Ironically, the same innovations we tend to regard as “creepy” (e.g., AI, algorithms, and Big Data) may help leaders make their workplace more inclusive. But there are reasons to be skeptical.
I don’t consider myself a techno-enthusiast, and I’m definitely not optimistic by nature. So, NO, this isn’t another overhyped post on how AI will save the world, or how Big Data (does anyone still use the term?) will make our world better by eliminating racism from society. Sadly, the only way to achieve that would be to eliminate humans, too. Indeed, if we were fully replaced by algorithms, racism would go extinct, much like if we replaced all human drivers with self-driving cars, traffic accidents would be rare – they would still happen, primarily due the unpredictability of human pedestrians.
Leaving aside these unlikely scenarios, and you can decide for yourself if they are more utopian or dystopian, it may be useful to understand a few of the more realistic ways in which technology could, if we truly wanted, help us keep racism in check, at work and beyond. After all, we have entered an age in which leaders’ willingness to reduce racism appears to have surpassed their ability to do so. So, if we want to move from condemnation to intervention, from criticism to solutions, we have to leverage every resource we have, and be open to new solutions, not least since traditional interventions have enjoyed limited success. Just look at the conclusion of a comprehensive academic reviews on the subject: “Of the hundreds of studies we examine, a small fraction speak convincingly to the questions of whether, why, and under what conditions a given type of intervention works. We conclude that the causal effects of many widespread prejudice-reduction interventions, such as workplace diversity training and media campaigns, remain unknown”.
There are at least four obvious ways in which technology, especially data-driven approaches to managing employees, could help us reduce workplace racism:
(1) Analyzing e-mail and messaging metadata (the context and social networks of communications): Without intruding in people’s privacy or reading what people say in their work-related communications, which in itself would not be illegal, algorithms can be trained to map the social networks in a team or organization, identifying individuals who are excluded from social interactions. Overlaying that data to demographic information on diversity (race or otherwise), could help organizations “model” inclusion digitally, using passive and non-invasive measures. Imagine a leader or HR professional, like a Chief Diversity Officer, who can access group-level data to assess whether race (or being part of any minority group or protected class) is statistically related to being left out, ignored, or ostracized from the team or organizational network. This granular level of evidence is likely to reveal exclusion where self-reports do not. People are not always aware of their unfair treatment of others, rationalizing their own actions to construct a benevolent self-concept, which is why the vast majority of people see themselves as “nice”, even when others don’t. And when they are, they are pretty good at disguising it, which is why the number of people who answer “yes” to the question “are you a racist?”, is far lower than the number of actual racists.
(2) Analyzing the actual content of communications (Natural Language Processing and red flags): Without getting bosses, or any human, to spend any time snooping on employees’ communications, AI could certainly be trained to reliably monitor the words people use when they interact in any digital medium. Of course, we did not need AI to deter people from misbehaving in traceable or recorded communications, and cautious employees have always found ways to keep offensive comments (including prejudiced and racist jokes) offline. But with an unprecedented level of work exchanges now happening online or in virtual environments, only AI could “keep an eye” on all the possible toxic, antisocial, or counterproductive comments. With rapid advances in Natural Language Processing, software that translates patterns of word usage into a psychological profile of the individual, including their potential level of prejudice, anger, and aggression, it is easier than ever for organizations to detect and sanction racist behavior. What happens offline tends to stay offline, but what happens online is recorded for posterity. Note the application of this technology to reducing racism could be twofold: you can check for actual offenses, which is what humans would do in the case of reported behaviors, or if they are actually spending their time reading everything people say; or you could check for potential, which means identifying signals that predispose or increase someone’s probability to misbehave in the future. The latter is ethically more questionable, but also enables prevention; whereas the former is mostly helpful for sanctioning behaviors after they happened.
(3) Mining the digital footprint of external job candidates, particularly for leadership roles (reducing selection bias): One of the best ways to reduce racism is to avoid hiring racist employees, particularly to be in charge. An inclusive culture is best harnessed top-down, with teams and organizations that are led by ethical, open-minded, altruistic, and compassionate individuals who show uncompromising commitment to equality and fairness, practice what they preach, and put their money where their mouth is. Throughout most of our human history, we lived in small groups where everybody knew each other well, and our models for understanding and predicting others were bullet-proof: if you systematically misbehaved, you just ended up with a terrible reputation. Fast forward thousands of years, to the typical demands of modern work, where we are forced to make high-stakes decisions about hiring, promoting, trusting, and following others who we barely know. Such are the complexities and ambiguities of work today that we have no way to know whether the person we have in front of us is truly the person we think we see. We do know that “you cannot judge a book by its cover”, so the only way to make seemingly logical evaluations of others with the rather limited information we have on them – for example, during a one-off job interview, is to rely heavily on our intuition, which is how we end up making prejudiced and racist decisions in the first place, even when we try to avoid it or persuade ourselves that that isn’t the case.
Imagine, instead, if we could access a candidates’ entire online footprint, consisting of everything they have done online in the past. The process would not be manual, of course, but algorithms could be trained to translate people’s digital history into a quantitative estimate of their open-mindedness, tolerance, authoritarianism, and empathy, etc. In some instances, you wouldn’t even need algorithms to detect whether someone “could” have a prejudiced profile, because their behaviors would just signal racism; as when human recruiters run a google search on a potential CEO candidate to gauge their reputation and assess “fit”. They may not be explicitly looking at indicators of prejudice, but may still want to exclude candidates, or at least one would hope, who don’t seem to have a strong reputation for integrity or ethics. As one would expect, there is a booming business for online reputation management, but these digital spin doctors are focused on helping you fool human assessors rather than machine-learning algorithms. And while prejudiced individuals may always find a way to fool both humans and computers, academic research suggests that if a well-trained AI were to access our complete digital footprint – everything from our Uber ratings to our Netflix and Hulu choices, our Facebook Likes, and of course our Twitter and Whatsapp exchanges – it would be able to predict with great accuracy whether we are likely to display any kind of prejudice or discrimination at work and beyond, and with what frequency. This could and should be deployed in an ethical way, asking candidates to “opt in” and put their data to the test. In fact, it may even be useful developmental feedback for them to find out whether they resemble more or less prejudiced individuals, which the algorithm could report.
(4) Exposing bias in performance ratings (eliminating the politics and nepotism in promotions and performance management systems): The most pervasive form of bias and discrimination people suffer at work, and one of the hardest to detect, is being unfairly evaluated and rated for their performance, whether consciously or not. This bias occurs even in well-meaning organizations, including ethical companies with mature diversity and inclusion policies, and meritocratic talent management intentions. Interestingly, this is an area where AI has attracted a great deal of popular criticism. For instance, when companies attempt to train AI to predict whether an employee is likely to be promoted, or algorithms are used to rank order internal candidates for potential, the likely result is that certain profiles, such as middle-aged White male engineers, over-index, while others, such as Black, Latino, or female, are underrepresented. However, in these examples, the problem is neither the algorithms nor the data scientists who develop them. If you train algorithms to predict an outcome that is itself influenced by systemic bias or prejudice, it will not just reproduce human bias, but also augment it. In other words, middle-aged White male engineers got promoted before AI was trained to reveal this, and they will continue to get promoted even if AI is not used.
Unlike humans, computers don’t really care about your race, gender, or religion: they don’t have a fragile self-esteem they need to boost by bringing other people down, and they don’t need to bond with other computers by stigmatizing certain classes or groups of humans (or computers). One thing they can do, however, and usually rather well, is to imitate the prejudiced preferences of humans. And yet, they can also be trained to not imitate them. Artificial intelligence may never match the breadth and scope of human intelligence, but it can certainly avoid replicating the vast collection of biases encapsulated in human stupidity.
If organizations are able to measure employees’ job performance objectively, then AI will not only predict it better than humans, it will also identify any distortion or interference introduced by bias. For instance, an Uber driver can be judged on the basis of (a) how many trips she makes, (b) how much money she brings in, (c) how many car accidents she has, and (d) what customer ratings she gets, all relative to other drivers working in the same location, and all done through AI. If two drivers with identical performance scores on (a), (b), and (c) differed significantly on (d), then a simple analysis would suffice to reveal whether the driver’s demographic background – e.g., white vs. black – inflated or depressed their scores on (d). So, if businesses, and particularly managers, were able to quantify output (what an employee actually contributes to the team and organization), and there was a formula to predict what someone is likely to produce, then algorithms won’t just be better than humans at applying this formula, they will also be better than humans at ignoring the wide range of extraneous variables (including race) that distract human managers from focusing on that formula. In judgments of talent, humans are not great at paying attention to the stuff that matters, or ignoring the stuff that doesn’t.
OK, so if all this sounds that simple, then what’s stopping organizations from adopting these measures? After all, technology keeps getting cheaper and cheaper, they are sitting on a growing volume of data (just think about the last 3-months of videoconferencing data), and there’s no shortage of data scientists who can help with this.
Three things, which do warrant a reasonable degree of skepticism, or at least a non-trivial amount of realistic pessimism.
The first is the double standards that make most people adopt the highest moral principles when they judge AI (or any new tech), while being much more lenient, and morally relaxed, when they judge humans. No technology will ever be perfect in predicting or detecting any form of human behavior, whether good or evil. That’s not the point. The point is that technology could be more accurate – and less biased – than humans, or at least that it could help humans be less biased. So, if we are comfortable understanding that even the latest, most advanced, tech will get things wrong, then let’s focus on what really matters, which is whether that tech can at least minimize human bias and discrimination, even by 1%. Given humanity’s historical record here, it is safe to say that the bar is pretty low, and that even rudimentary technologies may represent a hopeful improvement over the status-quo: pervasive prejudice and ubiquitous bias, courtesy of the human mind.
The second is that diagnosing or detecting the problem is necessary, but not sufficient, to fix it. So, suppose that some of the emerging technologies mentioned here reveal racism, or any other form of discrimination in a team or organization; what next then? Will leaders genuinely act on these data to sanction it, especially if it creates conflict or leads to negative short-term outcomes, particularly for themselves? Will they change the rules, the system, and those who oversee it, in order to improve progress and create a fairer, more meritocratic system? In general, people have little incentive to change the status quo when they are part of it. How comfortable will they be acknowledging that the status quo was rigged? Clearly, many leaders may prefer to avoid any new technology if it has the potential to expose the hitherto invisible toxic forces that govern the power dynamics in their teams and organizations. And AI, not just in HR but in any other area of application, is like a powerful X-ray machine that can reveal unwanted forces and socially undesirable phenomena: like opening a can of worms. It is always painful to make the implicit explicit, and the foundational grammar of any culture – both in companies and societies – is largely made of silent and subliminal rules, which makes them resistant to change. One of the disadvantages of living in a liberal world is that domination is rarely explicitly asserted, but hidden under the pretext of equality.
The third is that employers must be realistic about what can probably be achieved. Discussions on race and inclusion often focus on bias and prejudice, but it may be unrealistic to change the way people think, especially if they are part of a non-conformist minority, and how ethical would this be anyway? Crucially, people’s attitudes or beliefs are surprisingly weak as predictors of actual behavior. This means that discrimination – the behavioral side of prejudice – often happens in the absence of strong corresponding beliefs (conscious or not). Likewise, most of the people who are prejudiced – and hold derogatory attitudes towards individuals because of their race, gender, etc. – will rarely engage in overt discriminatory behavior. In short, we should focus less on changing people’s views, and more on ensuring that they behave in respectful and civil ways. Containing the expressed and public manifestations of racism at scale may be the best leaders can hope for, for they will not be able to change the way people think and feel.
Finally, we should not forget that not all leaders will be interested in reducing racism in their organizations, particularly if the process is slow and painful, and the ROI isn’t clear. Such leaders may not just neglect the potential value of new technologies for reducing workplace discrimination, but use it to perpetuate existing practices, amplifying bias and prejudice. In the long run, every decision leaders make will shape the culture of their organizations, and employees and customers will gravitate towards those cultures and brands that best represent their own values and principles.