AI and Diversity

I was sitting in a conference room last year when someone suggested we "just let the AI handle" our diversity recruiting. The comment landed with a thud. Around the table, a few people nodded—relieved, perhaps, that technology might solve what felt like an intractable problem. But I caught the eye of our head of HR, and we shared the same unspoken concern: What exactly would that AI be handling? And whose version of diversity would it encode?

That moment has stayed with me, especially after my conversation with Karoline Thompson, assistant professor at the Darla Moore School of Business. Her research, published in journals like the Journal of Applied Psychology and Leadership Quarterly, examines the messy reality of diversity work—how overlapping identities shape workplace experiences in ways that simple demographic checkboxes never capture. What struck me most wasn't her skepticism about AI in diversity initiatives, but her precision about where the real risks lie. We're not just deploying neutral tools. We're automating assumptions about what inclusion means, often before we've interrogated those assumptions ourselves.

The central challenge Karoline identified is one I've seen play out repeatedly: AI systems trained on historical data will, by definition, learn from historical inequities. If your company's "successful" candidates over the past decade were predominantly from certain backgrounds, your algorithm will optimize for more of the same. It's not malice—it's math. But the outcome is a kind of sophisticated gatekeeping, wrapped in the language of objectivity. Karoline's work on intersectionality makes this even more urgent. A Black woman doesn't experience the workplace as "Black" plus "woman" in some additive equation. Her reality is shaped by the interaction of those identities in ways that are contextual, relational, and often invisible to systems designed around single-axis thinking. When we ask AI to identify "diverse" candidates, we're often asking it to spot easily categorizable differences while missing the complex, lived experiences that actually shape how people navigate institutions.

This matters because the promise of AI in diversity work is seductive: scale, consistency, the elimination of conscious bias. And there are real applications. Natural language processing can flag gendered language in job descriptions. Structured interview platforms can ensure every candidate gets asked the same questions in the same order. These interventions address specific, measurable problems. But Karoline's research points to a deeper mechanism we need to understand. Bias isn't just individual prejudice that can be engineered away. It's embedded in systems—in who gets mentored, whose ideas get credited, which leadership styles are rewarded. AI tools can replicate those systems with devastating efficiency if we treat them as solutions rather than instruments that require constant human judgment.

I think about the organizations I advise, many of whom are experimenting with AI-driven talent management platforms. The question I've started asking isn't whether the algorithm is biased—all models reflect the data they're trained on—but whether the humans deploying it understand what they're optimizing for. Are we seeking demographic representation? Cognitive diversity? Equitable outcomes? Those are related but distinct goals, and conflating them is where AI implementations often go wrong. Karoline emphasized something that should be obvious but frequently isn't: technology amplifies our intentions. If our intention is to check a box, AI will help us check it more efficiently. If our intention is to fundamentally rethink who gets opportunities and why, then AI becomes one tool among many—but only if we're doing the hard cultural work alongside it.

The practical implication for leaders is uncomfortable but necessary. Before deploying any AI system in hiring, promotion, or performance evaluation, we need to audit what "success" has historically looked like in our organizations and ask whether that's actually the pattern we want to perpetuate. This requires transparency about outcomes—not just who gets hired, but who advances, who stays, whose ideas shape strategy. Karoline's work suggests that intersectional analysis can reveal patterns that aggregate diversity metrics obscure. For instance, you might hit your targets for both racial and gender diversity while still systematically undervaluing people who hold both identities. The algorithm won't catch that unless you've specifically designed it to look, and even then, human interpretation is essential.

What I've come to believe, informed by conversations like this one, is that AI's role in diversity work should be diagnostic and augmentative, not prescriptive. Use it to surface patterns you might miss—like the fact that certain managers' teams have lower retention rates for particular demographic groups, or that your promotion timelines vary significantly by identity. Use it to remove friction from parts of the process where bias is well-documented, like resume screening for name-based discrimination. But don't outsource judgment. Don't let the existence of a dashboard convince you that you've solved a problem that is fundamentally about power, belonging, and organizational culture.

Karoline's research reminds us that diversity isn't a technical challenge to be optimized. It's a continuous negotiation of who gets to define excellence, whose contributions are visible, and how we create conditions where people with different lived experiences can actually thrive. AI can make certain aspects of that work more rigorous. It can hold us accountable to goals we've set. But it can also calcify outdated norms under the guise of objectivity, especially when we're seduced by the illusion that data is neutral.

The companies getting this right aren't the ones with the most sophisticated algorithms—they're the ones using technology to ask better questions, not to automate answers. They're the ones where executives understand that every dataset has a point of view, and their job is to interrogate that view rather than defer to it.

If we're serious about inclusion, we need to be more skeptical of solutions that promise to make the hard work easy—because the work isn't supposed to be easy, and that difficulty is where the actual transformation happens.# AI and Diversity

I was sitting in a conference room last year when someone suggested we "just let the AI handle" our diversity recruiting. The comment landed with a thud. Around the table, a few people nodded—relieved, perhaps, that technology might solve what felt like an intractable problem. But I caught the eye of our head of HR, and we shared the same unspoken concern: What exactly would that AI be handling? And whose version of diversity would it encode?

That moment has stayed with me, especially after my conversation with Karoline Thompson, assistant professor at the Darla Moore School of Business. Her research, published in journals like the Journal of Applied Psychology and Leadership Quarterly, examines the messy reality of diversity work—how overlapping identities shape workplace experiences in ways that simple demographic checkboxes never capture. What struck me most wasn't her skepticism about AI in diversity initiatives, but her precision about where the real risks lie. We're not just deploying neutral tools. We're automating assumptions about what inclusion means, often before we've interrogated those assumptions ourselves.

The central challenge Karoline identified is one I've seen play out repeatedly: AI systems trained on historical data will, by definition, learn from historical inequities. If your company's "successful" candidates over the past decade were predominantly from certain backgrounds, your algorithm will optimize for more of the same. It's not malice—it's math. But the outcome is a kind of sophisticated gatekeeping, wrapped in the language of objectivity. Karoline's work on intersectionality makes this even more urgent. A Black woman doesn't experience the workplace as "Black" plus "woman" in some additive equation. Her reality is shaped by the interaction of those identities in ways that are contextual, relational, and often invisible to systems designed around single-axis thinking. When we ask AI to identify "diverse" candidates, we're often asking it to spot easily categorizable differences while missing the complex, lived experiences that actually shape how people navigate institutions.

This matters because the promise of AI in diversity work is seductive: scale, consistency, the elimination of conscious bias. And there are real applications. Natural language processing can flag gendered language in job descriptions. Structured interview platforms can ensure every candidate gets asked the same questions in the same order. These interventions address specific, measurable problems. But Karoline's research points to a deeper mechanism we need to understand. Bias isn't just individual prejudice that can be engineered away. It's embedded in systems—in who gets mentored, whose ideas get credited, which leadership styles are rewarded. AI tools can replicate those systems with devastating efficiency if we treat them as solutions rather than instruments that require constant human judgment.

I think about the organizations I advise, many of whom are experimenting with AI-driven talent management platforms. The question I've started asking isn't whether the algorithm is biased—all models reflect the data they're trained on—but whether the humans deploying it understand what they're optimizing for. Are we seeking demographic representation? Cognitive diversity? Equitable outcomes? Those are related but distinct goals, and conflating them is where AI implementations often go wrong. Karoline emphasized something that should be obvious but frequently isn't: technology amplifies our intentions. If our intention is to check a box, AI will help us check it more efficiently. If our intention is to fundamentally rethink who gets opportunities and why, then AI becomes one tool among many—but only if we're doing the hard cultural work alongside it.

The practical implication for leaders is uncomfortable but necessary. Before deploying any AI system in hiring, promotion, or performance evaluation, we need to audit what "success" has historically looked like in our organizations and ask whether that's actually the pattern we want to perpetuate. This requires transparency about outcomes—not just who gets hired, but who advances, who stays, whose ideas shape strategy. Karoline's work suggests that intersectional analysis can reveal patterns that aggregate diversity metrics obscure. For instance, you might hit your targets for both racial and gender diversity while still systematically undervaluing people who hold both identities. The algorithm won't catch that unless you've specifically designed it to look, and even then, human interpretation is essential.

What I've come to believe, informed by conversations like this one, is that AI's role in diversity work should be diagnostic and augmentative, not prescriptive. Use it to surface patterns you might miss—like the fact that certain managers' teams have lower retention rates for particular demographic groups, or that your promotion timelines vary significantly by identity. Use it to remove friction from parts of the process where bias is well-documented, like resume screening for name-based discrimination. But don't outsource judgment. Don't let the existence of a dashboard convince you that you've solved a problem that is fundamentally about power, belonging, and organizational culture.

Karoline's research reminds us that diversity isn't a technical challenge to be optimized. It's a continuous negotiation of who gets to define excellence, whose contributions are visible, and how we create conditions where people with different lived experiences can actually thrive. AI can make certain aspects of that work more rigorous. It can hold us accountable to goals we've set. But it can also calcify outdated norms under the guise of objectivity, especially when we're seduced by the illusion that data is neutral.

The companies getting this right aren't the ones with the most sophisticated algorithms—they're the ones using technology to ask better questions, not to automate answers. They're the ones where executives understand that every dataset has a point of view, and their job is to interrogate that view rather than defer to it.

If we're serious about inclusion, we need to be more skeptical of solutions that promise to make the hard work easy—because the work isn't supposed to be easy, and that difficulty is where the actual transformation happens.