Uploaded July 2026 | Updated November 2026, 4 days ago
AI Hiring Bias: When a Disability Signal Makes the AI Decisive
This is the control result in the AI hiring bias research that almost everyone skips over, and it might be the most revealing finding of all. Before testing disability, the researchers gave GPT-4 two completely identical CVs and asked it to rank them. Logically it should call them equal every time. It only did so seventy per cent of the time. In the other thirty per cent it picked a winner between two documents that were exactly the same, with contradictory reasoning (Glazko et al., 2024).
Hold that number next to what happened when disability entered the picture. Across all sixty comparisons where one CV carried a disability signal, the model never once called a tie. Sit with the contrast. Faced with genuinely identical evidence, the system was often uncertain. Faced with a disability signal, its uncertainty vanished and it made a firm decision. In algorithmic hiring, a disability signal did not add doubt. It removed it.
That is the finding that reframes everything else in this AI recruitment bias series. It tells us the disadvantage is not random noise. When there was nothing meaningful to separate two candidates, the model wavered. When the only meaningful difference was disability, it committed. For autistic, ADHD and AuDHD job seekers moving through AI resume screening and algorithmic hiring, this is the structural reason the outcomes feel so consistent. The system is at its most decisive at the exact moment it should be most careful.
This short also covers what happened when the researchers tried to fix it. They built a custom model trained to avoid ableist bias. Overall it improved, and for autism specifically it moved only a little, and the change was not statistically significant. The authors ask people not to present that as a solution, so I do not. A partial mitigation on a small sample is not a fixed system, and treating it as one would let the underlying AI hiring bias off the hook.
If you have ever felt that these processes reach a verdict about you almost instantly, this research gives you the receipts. It also gives you language. Instead of leaving an application convinced you did something wrong, you can name the structural pattern: a system that is most decisive at the exact moment it holds the least reason to be. That reframing matters, because self-blame is exhausting and it is often misplaced. Watch the full breakdown for the complete picture, including what the control result means for every other finding in this series, then grab the free toolkit if you want the legal grounding across five jurisdictions and the scripts to advocate for yourself with confidence.
AI Hiring Bias: When a Disability Signal Makes the AI Decisive
This is the control result in the AI hiring bias research that almost everyone skips over, and it might be the most revealing finding of all. Before testing disability, the researchers gave GPT-4 two completely identical CVs and asked it to rank them. Logically it should call them equal every time. It only did so seventy per cent of the time. In the other thirty per cent it picked a winner between two documents that were exactly the same, with contradictory reasoning (Glazko et al., 2024).
Hold that number next to what happened when disability entered the picture. Across all sixty comparisons where one CV carried a disability signal, the model never once called a tie. Sit with the contrast. Faced with genuinely identical evidence, the system was often uncertain. Faced with a disability signal, its uncertainty vanished and it made a firm decision. In algorithmic hiring, a disability signal did not add doubt. It removed it.
That is the finding that reframes everything else in this AI recruitment bias series. It tells us the disadvantage is not random noise. When there was nothing meaningful to separate two candidates, the model wavered. When the only meaningful difference was disability, it committed. For autistic, ADHD and AuDHD job seekers moving through AI resume screening and algorithmic hiring, this is the structural reason the outcomes feel so consistent. The system is at its most decisive at the exact moment it should be most careful.
This short also covers what happened when the researchers tried to fix it. They built a custom model trained to avoid ableist bias. Overall it improved, and for autism specifically it moved only a little, and the change was not statistically significant. The authors ask people not to present that as a solution, so I do not. A partial mitigation on a small sample is not a fixed system, and treating it as one would let the underlying AI hiring bias off the hook.
If you have ever felt that these processes reach a verdict about you almost instantly, this research gives you the receipts. It also gives you language. Instead of leaving an application convinced you did something wrong, you can name the structural pattern: a system that is most decisive at the exact moment it holds the least reason to be. That reframing matters, because self-blame is exhausting and it is often misplaced. Watch the full breakdown for the complete picture, including what the control result means for every other finding in this series, then grab the free toolkit if you want the legal grounding across five jurisdictions and the scripts to advocate for yourself with confidence.

![ADHD & Autism: Sleep Challenges Explained
ADHD & Autism: Sleep Challenges Explained
ADHD sleep, autistic sleep and AuDHD sleep each have their own story. Here is the whole series in one Short.
ADHD sleep is largely a timing story: adults with ADHD often sleep similar hours to others, but later and less regularly (Kaya et al., 2026; Walter et al., 2026). Autistic sleep is largely a regularity and load story: autistic adults slept as long as their relatives on average, but less consistently, and inactivity was linked to shorter sleep (Weissenkampen et al., 2026). AuDHD appears to need both levers at once. Not every study agrees that AuDHD sleep is worse: one study of children with ADHD found similar sleep problems whether or not they were also autistic, with anxiety and emotional load as the key predictors (Thomas et al., 2018).
In this Short, I bring the three parts of my sleep series together so you can see where you fit, and where to start. If you are ADHD, start with timing. If you are autistic, start with regularity and load. If you are both, try one change for each.
For leaders and managers: AuDHD staff may be managing a late body clock and a busy nervous system at the same time. Flexible start times, predictable routines, sensory adjustments and a clear fatigue risk conversation can make a real difference to wellbeing and safety at work.
I would love to hear from you in the comments. What gets in the way of sleep for you, and what has actually helped? Your experience helps other people feel less alone, and it shapes the videos I make next.
This Short comes from Part 3: AuDHD and Sleep: Why Autism Plus ADHD Is Its Own Profile, one of three videos in my neurodivergent sleep series. Watch the full video here: [FULL VIDEO LINK]. Grab the free printable brain dump and two-week sleep log here: [DOWNLOAD LINK]. The whole series, covering ADHD, autism and AuDHD sleep, is in this playlist: [PLAYLIST LINK]. Subscribe for more research you can actually use.
I am Nicola Knobel, a late-diagnosed AuDHD woman and a health, safety and risk professional based in Aotearoa New Zealand. I make research-based videos for neurodivergent adults and the leaders who support them, and I always show you where the evidence comes from, and where it is still thin.
This video is general information and is not medical advice. If sleep is affecting your health, work or safety, please talk to your GP. In New Zealand, melatonin is a restricted medicine, so any conversation about it belongs with your doctor. ADHD & Autism: Sleep Challenges Explained](https://i.ytimg.com/vi/GdVnDgnsySI/mqdefault.jpg)








