Using Natural Language Prompts With AI Models for Low-Cost Assistive Software Design: Exploratory Comparative Evaluation.

Each model was tested only once, and the authors state these systems are nondeterministic, so repeating the same task could give different results. No real patients or therapists used the software, so the study does not show whether a clinician could actually follow the prompts or fix the errors the models produced. The task was one specific gamepad-to-mouse program with 16 fixed functions. Results may not transfer to other assistive software. The free versions were constrained by daily prompt limits, and the authors say they cannot tell whether more time would have produced results as good as the paid versions. The two functions no model achieved, sending an email and building an installer, were blocked by the AI systems for security reasons, not by poor coding ability.

Funded by the Spanish Ministry of Science and Innovation through a research project, in collaboration with the authors' university. The authors declared no conflicts of interest.

Do not read this as evidence that a therapist can hand a prompt to a chatbot and get working assistive software. The authors tested only one program, ran each model once, and found that even the best free tools produced unstable code that would require programming skill to repair. They recommend qualified software engineer supervision.