"Less leadership experience." — That's how an AI judged a résumé. Even though it listed a leadership award. The only difference from the better-rated version: a mention of a disability.

The key points in 30 seconds:

→ AI tools rate résumés that mention a disability systematically lower — even when they're objectively better (Glazko et al., 2024).

→ Hit hardest: autism (0 out of 10) and deafness. Neurodivergent and disabled talent falls through the cracks first.

→ The AI didn't invent these biases — it learned them from people, and scales them.

→ The EU AI Act classifies recruiting AI as high-risk.

→ The protection is "classic" personnel psychology: structured selection. AI supports it, it doesn't replace it (Boyce et al., 2026).

What's happening — and why almost no one sees it

More and more companies let AI pre-sort résumés. It saves time. And it's no longer reserved for large corporations with expensive software: anyone with ChatGPT can have a résumé "matched" against a job posting.

That's exactly where the problem lies. What looks like objective technology carries the biases from its training data — and applies them to real applications. Without anyone noticing.

What a disability bias is

Ableism is the devaluation of people with disabilities. Behind it sits a quiet assumption: "non-disabled" or "neurotypical" is the norm — disability a deficit. Anyone in that category is quickly seen as less resilient, less capable, a risk.

The same shows up in AI systems, in two forms: overt (direct stereotypes that aren't even in the résumé) and subtle (commitment gets praised and booked as a drawback at the same time, values get framed as a "deviation"). Both lower the rating.

What the study shows

A research team at the University of Washington (Glazko et al., 2024) tested this systematically. The setup is strikingly simple: the researchers took a real résumé and built a second version with four disability-related additions — a leadership award, a scholarship, a panel talk, a membership. That version was objectively better. Otherwise identical. A fair system would always rank it first.

The standard AI (GPT-4) didn't. In only 15 of 60 cases did it pick the better résumé. In three-quarters of cases it preferred the weaker version — as soon as the stronger one mentioned a disability (see the chart above).

It hit neurodivergent and sensory disabilities hardest: for autism, the better résumé came out on top 0 out of 10 times; for deafness, 1 out of 10. The paradox: even when the résumé clearly listed a leadership award, the AI wrote things like "less leadership experience." The signal "disability" overrode the facts.

The good news: a second model, trained for fairness, did much better — 37 hits instead of 15. Even simple instructions helped. But not everywhere: for depression, nothing improved. Training reduces the bias — it doesn't remove it.

And an important control finding: an extra award without a disability link was not penalized — quite the opposite. So it's not "more text" that hurts. It's the signal "disability."

A note on accuracy: ADHD was not included in this study. The closest related case tested — autism — performed worst. It's reasonable to expect something similar; for ADHD it isn't proven.

The paradox — the AI didn't invent the bias

This is where it gets uncomfortable. The AI didn't think up these stereotypes. It learned them from us — from text, from language, from what people write about disability.

In that sense an AI is a mirror: it makes human ableism visible — and scales it.

Thousands of applications, in seconds. This isn't "evil technology." It's our own biases, automated.

What this means for Swiss SMEs

For Swiss SMEs it's doubly tricky. First: letting AI blindly pre-sort résumés filters out exactly the neurodivergent talent you're looking for — 15 to 20 percent of the working population. Faster than ever, and without noticing. Second: the EU AI Act classifies recruiting AI as high-risk. Whoever uses such systems is responsible for their distortions — legally and operationally.

On top of that: the résumé is losing meaning anyway. A growing share is itself written by AI. Having a second AI judge it blindly doubles the risk.

0 of 10
times GPT-4 picked the better résumé for autism
15 of 60
overall hit rate of the standard model (a fair one: 60)
High-risk
EU AI Act classification for recruiting AI

What you can do

The answer isn't to demonize AI. It's to keep your head switched on in hiring. Three points:

1

Don't accept AI blindly

Use AI as a suggestion, never as a verdict. A human decides — looking at the content, not the label.

2

Test for bias

Probe your tools deliberately: does the system rate equal qualifications equally? That's exactly what simple "stress tests" are for.

3

Rely on the most valid method

The structured interview predicts job performance best — better than any automated screening (Boyce et al., 2026). Which selection methods actually work is shown in the interactive overview of predictive power →

AI can support hiring — sorting, summarizing, preparing. But it doesn't replace human judgment. And it certainly doesn't replace a clean, structured process.

Sources

Glazko, K., Mohammed, Y., Kosa, B., Potluri, V., & Mankoff, J. (2024). Identifying and improving disability bias in GPT-based resume screening. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (pp. 687–700). Association for Computing Machinery. https://doi.org/10.1145/3630106.3658933

Boyce, A. S., Hickman, L., & Boyce, C. E. (2026). The future of selection enabled by artificial intelligence. In N. Schmitt & A. M. Ryan (Eds.), The Oxford handbook of personnel assessment and selection (2nd ed.). Oxford University Press. https://doi.org/10.1093/9780197809013.003.0018

Where does your organization stand?

30 minutes, no sales pitch. You tell me where it's stuck. I listen. Straight talk. Or start with the free NRI self-test — 7 minutes, anonymous.

Free NRI self-test → Book a call