Applying for jobs is stressful for most people. And it does not help when you know that “people like you” are likely to have a gigantic minus placed on their application—increasingly, by a machine.

The automated screening process is frequently met with applicant suspicion (1, 2). And in many cases, for good reason—the process is indeed fraught with bias. As just one example, older applicants are now given advice on outsmarting ageist bias, which appears to be built into many automated applicant screening systems and AI-saturated hiring models, by removing or condensing information on earlier experience and education (2). Sadly, even if removing older positions will help one person get a job, it does nothing to address ageism. If anything, it further embeds the connection between “too much experience” and “unhireable” into the system—and the cycle of ageism continues.

The cycle of biased practices begetting more biased practices can be clearly illustrated by analyzing neuroexclusion at work.

Neurodivergence, measurement validity, and the irony of human bias

Many automated screening systems reproduce biases that exist in manual resume screening. However, expanded versions of automated screening may include additional barriers traceable to the lack of validity—the correspondence between measurement and what is supposedly being measured. One of the validity challenges here is that systems with built-in assessments, such as personality testing, are rarely validated for neurodivergent candidates. Another issue is that these assessments may also not sufficiently correspond to the specific position and what is required for performance (3).

Of course, negative assumptions about neurodivergent people were prevalent long before the widespread use of automated screening. Yet, the world of human bias also presented some workarounds that, ironically, could allow for a more valid assessment of actual ability than “official” systems. Consider Temple Grandin, the famous autistic author and professor. For much of her early career, she used portfolios of her work to impress hiring managers and sidestep the likely discrimination based on her autistic differences (4). But at that time, Temple Grandin was not dealing with automated screening systems and could show her drawings and photos of completed projects to hiring managers. Would she have made it through and been able to show her portfolio in today’s world? Or would she have been automatically screened out?

When bias is located in a human, it can be corrected locally. When bias is located in the black box of algorithmic screening provided by monopoly vendors to a large number of employers, the potential to harm rises exponentially.

Data-based is not the same as fair

Most automated systems are trained on data that historically predicted employee success. But data-based is not the same as fair: data quality and the way it was collected matter. Because of the long-term exclusion, current applicant screening systems are not trained on neurodivergent employee data. They are also not trained on data that would apply to neuroinclusive environments. The impact of long-standing exclusion on data used to predict success is a fairness concern that may take a long time to address.

Again, the role of bad data in exclusion is best understood as a cycle. Organizational practices that keep neurodivergent people out of work may contribute to a lack of information on whether existing screening and selection procedures are valid and can accurately predict their success in the workplace. In turn, the lack of screening approaches validated for neurodivergent populations may perpetuate unemployment. The cycle closes.

Here’s an example of what happens when we use typically available data to train automated systems. Imagine that we are predicting the success of individuals in doing cooking tasks. We give them all an opportunity to cook in a kitchen and measure their speed and accuracy in the food prep. Sounds fair, right? On further examination, let’s say the kitchen is built for people who are around 5½ feet (168 cm) tall with no disabilities. We use that same kitchen to test the success of everyone: average-height people, those who are 4 feet tall, those who are 7 feet tall, and wheelchair users. Although the people who do not “fit” the kitchen show incredible ingenuity to accomplish their task, their success requires extra time and effort. Very few pass the test and are hired. Even fewer become master chefs. Over time, we have enough data to justify that being 5½ feet tall is a valid predictor of an aspiring chef’s success. Which it is, but only because the kitchen is built for these people.

Should we continue hiring 5½-foot-tall chefs, allowing the automated screening to exclude applicants who don’t meet the height requirement because it is technically “predictive”? Or should we make kitchens more flexible and accessible to avoid automatically screening out a wider pool of applicants?

By the same logic, what makes more sense: Should organizations continue to exclude neurodivergent people based on the data collected in exclusionary environments? Or should we make organizations neuroinclusive and collect better data? Algorithmic selection won’t make hiring more just unless human justice comes first.

If we don’t break the cycle of using prior unfairness to excuse present and future unfairness, our data and technology serve to make entrenched biases even more entrenched. That hardly qualifies as progress. But designing systems for human dignity that integrate the best of appropriately used technology and the best of human judgment, has the potential to support both progress and fairness.

This post is an edited excerpt from my book, The Canary Code: A Guide to Neurodiversity, Dignity, and Intersectional Belonging at Work. A shorter version also appeared in the Best Work for Your Brain newsletter.