Main Findings

It’s widely understood that the Longtermist ecosystem is quite racially homogenous. This analysis seeks to quantify and elaborate on that observation, to enable various stakeholders to make more informed decisions.

My review of 459 unique individuals listed on the team pages of 21 Longtermist organizations produced the following findings:

• People of Color (POC) make up 13-14% of staff at the 21 Longtermist organizations in my sample. This is less diverse, and in many cases much less diverse, than the communities from which these organizations tend to draw from (such as top US Computer Science programs where POC often make up a majority of students).
• At about half the organizations in my sample (10 out of 21), the core staff is entirely white. At every organization except Open Philanthropy’s AI Fellowship program, POC account for <30% of staff.
• Open Philanthropy’s AI Fellowship program was an extreme outlier, with 9 out of 15 fellows (60%) being POC. This fellowship accounted for just 6% of the core staff in my sample, but 26% of the POC. One of the program’s explicit goals is to improve diversity in the ecosystem, suggesting this sort of intentional effort can successfully bring talented POC into the Longtermist community.
• Organizations/programs that are based at universities are much more racially diverse than other types of Longtermist organizations (typically independent charities). At “University” organizations, POC make up 20% of core staff positions, compared to only 3% at “Non-University” organizations.
• POC appear to be particularly underrepresented in the most influential roles at Longtermist organizations. Only 4% of leadership positions are held by POC. However, POC account for 19-25% of support staff.
• Collectively, these findings suggest that the Longtermist ecosystem is significantly underrepresenting and underutilizing POC. At the same time, Open Philanthropy’s AI Fellowship program suggests improvements are possible if prioritized. Potential strategies for improving diversity are discussed below.

Methodology

Sample

To conduct this analysis, I looked at the “team” pages on the websites of 21 Longtermist organizations/programs, which encompassed 459 unique individuals filling 558 positions (some people hold roles at multiple organizations). This data was collected in late February 2020.

I selected the organizations by coming up with a list of prominent groups working in the ecosystem, supplemented this initial list with grantees of the Long-Term Future Fund, and made some adjustments based on feedback from people working in the Longtermist ecosystem (none of whom suggested adding other organizations). This process required some judgment calls, such as including CFAR (because while it is not explicitly Longtermist, it has received grants from the Long-term Future Fund and has trained many people in the Longtermist space). Similarly, I chose to include the Long-Term Future Fund itself (because despite being a volunteer-run effort rather than an organization, it’s role as funder makes it particularly influential).

In almost every case, I included all team members listed. The only exception was Open Philanthropy, where I included the leadership and people whose job titles/descriptions explicitly suggest they work full-time on Longtermist causes. I did this because Open Philanthropy plays a critical part in the Longtermist ecosystem, but has a scope that exceeds Longtermism. This methodology excluded some Open Philanthropy employees, including some POC, who work part-time on Longtermist causes. (CEA and 80,000 Hours also do work besides Longtermism, but I included all of their employees as it was not clear which of them do Longtermist work.)

Racial Classification

For each person in my sample, I tried to determine whether they are white or a person of color (POC) based on the pictures displayed on the relevant team page (in some cases, I performed a quick google search to try to find additional information). I included people I believed to be multiracial in the POC category. This is admittedly an imperfect approach, and I expect I misclassified some people. I sincerely apologize to anyone I misclassified. I believe it is unlikely that this imprecision materially changes the findings of this analysis, as there generally didn’t seem to be much ambiguity.

Note: While there are of course many dimensions of diversity, for purposes of brevity my use of that term throughout this analysis will refer to racial diversity unless otherwise stated.

Role Classification

Different organizations have different criteria for listing people on their team pages. Some organizations use loose criteria, and include interns, volunteers, and even past employees.

Using job descriptions listed on the team pages, I distinguished between different types of team members in two ways. First, I tried to determine whether or not each person was part of an organization’s “core staff”. I excluded from the core staff designation roles such as board members, freelancers, interns, and research affiliates. If someone was listed on a team page in any capacity, I include them in the “all staff” category.

I also tried to categorize the type of work each person did, distinguishing between “Leadership” roles (e.g. CEO, COO, board members), “Professional Staff” roles (e.g. analysts and researchers), “Support Staff” roles (e.g. operations and administration), and “Associate” roles (e.g. alumni and students).

As an example of how these classifications play out, a board member would be classified as a leadership role, but would not be considered a member of an organization’s “core staff”.

Shortcomings

My methodology includes several shortcomings, including relying on subjective judgements in assessing peoples’ race, who constitutes an organization’s “core team”, and which people are in which type of roles.

To mitigate these shortcomings, I have tried to provide multiple perspectives when possible. For instance, looking only at core team members relies on my assessment of who falls into that category. By contrast, looking at everyone listed on an organization’s team page does not require that assessment, but places equal weight on people who are essential to an organization’s work and those who are somewhat tangential. Each approach has its weaknesses, but looking at both perspectives helps mitigate those weaknesses. In the same vein, I supplement the overall numbers I provide (which effectively weight organizations by their staff sizes) with the average and median figures across organizations (which give organizations equal weight).

Other weaknesses are harder to mitigate. My analysis relies on the team pages it is based upon being (reasonably) accurate. It also relies on these team pages as they stood in February 2020, as it was not practical to continuously update my sample. In some cases teams have shifted since then (e.g. I’m aware that 80,000 Hours has added a POC to their core team, and the Long-term Future Fund has replaced the sole POC on its management committee).

Staff Demographics

Of the 459 unique individuals in this sample, 65 (14%) are POC. POC make up a slightly lower percentage of total roles (13%) and core staff positions (13%).

It’s not obvious which reference point(s) should be used to contextualize these figures. The following table lists a variety of possibilities.

Sources: EA Community, UK students inc. Oxford, UK faculty, US Doctoral Universities (overall and high research), All US Doctorates and APA members, Stanford undergrad, Stanford, Harvard, US Faculty, Philosophy degrees, CS Degrees, Harvard CS, Stanford CS, Economics, Google, Facebook.

Longtermist organizations appear less diverse than even the most homogenous of these benchmarks: UK university faculty (15%), the membership of the American Philosophical Association (17%), Oxford’s undergraduate population (18%), and the EA community (19.5%).

It’s plausible that Longtermist diversity is very roughly on par with diversity at top UK philosophy departments (while I could not find demographic data for the latter, it seems reasonable to expect that philosophy programs are less diverse than university populations as a whole in the UK, as is the case in the US). But while UK philosophers are certainly well represented at Longtermist organizations, it seems a stretch to argue that they’re the appropriate benchmark for the ecosystem as a whole. Roughly a quarter of the organizations in my sample have an explicit focus on AI and are based in the Bay Area; for these organizations one might reasonably expect half the staff or more to be POC. Moreover, there’s significant variation in job function both within and across organizations in my sample, which argues for considering a wider range of disciplines as a reference point.

The significant gap in racial diversity between the US and the UK (which collectively are the home of all the organizations in my sample) admittedly complicates the benchmarking process; later, I take a closer look at the relationship between geography and diversity.

Demographic differences across organizations

Racial diversity varies significantly across the 21 Longtermist organizations in this sample.

At about half the organizations, core staff is entirely white. At some of the most diverse organizations, a quarter of core staff are POC. The Open Phil AI Fellowship is a huge outlier, with 60% of fellows being POC (twice as much as any other organization.)

There appears to be a strong pattern whereby university-based research centers/programs have much more diverse teams than other types of Longtermist organizations (which are generally standalone registered charities). At “University” organizations, POC make up 20% of core staff positions, compared to only 3% at “Non-University” organizations.

This pattern is quite consistent across organizations. None of the 14 “Non-University” organizations has more than one POC on its core staff, and 10 (71%) have none. By contrast, each of the 7 University organizations has multiple POC on their core staff, and POC make up 24% of the core staff at the average organization.

Demographic differences by role

Overall, POC account for ~13% of all roles at the Longtermist organizations in my sample. Certain types of roles, however, have much different levels of diversity. There appears to be a strong pattern in which the more influential a role is within an organization, the less likely it is to be held by a POC.

Only 4% of leadership roles are held by POC. Diversity among professional staff (14%) is in line with overall levels. But for “Support staff” roles (operations, administration, freelancers, part-timers, interns, and volunteers), POC account for 25% of positions.

Interestingly, POC make up 17% of staff alumni. Since this rate is higher than the representation of POC across all roles (13%) and much higher than the representation at organizations excluding the Open Phil AI Fellowship (10%), this raises the questions of whether POC are more likely to leave Longtermist organizations than their white counterparts, and if so, why.

Note: Within the “Professional Staff” category, differences in diversity between “Staff” and “Researcher/Fellow” roles are in large part a function of the Open Philanthropy AI Fellowship’s high percentage of POC. Excluding that program, POC make up 11% of Researcher/Fellow roles.

Demographic differences by geography

As discussed above, the difference in diversity between US and UK hiring pools complicates the assessment of diversity in the Longtermist ecosystem. And since these two countries house a different mix of University-based and Non-University-based organizations in my sample, it is important to tease apart this confounding effect. The following table does just that (looking only at Core Staff since non-Core Staff numbers include many research affiliates who are not located where the organizations are).

As expected, US-based organizations have a higher percentage of POC than UK-based organizations. A large part of this is because of the outsized influence of the AI Fellowship program. Excluding those fellows, UK organizations are actually more diverse than US organizations overall, but only because UK organizations are more likely to be based at universities. Making an apples-to-apples comparison (e.g. university-based organizations in the US vs. UK), we again see US organizations are more diverse as expected given their respective hiring pools.

There are just two University-based organizations based in the US, the AI Fellowship and CHAI. Together, they have a similar percentage of POC (42%) to enrollment at US Doctoral Universities with the highest research activity (40%), though that is largely driven by the high rate of diversity within the AI Fellowship. Representation at CHAI alone (25%) is 38% below the same benchmark. This shortfall is roughly similar to what we find at University-based organizations in the UK, where POC’s 15% representation is 31% below the 22% POC enrollment rate at leading UK universities.

At non-University-based organizations, POC are extremely rare (5%) in the US, but are completely absent in the UK in my sample. (As mentioned in the methodology section, I’m aware of one POC who was hired by a UK organization after my sample was collected, which would bring the aggregate rate of POC at non-university-based UK organizations to 2%.) One could argue that the lack of representation in the US is actually worse than in the UK, given the US’s more diverse hiring pool (especially when one considers that many of the US-based organizations are focused on AI and likely frequently hire from relatively diverse CS programs).

Suggestions for improving racial diversity

• Proactively seek out diverse candidates. Steps could include surveying networks to identify talented POC, encouraging POC to apply for open positions or start their own organizations, inviting POC to board roles, and advertising job openings in communities with strong POC representation (e.g. the EA Global meetup for POC).
• Survey POC about their experiences in the Longtermist community. This survey (which should permit anonymous answers) could capture valuable information on whether and why POC may feel excluded, and ideas for improving diversity.
• Look for lessons from the Open Philanthropy AI Fellowship program. This program, which explicitly sought to increase diversity, had more than twice as many POC (in percentage terms) as any other organization in my sample. There are likely aspects of its operations, including how it sources and selects candidates, that could help other organizations become more diverse. Scaling this program could also help improve the broader ecosystem’s diversity over time. That said, it’s important to recognize that the AI Fellowship program is not exceptionally diverse; it’s diversity is simply in line with that at Computer Science programs at top US schools.
• Promote formal diversity and inclusion programs. My analysis doesn’t provide a definitive explanation of why being based at a university appears to have such a sizable effect on an organization’s diversity. But it seems reasonable to assume that the formal policies universities typically have in place may play an important role. Wider adoption of such policies could improve diversity throughout the Longtermist ecosystem.
• Shift funding, on the margins, to more diverse organizations. On the margins, some donors may want to support university-based Longtermist organizations, which generally have much more racial diversity than other organizations. Similarly, donors may want to support US-based organizations rather than their less-diverse UK counterparts. Funders who value diversity should communicate this priority to the organizations they support or are considering supporting, particularly large funders who can influence organizational direction.
• Consider POC who are already working with Longtermist organizations for core staff positions. Nearly 30% of volunteers, interns, freelancers, and contractors (past or current) are POC. Adding top performers from this cohort to core staff positions would be a simple way to improve diversity.
• Draw advice from people and organizations that excel in diversity. As mentioned above, the Open Philanthropy AI Fellowship likely has lessons that can be applied throughout the Longtermist ecosystem, and the same can be said for people and organizations from outside that ecosystem. Research into best practices (e.g. here and here) can also help this effort.
• Provide mentorship opportunities for POC. WANBAM provides mentorship for women and non-binary members of the EA community. An analogous program for POC could provide important support and training, while also signaling that improving racial diversity is a priority. Mentorship that could help POC take on leadership positions could be particularly valuable. I reached out to WANBAM’s founder, Kathryn Mecrow-Flynn, who would be willing to help a mentorship program for POC get started: “I would love suggestions [for] women of color who would like to mentor with us going forward. I would additionally welcome a mentorship program for women of color and people of color who are interested in pursuing high impact career paths and I would personally support and lend resources and lessons learned to such a program. You can reach me at <[email protected]>.”

Conclusion

This analysis has found that POC are significantly underrepresented at Longtermist organizations and projects relative to the populations they typically draw from. At many organizations, and at most of the organizations that are not based at a university, core staff is entirely white. Where POC are employed, it is disproportionately in support roles and very rarely in leadership roles.

I believe this situation undermines the efficacy of Longtermist organizations, which depends on the talent they employ. Their high degree of racial homogeneity suggests these organizations have been missing out on talented employees and alternative perspectives, and may continue to struggle to attract strong POC candidates in the absence of changes (see here for related discussion).

I also believe this lack of racial diversity undermines the legitimacy of Longtermist efforts. If you were a policy maker in the Global South, how credible would you view “global priorities research” from an ecosystem that’s as white as South Dakota? If you were an AI strategist in China, and knew that Asians outnumbered whites in many top CS programs in the US, would you be skeptical about an overwhelmingly white ecosystem producing research on topics like “distributing the benefits of AI for the common good”?

To be clear, I don’t think the racial homogeneity of Longtermist organizations is in any way due to explicit racism or discrimination on the part of those organizations or their employees. Rather, as CEA’s stance on diversity and inclusion states, “Without conscious effort, groups tend to maintain a similar demographic makeup over time. Counteracting that tendency toward narrowness and homogeneity takes attention and effort.”

My analysis also found that when attention and effort is applied, it can lead to positive outcomes. The example of Open Philanthropy’s AI Fellows clearly shows that intentionally prioritizing diversity works. On a similar note, the efforts of CEA’s Community Health Team have paid off in terms of improving speaker diversity at EA Global. I hope this analysis helps motivate further attention and effort toward improving diversity, and ultimately efficacy, in the Longtermist ecosystem.