Accelerating DEI with O.R. in the Age of AI

An analytic, data-driven approach enables diversity, equity and inclusion (DEI) efforts to succeed where they otherwise might suffer from a lack of focus or implementability. By clearly defining and measuring desired outcomes alongside the design of DEI improvement initiatives, organizations may then apply operations research (O.R.) techniques to focus their efforts and accelerate progress. With the recent explosion in popularity of generative AI, such acceleration is more critical than ever to prevent biases from being locked into AI foundations.

In the computer and mathematical sciences, overall diversity is poor within the U.S. and only improving very slowly. The 2023 NSF Report on Women, Minorities, and Persons with Disabilities in STEM [1] shows the percentage of female, Hispanic and African American employees in these areas were only 28%, 8% and 8% respectively in 2021. These numbers show little change from the corresponding statistics for 2019: 27%, 8% and 8%.

One key reason for this lack of diversity is that women and minority groups disproportionately drop out of the STEM pipeline because of a shortage of inspiring role models. A 2017 Opportunity Insights study [2], for example, found that “if girls were exposed to female inventors during childhood at the same rate that boys are to male inventors, the gender gap in innovation would fall by half.” Creating inspiring role models from underrepresented groups requires providing members with opportunities to fully participate, grow and progress. That is the essence of inclusion.

A possible explanation for slow progress in many DEI programs may therefore be the focus on a simple static diversity metric rather than on more dynamic metrics measuring inclusion and equity. Diversity can be easy to calculate as a percentage of an organization at a point in time, but measuring inclusion and equity are more challenging. Respectively, these require quantifying how well an organization enables its diverse members to grow professionally and how well an organization provides resources to achieve inclusion. Essentially, if diversity is akin to position, inclusion and equity are like speed and acceleration.

Measuring Inclusion and Equity

An article co-authored with Beata Kilos, “An Analytics Approach to Diversity, Equity and Inclusion,” explores potential metrics for inclusion and equity [3]. For example, one might measure inclusion based on the rate at which a group participates relative to its size. For example, if 25% of staff at a given level are women but only 20% of projects are led by women, the inclusiveness of project leadership is .2/.25 = .8, or 80%. If this metric increases year over year to 85% because of initiatives, one might measure equity as 25%, because one-quarter of the inclusion gap to 100% (5% of 20%) was eliminated.

That article also outlines O.R. tools for improving DEI. Simulation, for instance, can measure the impact of equity initiatives on inclusion and diversity over time. Meanwhile, the theory of constraints can be used to focus resources to optimally design such initiatives for maximal impact. For example, what type of inclusion most limits diversity, and what actions can most reduce this lack of inclusion? Such actions typically fall into two categories: 1) making more opportunities available and 2) improving how these opportunities are allocated. Continuing our position-speed-acceleration metaphor, the former is like adding more lanes to the road, and the latter is like making it easier to pass rather than perpetuating the queueing order.

AI and Bias

Accelerating DEI progress is particularly important today as we risk building today’s biases into the foundations of tomorrow’s AI. AI systems based on reproducing the statistically most likely recommendations are prone to subjugate minority opinions and reinforce unfair biases against minority groups. A diverse group of data scientists is more likely to be sensitive to such unfairness and design it out or detect it. For example, suppose an AI system was asked to recommend award recipients based on a data set in which 99% of awards went to males. Without proper supervision, such an AI might incorrectly conflate maleness with award-worthiness.

Preventing such biases isn’t just a moral imperative – it also makes good business sense. The more limited and provincial the data set available to an AI, the more likely it is to come up with suboptimal responses. The benefits of diversity in ideas might be a product of age, experience, gender, race or many other factors. For example, an engineer attuned to the mountain roads of California might train an AI with data on such roads, leading to a much better-handling car than one developed by an engineer who grew up with only the straight, flat, potholed roads of Michigan.

In a rapidly changing, globalized world, DEI is more critical than ever, both to empower a broader section of society to fully reach their potential and to enable organizations to harness a full spectrum of insights and creativity. Even though progress has often been slower than desired, a data-driven approach enabled by O.R. offers the potential for acceleration. Moreover, it offers an opportunity to enshrine diversity in today’s emerging AI systems rather than enshrining the biases of the past.

References

1. https://ncses.nsf.gov/pubs/nsf23315/
2. Bell, A., Chetty, R., Jaravel, X., Petkova, N. and Van Reenen, J., 2018, “Who becomes an inventor in America? The importance of exposure to innovation,” https://tinyurl.com/4cyhn4zc.
3. Reaume, D. and Kilos, B., 2021, “An analytics approach to diversity, equity and inclusion,” OR/MS Today, October 12, https://doi.org/10.1287/orms.2021.05.25.