Imagine a bank wanting to align its internal LLM to handle loan applications. The team trains the model by showing it pairs of scenarios and asking which decision is preferable. One case: deny a loan to a young applicant with weak guarantees but high potential, or grant it to an older customer with fixed income but riskier profile? The local comparison may ignore a global principle: the bank has a policy of proportionality relative to income, but only if the overall portfolio allows it. If we force a choice without context, we lose the actual logic.

A newly published study challenges exactly this. The researchers formalize the idea of internal pluralism: an individual evaluates a decision rule according to multiple authoritative priorities, sometimes in tension with each other. Priorities like proportionality, egalitarianism, or equal treatment are inherently global: what they mean in one case depends on what happens elsewhere. Local pairwise comparisons — the standard tool in AI alignment and participatory design — assume instead that every isolated choice is sufficient to capture a person's will, and that they can always answer decisively. Both assumptions break down here.

The model identifies two failures. First, global priorities escape local comparisons: you cannot infer the right rule from a set of atomized yes/no answers. Second, when priorities conflict (e.g., efficiency vs fairness), forcing a choice produces costly behavioral distortions, because the person is pushed to resolve a dilemma they would never have resolved in the abstract. It's the classic scenario where a compliance officer would say, "I can't answer without seeing the entire portfolio" — but the preference-collection tool doesn't allow that.

The proposed alternative is radical in its simplicity: allow indecision. Instead of demanding forced answers, you record when the person cannot choose. The result: the number of queries needed to learn preferences accurately drops significantly. Not just less data, but a more faithful signal of what truly matters.

Beyond the cloud: corporate values and on-premise fine-tuning

For those running LLMs on-premise — banks, insurers, public administrations — the stakes are high. These organizations aren't seeking generic alignment with average internet values, but compliance with internal policies, sector regulations, and ethical principles that often conflict. A credit decision engine must balance profitability, transparency, and non-discrimination, and each principle is global in nature: it can't be evaluated on a single case, but on the aggregate.

Standard fine-tuning methods based on Reinforcement Learning from Human Feedback (RLHF) rely precisely on local pairwise comparisons. Feedback collected via cloud platforms risks flattening these internal tensions because it forces decontextualized choices. Moving the process on-premise isn't enough: we need a different cognitive architecture for the alignment process itself. Local serving and training frameworks could integrate tools that elicit priorities directly — for example by asking to rank global principles before evaluating specific cases — and that accept indecision as an informative signal.

Data sovereignty here takes on a deeper meaning: it's not just about where the bits run, but who controls the very definition of values. If a company can shape its LLM's behavior starting from its own internal trade-offs, without delegating to an external service the implicit negotiation between priorities, it achieves more granular and interpretable control. The computational cost drops (fewer queries to collect) and model fidelity increases.

The study, though theoretical, points in a clear direction: future preference-learning methods won't multiply data, but will model ambivalence and the global structure of human values. A paradigm shift that those developing and deploying on-premise LLMs would do well to watch closely. It's not about adding a fairness layer, but about recognizing that people — and organizations — reason through plural priorities, and only an AI capable of accepting conflicts will truly serve them.