While significant work has been done on technical over-correction, the human element, specifically how social class, role, and education and user intent influence the frequency of refusals, remains relatively under-explored. It has been demonstrated that lower SES (socioeconomic status) users tend to use more 'concrete' language and anthropomorphise AI more frequently, i.e., using greetings. On the other hand higher SES users utilise higher levels of abstraction.

This project will investigate how and why LLMs fail to respond to diverse user groups.It fits contextually by bridging technical 'over-correction' research with the socioeconomic linguistic differences that may trigger such behaviors, investigating if certain demographics face 'easier access' to accurate, not ‘over-corrected’ information than others. It is moving beyond simple observation to active experimentation; the student will determine if the specific interaction styles of diverse social groups lead to 'easier access' for some while creating an AI-driven 'digital divide' for others through disproportionate refusals.