A look at “A Framework for Context-Aware and Adaptive Large Language Models for Autism Spectrum Disorder Therapy” by Achhiya Sultana, Mohammad Shidujaman and Md Tarek Habib (2026)
This paper comes from computer scientists in Bangladesh. It proposes using AI language models, the same family of technology behind tools like ChatGPT, to support autistic children who speak Bangla (Bengali).
What it proposes
The authors start from a real problem. Support for autistic children is often scarce, generic and available mainly in English or other widely spoken languages. Bangla is spoken by hundreds of millions of people, yet it has relatively few tools tailored for autism support.
Their solution is a conversational system built on several existing Bangla and multilingual AI models. It is designed to:
- understand and respond in natural Bangla, including emotional tone;
- adapt to each child’s context, needs and responses in real time;
- be available across clinical, educational and home settings.
They describe how they trained and fine-tuned the models, including using data from online autism groups, and report technical performance scores. They also describe testing the system with autistic users and gathering feedback from caregivers and clinicians.
A grain of glitter ✨
The goal of making support accessible in under-served languages is genuinely worthwhile. But this paper raises more questions than it answers.
Most importantly, the paper describes testing the system with autistic children, yet its ethics and consent statements both say “nothing to report”. For research involving autistic children, that’s a significant gap. It isn’t clear how many children took part, how they were recruited, or what safeguards were in place. One section even refers to participants discussing “workplace communication issues”, which doesn’t fit a study about children and raises questions about how carefully the paper was put together.
It’s an engineering paper, so “performance” mostly means technical accuracy scores, not whether children’s wellbeing actually improved. It also relies on data from online autism groups, which may not represent the children it aims to help, as the authors acknowledge. And the framing is firmly deficit-based, opening with gene mutations and toxins as possible causes, with no neurodiversity-affirming perspective or input from autistic people.
Why this matters
AI tools for mental health and neurodivergent support are multiplying fast, and some may genuinely help, especially where human support is scarce. But a chatbot isn’t a therapist. If you or your child use an AI tool, it’s wise to ask who built it, what evidence supports it, how your data is used, and whether autistic people helped design it. Human connection, adapted to how you communicate, still matters enormously.
Robots may be getting cleverer, but no algorithm has yet worked out how to give a proper unicorn hug.