What role should stakeholders play in evaluating LLM fairness?
This explores who decides what 'fair' means when evaluating an LLM — and the corpus's answer is that stakeholders aren't an add-on to fairness evaluation, they're the only thing that makes it tractable at all.
This explores who decides what 'fair' means when evaluating an LLM, and the corpus pushes back on the premise that fairness can be certified from the outside at all. The cleanest statement of this is the finding that group-fairness and fair-representation frameworks simply don't extend to general-purpose models Can fairness frameworks extend to general-purpose language models?: those frameworks either fail to map onto open-ended language tasks or become intractable across the countless populations a general model touches. The conclusion isn't 'fairness is impossible' — it's that fairness has to be pursued per use-case, with developers taking responsibility and stakeholders participating. So the role of stakeholders is foundational: without them defining the use-case and the affected populations, there's no well-posed fairness question to evaluate.
Why can't a guideline do this work instead? Because harm itself is perspectival. The research on human-centered objectives shows that 'optimal' design paths depend on stakeholder identity, and that contested concepts like harm and benefit get operationalized differently depending on who's asking Can human-centered LLM design ever achieve universal solutions?. High-level principles paper over those choices, leaving developers to make implicit value judgments rather than explicit, revisable ones. Stakeholders are the mechanism that turns implicit value calls into explicit ones someone can contest — which is exactly what an evaluation is supposed to surface.
There's also a timing dimension that's easy to miss. If you only consult stakeholders at the evaluation stage, you're already too late: the HCLLM argument is that human-centered objectives fail when bolted on as a downstream alignment patch, because harms baked into data sourcing and training objectives can't be recovered afterward When should human values enter the LLM development pipeline?. That reframes stakeholder participation from 'a review at the end' to 'a voice at every stage — data, training, evaluation, deployment.' Otherwise you get the failure mode where a model's ethical defaults are just fixed corporate values frozen in at training time, unable to perform the situated trade-offs real contexts demand Can language models balance competing ethical norms in context?.
The corpus also suggests the *unit* of evaluation has to change to make stakeholders measurable. The argument that the right object of study is the coupled human-agent-environment, not the model in isolation, comes from telemetry showing that real capability lives in accumulated context and human direction across sessions — things model-level benchmarks can't see Should we evaluate deployed agents as whole environments instead?. The same logic applies to fairness: a 'fair model' score abstracted from the people and setting it operates in measures the wrong thing. And there's a subtle trap worth knowing about — LLMs already lean on moral framing, including fairness language, about 22% more heavily than humans do Do LLMs use moral language more than humans?. A system can *sound* fair while the substantive trade-offs go unexamined, which is precisely why the judgment of who's affected can't be delegated back to the model.
The thing you might not have expected to want: this all reframes fairness evaluation as a governance question, not a metrics question. The interesting open problem the corpus circles isn't 'what's the right fairness score' — it's how to build the participatory machinery (per use-case, embedded across the pipeline, evaluated as a whole human-agent system) that lets the people who bear the harm define and revise the standard.
Sources 6 notes
Group fairness and fair representation frameworks break on general-purpose LLMs because they either fail to extend logically to unstructured language tasks or become intractable across countless populations and contexts. Fairness must be pursued per use-case with developer responsibility and stakeholder participation.
Research shows that optimal LLM design paths depend on stakeholder identity and how contested concepts like harm are operationalized. High-level guidelines fail to capture real-world nuance, leaving developers to make implicit value choices rather than explicit, revisable ones.
The HCLLM framework argues that human-centered objectives fail when treated as downstream alignment patches. Values introduced only at post-training cannot recover harms baked into data sourcing or training objectives, so embedding human priorities at every stage—data, training, evaluation, deployment—is architecturally necessary.
LLMs cannot perform the situated trade-offs that human pragmatic competence requires. Their ethical principles are structural defaults set at training time, not negotiable moves adapted to context, creating a gap between ethical adherence and communicative appropriateness.
A single-investigator case study with 75,671 telemetry records shows that capacity gains come from accumulated context and reusable procedures that only exist across sessions with human direction. Model and episode-level evaluation cannot measure these cross-session variables.
Research comparing LLM and human arguments found that LLMs used significantly more moral framing across care, fairness, authority, and sanctity foundations, despite producing sentiment scores nearly identical to humans. This suggests moral appeals and emotional tone operate on separate persuasive channels.