Results

Results

Attrition

Domain: Attrition

Research Question 1: Why do women/minoritized folks leave AI teams?

The predominant themes related to attrition that emerged from participant responses were:

1. A Toxic Work Environment

Participant Identifier

This report denotes each participant with a unique identifier consisting of a letter and a number. The letter corresponds with the type of participant.

  • F: General participants
  • M: Managers
  • D: Those working in DEI roles

Several factors worked together to create a toxic environment on the AI teams of participants that we interviewed. Some of these factors were more generalizable to other tech or non-tech workplaces, while some uniquely fit the AI ecosystem. For instance, several participants described not getting proper credit for their work and others described a negative sense of competition within teams. One participant said, “It felt like a war zone. Like you weren’t working together to build a product. You were trying to tear other people down to get yourself ahead, which is not, not my mantra.” (F1) Some participants highlighted the highly competitive nature of AI teams specifically. According to one, “so I hear people aren’t exactly happy on that team and it works in AI because it’s a product team, but [for a] conversational agent. So it’s high pressure. The stakes are high because they have to compete against you know, [AI product name] and other teams like that. But I know that multiple people on that team have taken on medical leave for work, work-related stress and burnout and well.” (F8) Workers stressed that managers heavily influenced whether they stayed or left due to these positive or toxic environments.

Managers heavily influenced whether workers stayed or left due to positive or toxic environments

Participants also described managers’ approach to building and leading AI teams having a strong impact on whether they stayed or left. Some were on teams that fostered a sense of interdisciplinary collaboration where people from both technical and non-technical backgrounds worked together to build products, explore the implications of AI, and use AI to tackle social problems such as inequality.

Participant F21 said:

“It was really just like one of the big reasons why I left. It was like people making, not allowing me to be fluid and very much being like, ‘Hey, you need to pick a role.’ And that was mostly because of organizational structure. [Because] I was no longer in an interdisciplinary organization.”

Conversely, participant F19 described reasons for staying:

“What makes me want to stay longer? I think I really resonate with the mission the team is working on. I think I have a good manager and I have a great team. I think it’s, it’s the dynamic. Although, like, there, there are things that can be improved, in terms of team culture and team dynamics. But… I think this is probably above the industry average. I wouldn’t leave at this point and then there are things I want to do. There are researchers that are really, really interesting, really fascinating. They don’t want to continue on, [but] I do, I do think what I’m doing is going to have a positive impact on people and people I care about. “

Environments in which diverse participants thrived tended to encourage exploration of an interdisciplinary approach to AI that incorporated social or sociotechnical issues.

2. Experiences of Prejudice

As in other work environments, AI teams were not immune to the effects of prejudice. Participants described microaggressions, overt interpersonal examples of prejudice, and more systemic ways in which their teams and organizations discriminated against them due to their identities.

As one participant who worked within a DEI role stated:

“I think a lot of people burn out and leave as a result, I think that’s one layer. I think the other layer is just, even if you’re not involved in DEI work, being in a product area or even at a company that isn’t particularly diverse means that you’re surrounded by people who in a lot of ways, just aren’t like you, they haven’t had similar life experiences. They don’t know what it means to be you, and they often have a lot of ideas about who you are. And so you’re constantly trying to prove yourself that you’re not, like, a diversity hire that you’re actually meant to be in the room and et cetera.” (D2)

One female participant of color, F17, described a case in which:

”It took a fight to have my manager [to] be okay with letting go of a partner because it literally took me and my female coworker for him to sit in on three different meetings and realize how sexist he was. And it was, you know, comments… like, ‘You guys are women. You don’t understand how the space works.’ Or like, ‘Women are better at planning dinners than men, like, you guys should just worry about the menu and like the food.’“

3. A Need for Growth

Many participants expressed a need for continual growth within the field of AI and left teams or organizations that did not provide pathways for this or those that hindered their professional growth. Managers and AI teams that encouraged this development tended to encourage workers to stay. According to F28:

“So I would say the managers and leaders in that previous [REDACTED] company, they were very much about bottom line and customers and delivering good products. They did not pay too much attention to the culture of developing people or retaining them.”

Key Takeaway From This Domain:

When taken together, the data suggest that these participants belonging to minoritized identities either left or intended to leave organizations that did not support their continual career growth or had values that did not align with their own.

Participants belonging to minoritized identities either left or intended to leave organizations that did not support their continual career growth or had values that did not align with their own

Consistent with this, participants described examples of their organizations not valuing the content of their work. Participants also tied their desire to leave with instances of prejudice or discrimination, which may also be related to a toxic work environment. Some participants reported examples of being tokenized or being subject to negative stereotypes about their identity groups.

Culture

Domain: Culture

Research Question 2: What are the characteristics of teams where folks do/don’t thrive?

The predominant themes related to culture that emerged from participant responses were:

1. Collaborative Work Environments

Many participants described their current work environments as generally collaborative and collegial. They generally characterized their teammates as willing to help each other. Typically, participants said their previous teams started out this way but eventually started to exhibit toxic aspects to their culture that made them leave. The culture of the teams seemed to depend on the tone set by the manager. According to participants, the managers heavily influenced whether AI teams were inclusive or toxic. F13 described their own team thusly:

“It’s a pretty personable team. The word ‘family-oriented’ is coming to mind, not so much that my team’s a family, but everyone on the team has families and they’re pretty accommodating of that, which I feel has been not always the case in positions or teams I’ve had, so that stands out… I think my manager has influenced the climate on my team by overdoing empathy and making sure that from the first day that I met her, she tried to get to know me as a person and not an employee and kind of asked, ‘What do you want your work-life balance to look like? What is important to you to work on? Like why do you want to work on this team?’ From then she sometimes will check in, and just ask if she noticed [something unusual] like in a meeting. I’m thinking of a recent time, I kinda got spoken over in a meeting and my manager immediately messaged me after. And she was like, ‘Hey, like, you know, that was a great point you were making and, I heard you.’ So she’s, she’s very proactive, I think with her empathy.”

Other participants emphasized that the collaborative and positive environment felt restricted to their own “bubble” that had been carefully curated by their managers. For instance, F16 said that:

“I think people see our group as very close together and cohesive because we work on projects together and we communicate all day, and we’re quite separate to the rest of the organization, like just the [REDACTED]. So yeah, it kind of feels like we’re all sometimes in our own little bubble.”

2. Experiences of Prejudice

Several participants described ways in which their AI teams exhibited prejudice, especially female participants discussing specific examples of sexism or ways in which others were allowed to undermine their technical expertise as women.

F9 relayed an experience that several other participants in AI mentioned, specifically:

“There have been times where people, and this is actually not just like, you know, like non-diverse crowd, like a broad spectrum will be like, ‘Oh, are you technical?’ And then I think there has been times when I’ve had to say — and not so much like gender or ethnicity or race or like sexual orientation — but more from like diversity of talent that contributes to the design of AI, which is an area passion of mine where people say like, ‘Oh, so you’re not technical’ or like, ‘Oh, so you’re just the program manager’ in the AI world, and it’s like males and females will say that. And I think there’s definitely this kind of belief that this is a world only for technical people.”

3. Diverse and Inclusive Teams (or Lack Thereof)

Many participants described the importance of interdisciplinary and identity diversity for fostering a positive team environment. For instance, F14 said:

“So I think my manager is very proactive and sort of leads by example. In larger group meetings or meetings with other folks she’ll point out particular social concerns or point out, you know, who’s not speaking and sort of gently prodding for certain folks to chime in and, or really gracefully getting other people to stop talking. I think also in one-on-one meetings, she’s very upfront about saying, ‘Hey, if this thing is an issue or if anything is an issue, let’s talk about it.’ And for me personally, it helps that she’s a person of color.”

Beyond diverse managers, participants emphasized the importance of managers intentionally creating environments where minoritized workers felt empowered to contribute to their teams using their technical and interdisciplinary skills. As F1 said:

“My old team people would try to steal credit from each other all the time. I had a couple of big things technically stolen from me and someone else got the credit for it. On this team, the managers and the people make sure in every presentation, like, ‘Hey, X, Y, and Z, this person contributed, they did amazing. Here’s all this great work we did because we all did it together.’”

One manager described taking time to intentionally create a supportive and inclusive team culture where people with diverse backgrounds could thrive, drawing from diverse perspectives and interdisciplinary expertise to cultivate these qualities. The value of this approach to work was echoed by several comments from participants who found positive AI teams to work on.
Key Takeaway From This Domain:

Participants consistently expressed that managers who belonged to minoritized identities or who took the time to learn about working with diverse identities were more supportive of their needs and career goals. This filtered into teams that were also more diverse and inclusive, engendering a positive team culture/climate where folks belonging to minoritized identities thrived. In contrast, there was a recurrent theme of work environments that were sexist or ones where participants experienced acts of prejudice such as microaggressions. This could manifest as decision-making from leaders who did not take folks who were not White into account, or times where credit was misattributed to those in power, instead of female or minoritized folks who did large amounts of the work. While collaborative or positive work environments were also a common theme, these did not negate predominant cultures which deprioritized “DEI-focused” work or work that did not serve the dominant group.

Efforts to Improve Inclusivity

Domain: Efforts to Improve Inclusivity

Research Question 3: What are things that can be done to make organizations/teams more inclusive?

The predominant themes related to efforts to improve inclusivity that emerged from participant responses were:

1. DEI Trainings

Participants described a broad range of DEI trainings but few spoke confidently about their effectiveness. In general, trainings that they described as specific received more praise than those that spoke to very general “DEI” topics such as race, culture, or implicit bias.

2. Employee Resource Groups (ERGs)

Many participants mentioned Employee Resource Groups as one aspect of their organizations that effectively promoted inclusivity. These ERGs seemed to serve two purposes as described by the participants. They provided an opportunity for workers from minoritized identities to find fellowship and build community, and they provided a space for workers to find others who practiced interdisciplinary work that incorporated AI with tackling social issues that affected diverse communities.

One participant said that they benefitted daily from ERGs, which acted as valuable spaces for camaraderie, knowledge-sharing, and guidance that foregrounded minoritized identities inside larger organizations that did not. Given their value, the participant thought ERGs should have greater institutional support.

3. Bias Against DEI efforts

There still seemed to be a stigma attached to top-down DEI efforts at all levels. Participants described efforts around diversity as not being taken as seriously as more technical topics within AI teams. F8 said:

“I think that’s just a general, like I’ve noticed that [REDACTED] pretends to have certain values about inclusion and diversity, but at the end of the day, the types of behaviors that get rewarded tend to be the behaviors of the dominant culture.”

On a systemic level, D5 said:

“There are a ton of factors. Certainly there is on the most sort of individual-level interpersonal interactions that are laced with micro- or macro-aggressions. And then all way up to the highest most institutional systemic level. The tech industry is one that has a culture of white supremacy and not in the sense of explicit racism and interpersonal hostility, but one where the communication styles, decision-making styles, resource allocation decisions are all fueled by the predominantly white and Asian male people who have led in the industry for decades, and is not accommodating of a lot of other perspectives, worldviews approaches, needs identities, et cetera. And so there’s a perpetual conflict intention that an individual can feel operating in these spaces. And then, you know, certainly more acute instances of implicit biases leading to impacting hiring decisions or promotion decisions who gets to work on what project, how teams are organized and supported, what work is and isn’t valid, and that may or may not align with the goals, priorities and experiences of people from marginalized identities. And so, yeah, that perpetual tension is one that prevails.”

4. Manager Supports

Ultimately, managers and leadership played a huge role in making their teams more inclusive. Participants tended to attribute much of the success of their teams in creating a diverse and inclusive climate to the intentional actions of their managers. In contrast, participants that complained of toxic environments that failed to foster supports for diverse workers tended to attribute this to the lack of concern from their managers. The managers in the study said that it was at times challenging to maintain a diverse and inclusive culture when the leadership at higher levels was not supportive. Several managers interviewed also referred to the importance of mentoring their reports, especially those who belonged to minoritized identities in the AI field. They reported it was professionally and personally important to them to continue to build an AI community that draws from a variety of disciplines and encourages minoritized workers (both with technical skills and those without) to find their place in this growing field.

5. DEI Hiring Practices

Participants typically said that intentional DEI initiatives around hiring were effective. Although they acknowledged that the pipeline into their companies was a separate issue, they mentioned it as being relevant to increasing the inclusive and diverse culture of AI teams in general. F22 said:

“It’s very unfair in the sense that you can work with people, for example, the data science space, but AI space is quite fully occupied by people with masters and PhDs, but you work in corporates whereby you, you get a 25-year-old, 23-year-old woman, who’s got no experience, no degree, no masters. And only because they’re White and they’re well-connected so it’s still an issue. And I must say it’s systematic, it’s, it’s more systematic than the recruitment, I think the recruitment. I mean, there’s a way, or we can use artificial intelligence and machine learning to, deal with these issues.”

Key Takeaway From This Domain:

Although many participants reported having to do DEI trainings, these varied widely in terms of how effective they were perceived, with some participants discussing them as having little to no effect and others describing real and observable changes due to these trainings. For instance, several participants discussed DEI trainings that were very specific to some groups (e.g., gender diverse folks, Black people) being the most effective. Participants who mentioned ERGs uniformly praised them, discussing the huge positive impact they have had on them, forming the basis for their social support networks in their organizations. Although this study did not focus on the pipeline, DEI hiring practices, such as recruiting at Historically Black Colleges and Universities, and intentionally diversifying the interview pool of applications, were recurrent themes, potentially related to increasing the diversity within organizations and thus improving climates in general for minoritized folks. Equally recurrent was bias against DEI efforts, with participants discussing that other workers in their organizations would dismiss or discredit these efforts as frivolous or unimportant. Managers also had a huge role to play in the execution of these DEI efforts. Minoritized folks often looked to managers as mentors. The predominant types of mentorship that emerged were:

  1. Formal training from mentors,
  2. Informal mentorship, often involving minoritized folks looking out for each other, and
  3. Mentors supporting more junior workers who belonged to similar identities.

Key Takeaways From Across All Domains:

  • AI teams which fostered an interdisciplinary and diverse environment, supported by managers and senior leadership, tended to have positive and healthy cultures for workers. These diverse workers — who had both technical and non-technical roles — went into the AI field for a variety of reasons, but often expressed a desire to use their interdisciplinary backgrounds to contribute to social good, especially for minoritized communities, and to grow within their own careers, building products and working on research that advanced innovative and ethical AI methods.
  • Other AI teams failed to build this type of environment, instead allowing cultures where:
    1. Microaggressions, sexism, racism, and other prejudice went unchecked,
    2. Teams in which women with technical skills in AI were undermined due to gender, and/or
    3. There was little regard for diversity or interdisciplinary backgrounds, which tended to create the perfect storm for minoritized workers to want to leave the moment their compensation no longer became worth it.
  • The key to build more inclusive spaces seemed to come from diverse leadership that valued a diverse and interdisciplinary AI team

    While organizations and teams have tried numerous efforts to make workplaces more inclusive, such as diversity trainings, the key to build more inclusive spaces seemed to come from diverse leadership that valued a diverse and interdisciplinary AI team.Other efforts such as diversity statements or ERGs seemed to be valuable to minoritized workers, but these alone were not enough to undo systems which did not value them or leadership that did not look like them or understand their perspectives.