UABDivulga - Barcelona recerca i innovació

When we talk about artificial intelligence, we almost always imagine “normal” systems: orderly, efficient and predictable. In our paper, however, we start from a more uncomfortable and provocative question: what if, to tackle complex problems, we also needed “neurodiverse” AIs, inspired by traits we associate with autism, ADHD or dyslexia?

The idea of neurodiversity reminds us that not all brains work in the same way. Autistic people, people with ADHD, dyslexia or other profiles often process information differently: with stronger hyperfocus on certain details, heightened sensory sensitivity, or alternative ways of connecting ideas. These differences can create serious challenges in everyday life, but they can also give rise to distinctive strengths.

Most current AI systems, by contrast, are based on an idealised “neurotypical” brain. They optimise for average performance, stability and regularity. That works well for many tasks, but it leaves out a large part of the possible “catalogue” of ways of thinking. In our work, we ask what an AI would look like if it were designed from the start to integrate multiple cognitive styles, including those we link to neurodivergence.

Instead of building a single homogeneous artificial mind, we imagine ecosystems of AI made up of specialised modules. Some would display a kind of hyperfocus that helps to detect subtle patterns in huge datasets; others would explore rare or unlikely solutions, echoing more chaotic forms of thought; yet others would combine information in non-linear ways, creating unexpected connections. This internal diversity makes the overall system more robust, creative and adaptable.

Crucially, we are not trying to “build an autistic AI”, nor to romanticise conditions that, for many people, involve real suffering and discrimination. What we do is to extract functional principles —such as hyperfocus, sensitivity to patterns or tolerance for uncertainty— and ask how they might be implemented in AI architectures, while keeping respect for neurodivergent people at the centre and avoiding stereotypes.

We also examine the ethical side of this research. AIs that can concentrate so intensely on specific details could be powerful tools in medicine, personalised education or anomaly detection, but they could also magnify existing biases or be used for intrusive surveillance. For this reason, we argue that any move towards neurodiverse AI has to be accompanied by open public debate, clear regulation and real involvement of neurodivergent communities.

So when we ask whether we “want autistic AIs”, the answer is not a simple yes or no. Our proposal is to use the richness of neurodiversity as a source of inspiration to design more diverse and imaginative AIs, while remembering that behind these ideas there are real lives that deserve respect and rights, not just technological curiosity.