Research Summary: Gender Associations in LLM-Mediated ADHD Self-Diagnosis

A research study evaluating GPT-4o demonstrates that demographic cues influence large language model outputs during zero-shot ADHD self-diagnostic queries, with explicit gender cues amplifying specific subtype associations and potentially triggering over-correcting bias mitigation mechanisms.

Why it matters

This research matters because uncalibrated artificial intelligence responses in healthcare self-assessment tools can distort diagnostic outcomes along demographic lines. As public reliance on LLMs for medical triage grows, algorithmic over-correction presents operational and clinical risks that can misdirect patient pathways. Ensuring reliable and clinically aligned demographic handling is essential for maintaining safety and trust in AI-enabled health technologies.

Key insights

• Large language models (LLMs) are increasingly utilised by individuals for medical self-diagnosis, yet the influence of demographic cueing on diagnostic outputs remains insufficiently understood.
• Persistent gender disparities exist in ADHD diagnosis that do not entirely align with clinical evidence.
• Using an adaptation of the Context Association Test (CAT), researchers found that GPT-4o responds differently depending on the nature of demographic cueing.
• Explicit gender cueing consistently amplifies specific ADHD subtype associations in zero-shot settings.
• Implicit cueing generates diagnostic distributions that closely align with demographic proxies.
• Divergences between explicit and implicit cueing indicate that explicit cues may trigger bias-mitigation safeguards that over-correct diagnostic suggestions.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.33697