Can fairness frameworks extend to general-purpose language models?

Machine-learning fairness frameworks — group fairness, fair representations — were built for systems with well-structured inputs and outputs and a self-evident use (lending, recidivism, coreference). This work argues they break on general-purpose LLMs: each framework either does not logically extend to LLM tasks (unstructured natural-language data) or becomes intractable, because LLMs touch a multitude of populations, sensitive attributes, and use cases at once. The conclusion is sharp: it is not feasible to certify or guarantee that an LLM is generally "fair."

The constructive move is to lower the target from universal fairness to use-case-specific fairness, governed by three guidelines: the criticality of context, the responsibility of LLM developers, and stakeholder participation in an iterative design-and-evaluation process — with the speculative note that AI's general capabilities might eventually help address fairness as a form of scalable AI-assisted alignment.

The keeper is the impossibility-style framing: "fair LLM" is not a global property you can stamp on a model; fairness only has teeth relative to a context and its stakeholders.

This pairs with the vault's preference/alignment-pluralism thread. It mirrors Can human-centered LLM design ever achieve universal solutions? and complements Can a single reward model represent diverse human preferences? — both reject one-size-fits-all normative targets — and supports Should AI alignment target preferences or social role norms? as the contextual alternative.