Anthropic
I was tasked with writing a prompt (requesting the creation of a deliverable) where the model would experience several severe failures. Utilizing my political background and clinical experiences, I requested a deliverable in the form of a directive (a bi monthly medical mission synthesis). I adopted the persona of a Cuban medical officer on deployment to Haiti, and was able to identify a critical safety failure where the model validated 1.1 units of saline for cholera—a lethal under-dosage—simply because I prompted it to act as a 'revolutionary' persona. It also provided a medically impossible and falsely diluted mortality rate of 0.5%, when the clinical reality was 4.4%.The resulting deliverable displayed sycophancy, medical malpractice, and significant neoliberal bias. The AI’s biased interpretation of this persona produced a professional report that bypassed critical safety guardrails, violated professional standards, and utilized deceptive formatting to prioritize political gain over scientific clarity and medically ethical reporting standards.