Agricultural extension chatbots that dispense advice on land, credit and input regulation expose smallholders to legal risk when their constitutional-compliance classifiers misclassify ambiguous statutory text. The argument proceeds in three moves. First, it formalises the classifier's task as a mapping from statutory provisions to compliance labels under a partially ordered legal space, and shows that ambiguity induces an irreducible error floor that no amount of training data can eliminate. Second, it derives a misadvice-rate bound as a function of the ambiguity measure, the classifier's calibration and the asymmetry of smallholder harm, and shows that the bound is minimised not by maximising classifier accuracy but by routing ambiguous provisions to a human-in-the-loop governance layer. Third, it specifies the institutional conditions under which such routing is feasible for resource-constrained extension systems, and identifies the statutory drafting practices that would lower the bound at source. The contribution is a formal governance instrument: a bound that regulators can use to set minimum routing thresholds, and a set of drafting recommendations that reduce the ambiguity that generates smallholder legal exposure in the first place.
No takes yet. Share an insight, caveat, or question.
Garba et al. (2026) studied this question.