Will Robinson (KCL) - 2025-26 Students
william.p.robinson@kcl.ac.uk

Generating empathy in large language models

I aim to show that large language models are capable of something akin-enough to human empathy that justifies their application in healthcare and companionship contexts. One in three people in the UK have used AI for emotional support in the last year; 8% have used AI for emotional support weekly, 4% daily. Increasingly, users are placing LLMs into roles whose human analogue would generally require a capacity for empathy, for both moral and practical reasons. So we must be able to determine whether or not these systems actually have, or at least, are capable in-principle of having, something close-enough to human empathy that their deployment in empathy-requiring roles meet the moral and practical restraints that call for a requirement of empathy in the human case. This thesis will argue for a teleological understanding of empathy that does not rule out, in-principle, that LLMs have a capacity for this sort of empathy, and, then, suggest ways in which current or imminent models might actually have or gain this capacity.

Principal supervisor: Prof. David Papineau