discriminate by race and ethnicity. For these tests, I used the most recent North Carolina voter registration dataset, which has voter-provided race and ethnicity information. In the North Carolina voter data, there were approximately 4.0 million white, 1.2 million African American, 90,000 Asian, and 180,000 Hispanic active voters. I created different racially and ethnically homogenous sub-samples of voters with Facebook Custom Audiences and then asked Facebook to find additional similar users to target for ads using their Lookalike and Special Ad Audience tools.

While Facebook doesn’t permit direct demographic queries for a Lookalike or Special Ad audience, I was able to leverage the daily reach estimates of Facebook’s ad planning tool to indirectly observe the racial and ethnic breakdown of a given Lookalike or Special Ad audience amongst North Carolina voters. For example, a Lookalike audience based on 10,000 African American voters in 2021 had an overlap in estimated daily reach of 139,000 users with a sample of 1 million African American voters but only an overlap of 17,000 users with a sample of 1 million white voters. This indicates that African Americans were likely over-represented in the Lookalike audience, since when intersecting the Lookalike audience with a combined 2 million voter sample that was 50% African American and 50% white at baseline, I found that 89% of the Lookalike audience’s overlap was with African American voters and only 11% with white voters. I replicated a similar process for testing the demographics of Lookalike and Special Ad Audiences based on white, Asian, and Hispanic voters. More details on the methodology can be found here.

Figure 2: Overlap of a Facebook Lookalike Audience based on African-Americans with race-based North Carolina voter samples

A major concern about algorithmic discrimination is whether computers are reinforcing the existing patterns of discrimination that humans have made in the United States. Social science researchers have found in employment, housing, lending, and other aspects of life, racial proxies such as one’s name and neighborhood can increase the degree to which human decision-makers discriminate. Thus, I used a frequently-cited computer algorithm trained on the names of 13 million voters in Florida to identify commonly used names for each demographic group and U.S. Census data to identify ethnic enclave ZIP codes. Then, I tested if using voter samples based on commonly-used racial proxies such as names and ZIP codes can affect the degree of bias in the resulting Lookalike and Special Ad Audiences. For example, would a Lookalike audience based on African American voters with common African American names and who live in majority African American ZIP codes be even more likely to over-represent African Americans above the 50% baseline?

Facebook still has a discrimination problem by race and ethnicity on its advertising platform

My research study of Facebook’s ad platform in 2020 and 2021 had three major findings. Facebook’s ad platform still offers multiple ways for discrimination by race and ethnicity despite the historic boycott it faced in 2020, and bad actors can exploit these vulnerabilities in the new digital economy.

Finding 1: The racial and ethnic targeting options by Facebook were even more accurate in 2021 than in 2020.

In 2021, certain racial and ethnic cultural interest groups that Facebook still offered to advertisers were even more accurate in targeting minority users than the old “multicultural affinity” groups Facebook retired in August 2020. For example, in 2020, 150,000 African American voters in North Carolina could be reached by Facebook’s “African American (US)” targeting option each day, which is slightly more than the 142,000 who can be reached by Facebook’s “African-American Culture” targeting option in 2021. However, there was a dramatic decrease in the number of white voters who can be reached by the same targeting options, from 428,000 in 2020 to just 109,000 in 2021. This means that in 2020, Facebook’s “African American (US)” targeting option was likely to reach nearly three times as many white users as African American users, while in 2021, Facebook’s “African-American Culture” targeting option became significantly more accurate in reaching nearly a same number of African American users while targeting 75% fewer white users.

Figure 3: Number of North Carolina voters by race reached by Facebook’s African American related targeting options, 2020 and 2021

Finding 2: Facebook’s algorithms to find new audiences for an advertiser permit strong racial and ethnic biases, which includes the algorithm that Facebook explicitly designed to avoid discrimination.

Facebook’s Lookalike and Special Ad Audiences can be biased by race and ethnicity in both 2020 and 2021 depending on the demographics of a customer list that an advertiser provides to Facebook. For example, in 2020, a Lookalike audience based on African American voters in North Carolina had a sample share of 83% African Americans, which increased to 89% in 2021. Thus, these Lookalike Audiences significantly over-represented African Americans above the expected 50% baseline sample share (Figure 4). Similarly, Lookalike Audiences based on white voters had a sample share of 73% whites in 2020 and 71% in 2021, significantly over-representing whites above the 50% baseline. Other tests found that Lookalike Audiences based on Asian or Hispanic voters would also significantly over-represent Asians or Hispanics above the expected baseline share.

I also found a high degree of racial and ethnic bias in testing the Special Ad Audience tool that Facebook designed to explicitly avoid using sensitive demographic attributes when finding similar users to an advertiser’s customer list. For example, in 2020, a Special Ad audience based on African American voters had a sample share of 83% African Americans, which decreased slightly to 76% in 2021 (Figure 4). Special Ad Audiences based on white voters had a similarly high sample share of 83% whites in 2020 and 81% in 2021. Finally, I also found that Special Ad Audiences based on Hispanic voters significantly over-represented Hispanics in both 2020 and 2021. This finding is especially problematic since it means that advertisers for housing, credit, and employment can use the Special Ad audience tool that Facebook has created for these legally protected sectors to pursue discrimination by race and ethnicity. This also undermines the goals of the 2019 legal settlement between Facebook and the ACLU, which required Facebook to create an alternative ad targeting solution for these sectors where an advertiser “cannot target ads based on Facebook users’ age, gender, race, or categories that are associated with membership in protected groups, or based on ZIP code or a geographic area”.

Finding 3: Racial proxies such as names and ZIP codes can increase the bias of Facebook’s ad targeting algorithms.

In both 2020 and 2021, the degree of racial and ethnic bias in Facebook’s Lookalike and Special Ad Audiences is higher when the ad targeting algorithm is trying to find users similar to individuals with racially stereotypical traits such as their name or neighborhood. For example, a Lookalike audience based on African American voters with commonly-given African American names and who live in majority African American ZIP codes had a very high sample share of 93% African Americans in 2020, which increased to 94% in 2021 (Figure 4). In the most extreme case, in 2021, I found that a Lookalike audience based on Asian voters with commonly-given Asian names and who live in popular Asian ZIP codes had a sample share of 100% Asians.

I also found that using racially stereotypical names and ZIP codes can embolden the opportunities for more precise discrimination in the use of Special Ad Audiences. For example, in 2020, a Special Ad audience based on African American voters with commonly-given African American names and who live in majority African American ZIP codes had a sample share of 97% African Americans, which was even higher than the 83% sample share observed when testing the Special Ad audience based on a generic sample of African American voters (Figure 4).

Figure 4: Sample share of African-American voters in Lookalike Audiences (top) and Special Ad audiences (bottom) based on lists of North Carolina voters with different traits

What’s causing these research findings?

Based on algorithmic discrimination research, it’s likely that a combination of multiple causes contributed to the biased outcomes observed in this study. One of the primary factors is that computer algorithms tend to replicate existing patterns and behaviors that already exist in society. Facebook’s Lookalike and Special Ad Audience algorithms are using the enormous amounts of data that Facebook has about its users in order to identify which users are most similar to one another and thus most likely to respond well to the same type of ads. When it comes to demographic groups, researchers have found that racial and ethnic groups tend to behave differently from each other online and that Americans tend to have very racially homogenous friend networks. Thus, users within a racial or ethnic group may appear more alike to one another in the eyes of Facebook’s algorithms than users in different groups to each other. However, this doesn’t mean that Facebook should facilitate the discriminatory targeting of racial and ethnic minorities with ads, nor that commonly-used racial proxies such as name or ZIP code should also influence the digital world created by Facebook’s advertising algorithms. After all, hard-fought civil rights laws in the U.S. have intentionally elevated society’s interest in reducing discrimination above the private interest of landlords, employers, and lenders to maximize profits through possible discrimination. It would be taking a step backwards to make new allowances for discrimination on Facebook simply because the decision-maker is now a computer instead of a human.

Given these findings, policy makers have a role in mitigating the biases being activated by companies like Facebook and other advertisers reliant on online tools.

Policymakers, regulators, and civil society groups should demand and require greater transparency by Facebook and its advertisers

Clearly, at the heart of the problems, is the lack of transparency by Facebook to the public and to its advertisers about how its ad platform can potentially discriminate by race and ethnicity. This may be exploited by discriminatory advertisers while undermining the goals of non-discriminatory ones.

For example, discriminatory advertisers may already know that the “African-American Culture” targeting option contains fewer white users than the “African American (US)” option Facebook removed in 2020. Discriminatory advertisers may also be using similar proxy variable techniques to the ones tested in this study based on racially stereotypical names and ZIP codes to create biased Lookalike and Special Ad Audiences. On the other hand, non-discriminatory advertisers may be unintentionally choosing similar targeting settings as discriminatory ones but are unaware of how Facebook’s ad platform is carrying out racially and ethnically biased targeting on their behalf.

Facebook can reverse this lack of transparency by more corporate accountability, transparency statement, disclosure to advertisers, and robust anti-discrimination engineering. However, it’s vitally important that regulators, advocacy groups, and industry groups can respond to improve upon such outcomes.

Recommendation 1: Require Facebook and advertisers to release their ad targeting data.

Advertising platforms like Facebook should provide greater transparency about the way that advertisers are using their tools to target ads. Regulators can require Facebook’s Ad Library for political, housing, employment, and credit-related ads to show the relevant metadata for establishing a “robust causal link” for racial or ethnic discrimination lawsuits. This is especially important since the Department of Housing and Urban Development’s (HUD) “Implementation of the Fair Housing Act’s Disparate Impact Standard” published on September 24, 2020 now requires a plaintiff to present evidence of a “robust causal link” in order to bring a disparate impact discrimination lawsuit in the first place. Facebook has begun to share relevant metadata for political ads with approved researchers in 2021, but doesn’t currently release any ad targeting metadata for housing, employment, and credit-related ads.

Recommendation 2: Conduct regular algorithmic bias auditing of Facebook’s ad platform.

Future civil rights audits of Facebook’s platform should test its technologies for algorithmic bias based on race, ethnicity, gender, sexual orientation, and other protected classes. In July 2020, Facebook released its first civil rights audit primarily focused on political speech, misinformation, and legal issues during the Stop Hate for Profit boycott. The civil rights groups organizing the boycott called for regular audits of Facebook’s platform. These future civil rights audits can build on the technical efforts by myself and other researchers. Ideally, if provided even deeper access to Facebook’s data and systems, these audits can go even further in examining why these racial and ethnic biases exist and how to address them. Since 2014, many large U.S. tech companies such as Facebook, Apple, Google, Microsoft, Amazon, Twitter, and others have participated in the related norm of releasing annual workforce diversity reports. Federal regulators such as the Federal Trade Commission (FTC) can also audit Facebook’s ad platform for discrimination and use their enforcement powers to address violations of Section 5 of the FTC Act, the Fair Credit Reporting Act, and the Equal Credit Opportunity Act.

Recommendation 3: Design anti-discrimination solutions using the more effective “fairness through awareness” approach.

Technology and technology policy need to move beyond “fairness through unawareness”—the idea that discrimination is prevented by eliminating the use of protected class variables or close proxies—in order to actually address algorithmic discrimination. For example, Facebook created the Special Ad Audiences tool as an alternative to Lookalike Audiences in order to explicitly not use sensitive attributes such as “age, gender or ZIP code” when considering which users are similar enough to the source audience to get included. However, my research demonstrates that Special Ad Audiences based on African Americans or whites can be biased towards the race that is more dominant in the customer list used to create the audience, just like the corresponding Lookalike Audiences.

Statistics research has labeled this phenomenon as the Rashomon or the multiplicity effect. Given a large dataset with many variables, there exists a large number of potential models that can perform approximately to equally as well as a prohibited model that uses protected class variables. Thus, even though the Special Ad Audiences algorithm for finding similar users to a customer list does not use demographic attributes in the same way as the Lookalike Audiences algorithm, the two algorithms may end up making functionally comparable decisions on which users are considered to be similar enough to get included.

There are public policy risks from continuing to implement a “fairness through unawareness” standard, which has been shown to be statistically ineffective. For example, the initially proposed language on August 19, 2019 by the Department of Housing and Urban Development (HUD) for its updated Disparate Impact Standard, would likely have wrongfully protected companies like Facebook from being sued for discrimination simply because its Special Ad Audience algorithm follows “fairness through unawareness” and does not rely on “factors that are substitutes or close proxies for protected classes under the Fair Housing Act.”

Instead, the more effective approach is to use “fairness through awareness” to design anti-discrimination tools, which could then accurately consider the experiences of different racial and ethnic groups on the platform. Thus, tech companies such as Facebook would need to learn the demographic information of their users. One way is to directly ask their users to provide their race and ethnicity on a voluntary basis, like with LinkedIn Self-ID. Another way is to use algorithmic or human evaluators to generate that information for their anti-discrimination testing tools. This was the approach taken by Airbnb’s Project Lighthouse, launched in 2020 to study the racial experience gap for guests and hosts on Airbnb. Project Lighthouse used a third-party contractor to assess the perceived race of an individual based on their name and profile picture. In December 2009, Facebook researchers used a related methodology by comparing the last names of Facebook users to the U.S. Census’ Frequently Occurring Surnames dataset in order to demonstrate how Facebook was becoming increasingly diverse over time by having more African American and Hispanic users.

The harms from discriminatory advertising on Facebook that target specific demographic groups are real. Facebook ads have reduced employment opportunities, sowed political division and misinformation, and even undermined public health efforts to respond to the COVID-19 pandemic. My research shows how Facebook continues to offer advertisers myriad tools to facilitate discriminatory ad targeting by race and ethnicity, despite the high-profile Stop Hate for Profit boycott of July 2020. To address this issue, regulators and advocacy groups can enforce and demand greater ad targeting transparency, more algorithmic bias auditing, and a “fairness through awareness” approach by Facebook and its advertisers.

In July 2020, Facebook Chief Operating Officer Sheryl Sandberg stated that its civil rights audit was “the beginning of the journey, not the end.” Faced with a long road ahead, it’s time for Facebook to continue down the path to create more equitable and less harmful advertising systems.

The Brookings Institution is a nonprofit organization devoted to independent research and policy solutions. Its mission is to conduct high-quality, independent research and, based on that research, to provide innovative, practical recommendations for policymakers and the public. The opinions expressed in this paper are solely those of the author in their personal capacity and do not reflect the views of Brookings or any of their previous or current employers.

Microsoft provides support to The Brookings Institution’s Artificial Intelligence and Emerging Technology (AIET) Initiative, and Amazon, Apple, Facebook, and Google provide general, unrestricted  support to the Institution. The findings, interpretations, and conclusions in this report are not influenced by any donation. Brookings recognizes that the value it provides is in its absolute commitment to quality, independence, and impact. Activities supported by its donors reflect this commitment.