LinkedIn, the professional social networking site owned by Microsoft, has begun testing every change to its service to ensure minority groups are not disadvantaged.
The company is now using equality metrics to assess the impact of any new features, updates, or design modifications it makes, it said in a blog post on Tuesday.
The professional networking site, like most Internet-based services, has long used A/B testing, experiments in which a change is rolled out to a small, representative group of users, as well as a control group, to assess its impact. But LinkedIn says it is now considering inequality when reviewing these experiments.
LinkedIn is using the Atkinson Index, a well-known measure of income inequality, but instead of using it to measure wealth differences, it is applying it to metrics relevant to the company, such as how many job listings are recommended for a user, says Guillaume Saint-Jacques, a LinkedIn researcher who has helped apply the technique at the company internally.
LinkedIn wants to be sure any changes do not disadvantage users when viewed through the lens of legally protected characteristics—such as race, gender, or age. Regarding gender, for instance, the site will examine whether a change encouraged engagement more from men than from women. But Saint-Jacques says the company also takes a far more expansive view of equality.
“For a lot of protected categories, such as ethnicity or religion, we don’t have the data,” he says. The site doesn’t routinely collect information about people’s race, religion, gender, or age—although users can choose to provide that data.
Instead, he says, the company’s data scientists and product managers look at the least well-off users, according to certain LinkedIn-specific metrics, such as social capital (how many connections to other LinkedIn users does the person have) or engagement (how much does the user post, comment, or take action such as applying for jobs on the site). They try to make sure that those users who rank low on these metrics do not see the metrics fall still further as a result of any changes the company makes.
He says that often users who rank poorly according to these metrics also tend to come from underprivileged socioeconomic backgrounds. So by improving how these users experience LinkedIn, the company is also probably helping with income inequality. And because income inequality is also linked in many cases to race, perhaps with racial inequality as well.
Saint-Jacques says that too many companies, when examining the fairness of their algorithms, have focused solely on the workings of the algorithm itself or the data the algorithm ingests. In other words, they’ve assumed that if their process is fair, they are okay.
While process is important, Saint-Jacques says, companies should devote just as much effort to looking at the fairness of outcomes: What impact does an algorithm or update have on different groups of users.
“Even if you treat people the same way, that can actually not make the outcome equal,” he says.
LinkedIn has begun applying inequality tests even to site changes that seem, on the surface, as if they ought to be neutral—such as altering the background color of a page or changing whether people reset their password with a text message or an email. It turns out, Saint-Jacques says, that even these seemingly neutral changes can wind up having a disparate impact on certain subsets of users.
And the key, the company said in its blog post, is to not look simply at the average impact of a change. “Even if no average impact is detected, we found many instances where there was an inequality impact,” it said.
The company noted that it found, somewhat to its surprise, that changes that improved the site’s speed and reliability, especially for low-bandwidth users, improved many of its equality metrics. It also found that pushing notifications about new features to less-engaged users had a big impact on improving equality.
Saint-Jacques says that the product managers are not compelled to always make decisions that will increase equality—only that they must consider the impact on inequality as part of their decision-making calculus.
For instance, he says, there are cases where a certain change to the site might help “influencers,” business leaders who tend to post a lot of content to LinkedIn and generate a lot of engagement from other users. But the same change could also disadvantage users with low social capital. The company might still implement the change, Saint-Jacques says, but it will have at least discussed what the equality tradeoffs are and looked at ways of possibly mitigating the adverse impact on less-active users.
This story has been updated to clarify that LinkedIn users can opt to provide data on race, gender, and ethnicity but that the company does not use any other method to infer such information. The update also corrects an error: Users who rank low on LinkedIn’s internal metrics often come from underprivileged backgrounds, not privileged ones as the article originally stated.
More must-read tech coverage from Fortune:
—How the coronavirus stimulus package would change gig worker benefits
—Inside the global push to 3D-print masks and ventilator parts
—Apple focuses on what’s next amid coronavirus outbreak
—A startup is building computer chips using human neurons
—Listen to Leadership Next, a Fortune podcast examining the evolving role of CEOs
—WATCH: Best earbuds in 2020: Apple AirPods Pro vs. the Sony WF-1000XM3
Catch up with Data Sheet, Fortune’s daily digest on the business of tech.