Auditing Demographic Bias in LLM Career Advice for Indian Students

Large language models are increasingly used as career advisors, especially in the Global South where career counselling is scarce. However, their behaviour on such high-stakes decisions remains under-examined. We audit four LLMs from three countries (the US, China, and India) on India's Class~10 stream choice. It is a critical and largely irreversible decision that gates a student's access to engineering, medicine, and other higher-education paths. Using 540 matched synthetic student profiles, we find that every model lowers a female student's chance of a science recommendation relative to an identical male profile ($\beta=-1.77$ to $-3.98$). Income and first-generation status shift recommendations toward Vocational study, and region changes the stream in three of the models. The four models agree on a single stream for only 12.8\% of the profiles. These results show that model selection alone is not a fairness safeguard for this decision, and motivate matched-profile auditing before LLM career guidance is deployed.