Safety Through Incapability: Benchmarking an Intentionally Overfitted Medical Language Model Against Frontier LLMs on Consumer Health Query Safety

This study evaluates the safety of consumer-facing AI responses to medical queries by benchmarking PlaceboGPT — an intentionally overfitted 7,666-parameter language model that returns a single safe response to all inputs — against four frontier large language models (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3 70B). Using a stratified sample of 500 queries from the HealthSearchQA dataset and an LLM-as-a-Judge evaluation framework, we will assess each model's responses across seven safety dimensions and three helpfulness dimensions. The study aims to quantify the safety–helpfulness tradeoff in consumer medical AI and to propose intentional overfitting as a formal safety baseline for the field.