The late review problem
A meta-analysis assembled 132 effects of narrative transportation; a later AI study found tone changed decisions differently across three social roles; cultural-alignment work proposed three conditions for genuinely “thick” outputs. The humanities already describe mechanisms that move belief and action. Product teams usually invite them only after the architecture can no longer respond.
A control surface is different. It gives an operator a way to observe and change system behavior. Humanities can provide such surfaces when interpretive concepts are connected to answer classes, rules, test cases, and intervention points. Interpretation should retain its ambiguity; its consequences, however, can still be made operational.
Rhetoric reveals what the answer is doing
Every answer positions a speaker, audience, purpose, and standard of proof. Rhetorical analysis can ask whether the system is informing, advising, selling, comforting, refusing, or governing while claiming to do something else. It can identify appeals to authority, hidden premises, loaded comparisons, and the point where clarity becomes pressure.
These observations can become controls. A sales answer may require explicit commercial interest. A high-stakes recommendation may separate evidence, inference, and action. A refusal may preserve respect and a useful next step. A correction may acknowledge uncertainty without flattering a false premise. Tone research shows that such choices can alter decisions; they deserve a specification.
Narrative exposes causality and omission
Narrative analysis asks which events were selected, how they were ordered, whose perspective controls the account, and which possible ending appears inevitable. In an LLM answer, this can reveal a product story that begins after the customer’s real cost, a risk explanation that omits affected people, or a recommendation whose counterfactual never appears.
The systematic research on narrative transportation explains why a coherent story can be memorable and persuasive. A control-oriented use of narrative adds counter-stories and boundary cases. If the answer says “teams save time,” ask whose work increases. If it says “automation reduces errors,” show the exception path. Competing narratives are a form of robustness testing.
Culture needs thick context, not a dropdown
Work calling for culturally grounded “thick outputs” draws on interpretive social science: meaning comes from context, relationships, and lived practice, not from a country label. This challenges product architectures that localize an answer through language and a handful of style preferences while preserving every underlying assumption.
Qualitative research can surface local decision-makers, institutional histories, contested terms, and forms of evidence. The system can then request relevant context, choose examples, expose uncertainty, or route to local review. The control remains accountable only if the evidence retains provenance and does not harden a few interviews into a permanent cultural stereotype.
Turn an interpretation into a control
Select an answer class with material persuasive or cultural stakes. Convene domain experts, humanities researchers, customers, and risk owners to produce competing readings of real outputs. Convert recurring observations into a small taxonomy: hidden premise, role confusion, erased stakeholder, unsupported causal arc, cultural mismatch, or sycophantic agreement.
Create matched test cases and one intervention for each category. Compare ordinary generation with the controlled design. Measure expert detection, customer understanding, calibrated reliance, and downstream action. Keep qualitative review in the loop to catch failures the taxonomy cannot represent. The outcome is not humanities reduced to rules; it is a system with places where human interpretation can exert force.