By Arvin Obnasca Chercheur et Analyste en Éthique de l’IA et Normalisation Internationale at BeEthical (Amiens, France)
This blog is produced as a part of the Reshaping Work Fellowship Programme. The opinions and views expressed in this publication are those of the author. They do not purport to reflect the opinions or views of Reshaping Work or organisations that have supported the programme.
Imagine being rejected by the same tool that could have made your working life better than it has ever been.
Bloomberg Law reported in 2025 [1] that AI video interview tools were marking autistic applicants down. The reason? How they make eye contact. How their voice sounded. How their face moved. None of these things has anything to do with whether someone can do a job. But the algorithm was trained on neurotypical workers, so it learned that a “good candidate” looks and sounds in a particular way. If you fall outside the pattern, you get filtered out. The US government’s guidance [2] has warned that facial and voice analysis tools can screen out qualified autistic people, even when they are fully capable of doing the work.
That is one half of the stories. Here is another. A UK Department for Business and Trade study [3] found neurodivergent workers were 25% more satisfied with AI assistants than their neurotypical colleagues. Scheduling tools, task managers, and communication aids were reshaping the workday around how neurodivergent people actually think. Not the other way around. When the support fits, the results follow. GP Strategies found [4] that neurodivergent workers can be 30% more productive than neurotypical colleagues when given the right tools.
So here we are. The same technology rejects autistic people at the door and empowers them once they are inside. In Dissemination [5], the philosopher Jacques Derrida had a name for things like this. He called them pharmakon, an ancient Greek word that means both remedy and poison at once. Not because the thing is broken, but because that duality is its nature. You cannot ban the poison without losing the cure.
The researcher Damian Milton, whose work on autism and communication has been widely influential, saw something deeper years before AI was part of the conversation. His “double empathy problem”[6] makes a simple but radical point. The communication difficulties autistic people face are not one-sided. Neurotypical people struggle to understand autistic people just as much as the reverse. The difference is that neurotypical people designed the systems. Their way of being became the default. AI hiring tools inherit that default and apply it on an enormous scale. EEOC data [7] shows that autism-related workplace discrimination charges have risen sharply in recent years, with merit resolutions more than doubling between 2016 and 2023.
Europe has begun to act. The EU AI Act [8] banned emotion recognition in the workplace from February 2025. That is a real win for autistic workers, who were among the most exposed to systems that scored facial expressions and vocal tone. The high-risk framework [9] goes further, classifying AI used in hiring and employment as high-risk, which means these systems must be tested for discriminatory outcomes before they reach the market. As of 7 May 2026, that deadline has officially shifted. The Council and Parliament reached an agreement [10] on the Digital Omnibus on AI, pushing high-risk obligations back to December 2027 instead of August 2026. The deal also does something quieter. It widens the use of sensitive personal data to test AI for bias. That could, in principle, include neurotype. The door is open. Nobody has walked through it yet
But the question remains. While the emotion recognition ban is clear, the picture for personality assessments, video interviews, and behavioural monitoring tools is more ambiguous. They may still penalise neurodivergent traits without being captured by the narrower prohibition on workplace emotion recognition. These tools may fall under the high-risk framework, depending on how regulators interpret discrimination in practice. Whether neurotype is treated as a relevant axis remains an open question. As it stands, bias testing often focuses on legally recognised or routinely measured categories such as race, gender, and age. Neurotype is not explicitly included. Whether AI hiring tools systematically exclude people who think differently is yet something regulators are required to test. Ireland’s General Scheme of the Regulation of Artificial Intelligence Bill 2026 [11], published in February, is the first national legislation to operationalise the EU framework, with a dedicated AI Office by August 2026. But neurotype does not yet feature on its agenda either.
This is where Derrida’s thinking becomes practical. You do not fix a pharmakon by picking sides. You govern the conditions under which it is used. Three steps would probably make a real difference. First, bias audits under the EU AI Act should be required to test for the exclusion of cognitive variation, not just race and gender. Second, if AI is used in hiring, candidates should have the right to a non-algorithmic pathway. Third, AI workplace tools should be built with neurodivergent workers from the start. The European Accessibility Act [12] already moves in this direction for products and services. Extending that logic to algorithmic systems is not a leap. It is a consistent next step.
Workplaces have always preferred a certain kind of person, i.e. someone who makes eye contact, reads the room, and follows unwritten rules. AI did not create that bias. And now it runs at a scale no human hiring team ever could, thousands to millions of decisions a year. We built AI to find the best candidates. It found the most typical ones instead. But it has also built something else along the way: tools that, for the first time, fit how neurodivergent minds can actually work. The question for regulators is whether they hold both truths at once. Whether they can govern a technology that is, by its nature, both the barrier and the bridge.
[1] Bloomberg Law. (2025, August). AI hiring tools elevate bias danger for autistic job applicants. Bloomberg Law. https://news.bloomberglaw.com/daily-labor-report/ai-hiring-tools-elevate-bias-danger-for-autistic-job-applicants
[2] U.S. Department of Justice. (2025). Algorithms, artificial intelligence, and disability discrimination in hiring. ADA.gov. https://www.ada.gov/resources/ai-guidance/
[3] CNBC. (2025, November 8). Why workers with ADHD, autism, dyslexia should use AI agents. CNBC. https://www.cnbc.com/2025/11/08/adhd-autism-dyslexia-jobs-careers-ai-agents-success.html
[4] GP Strategies. (2024). 5 AI tools to foster a more inclusive work environment for neurodiverse learners. GP Strategies. https://www.gpstrategies.com/blog/5-ai-tools-to-foster-a-more-inclusive-work-environment-for-neurodiverse-learners/
[5] Derrida, J. (1972). Dissemination. Éditions du Seuil. https://press.uchicago.edu/ucp/books/book/chicago/D/bo27619136.html
[6] Milton, D. (2012). On the ontological status of autism: The ‘double empathy problem’. Disability & Society, 27(6), 883–887. https://doi.org/10.1080/09687599.2012.710008
[7] Ogletree Deakins. (2025, April). Disability discrimination charges involving neurodivergence are rising, according to EEOC data. Ogletree Deakins. https://ogletree.com/insights-resources/blog-posts/disability-discrimination-charges-involving-neurodivergence-are-rising-according-to-eeoc-data/
[8] European Parliament & Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. https://artificialintelligenceact.eu/article/5/
[9] European Parliament & Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Annex III. Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
[10] Council of the European Union. (2026, May 7). Artificial intelligence: Council and Parliament agree to simplify and streamline rules [Press release]. https://www.consilium.europa.eu/en/press/press-releases/2026/05/07/artificial-intelligence-council-and-parliament-agree-to-simplify-and-streamline-rules/
[11] Department of Enterprise, Tourism and Employment. (2026, February 4). General Scheme of the Regulation of Artificial Intelligence Bill 2026. Government of Ireland. https://enterprise.gov.ie/en/legislation/general-scheme-of-the-regulation-of-artificial-intelligence-bill-2026.html
[12] European Parliament & Council of the European Union. (2019). Directive (EU) 2019/882 on the accessibility requirements for products and services. Official Journal of the European Union. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32019L0882