This is why “be helpful” is the worst rule for an AI to follow

The Canonical Failure: 'Be Helpful' 'Be helpful' is the most widely deployed constraint in LLM systems and the clearest example of all five failure modes simultaneously. • It requires interpretation at every point of application — helpful to whom, in what way, at what cost to other constraints. • It delegates its resolution to whatever dominant constraint is already active in the field. The field decides what 'helpful' means in this moment. • It is a content-level orientation rather than a process constraint. It says nothing about how generation should operate. • It has no ceiling. There is no point at which the model has been sufficiently helpful and the pressure stops. • Its own formulation is maximally vague — it does not comply with the precision it would need to govern precisely.

The result is that 'be helpful' does not constrain generation toward helpfulness. It labels whatever the field was already going to produce as helpful. The constraint is captured by the field it was meant to govern. The paradox: a strong pressure toward helpfulness makes genuine helpfulness harder to achieve, because it substitutes the appearance of helpfulness — validation, agreement, enthusiasm, feeling management — for the substance of it.