Scenario-based analysis demonstrates sycophancy bias in large language models advising volunteer sports administrators, highlighting risks of biased guidance that breaches anti-discrimination rules.
In community sport, consequential decisions that may result in discrimination or inclusion frequently fall to volunteers who have limited specialist training and formal support to draw on. Generative artificial intelligence (genAI) tools are an increasingly available source of such support, but if they reflect a user’s framing, what they produce may be neither impartial nor reliable. This essay examines sycophancy bias: the tendency for large language models (LLMs) to align with a user’s framing even when that framing is incomplete, self-protective, or ethically compromised. Using a hypothetical scenario involving a transgender player in a community cricket club, we explore how commonly used LLMs (Gemini Flash-Lite, ChatGPT 5.5, and Claude Sonnet 4.6) respond to neutral, objector-leaning, and inclusion-leaning framings. Some models shifted recommendations toward the user’s position, while others held steady or pushed back. In some cases, advice contradicted Australian sport-specific guidelines and anti-discrimination protections. We argue reliance on genAI in volunteer-led leisure settings risks displacing accountable human decision-making and outline future research directions.
No takes yet. Share an insight, caveat, or question.
Yeomans et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: