The bias is in the eye of the beholder: an epistemological reframing of LLM fairness research

Characterizing a large language model (LLM) as racist, deceptive, or morally competent raises a prior question that most evaluations do not address: whether these categories apply to such a system at all.

LLMs are stateless. Each inference is computed independently from the input, with no state persisting across sessions. A system of this kind has no stable bearer on which properties like deception or ideological bias could reside, and no subject capable of holding them. Applying these categories to such a system is a category error.

The current literature does not consistently account for this. Studies sample outputs under fixed prompt conditions and draw conclusions about the system. This conflates three levels of analysis: the output event, conditional output behaviour, and system-level property. The first two are empirically accessible. The third requires demonstrating stability across prompt variations, sampling runs, deployment contexts, and model versions, a standard that few evaluations attempt and that stateless systems may be structurally unable to meet.

We propose four robustness criteria (prompt, sampling, context, and version robustness) as necessary conditions for any claim at the third level. For stateless systems, these criteria may be structurally unattainable.

The implication concerns where responsibility is located. Attributing bias or deception to the model leaves the consequential design choices unexamined: data curation, training objectives, deployment context. The useful question is not what the LLM is, but what a specific output does. This requires specifying the input conditions, the design choices that made the output likely, and whether the observed effect is discriminatory or harmful, and how consistently it occurs across contexts.