Critical Perspectives on Artificial Intelligence Research for Image-Based Autism Detection

Authors

Abstract

In recent years, possible applications of artificial intelligence (AI) in identifying neurodivergent people have received increasing attention. As an example, a publicly available database of children's facial images hosted on Kaggle has been used to train AI models for autism diagnosis. The database includes approximately 3000 images of putatively autistic and neurotypical children and has been used in around 50 AI articles over the past five years, with more than 500 citations in total. However, the data were compiled using online searches on autism websites and social networks. This article critically examines the scientific validity of such a dataset for determining a child's autism status, and highlights significant ethical and methodological flaws in the database construction and use. We argue that these flaws could have been prevented if the authors had partnered with community members and interdisciplinary experts. Without rigorous methods, appropriate ethical review, and genuine understanding of the populations being studied, AI models can easily become invalid. This article calls for a re-evaluation of the ethical implications and scientific integrity of using weakly validated datasets in AI for autism research.

No cited bibliography is currently integrated for this reference. This does not necessarily mean that no bibliography exists.