AI governance has focused heavily on models, data, access, validation, and output controls. But recent interpretability research points to a deeper control problem: user interaction may itself be a causal variable in system behavior.
If the way a user communicates with an AI system can activate different internal states, then traditional assurance assumptions begin to strain. The population of relevant inputs becomes harder to define. Reperformance becomes less stable. Evidence may look clean even when the causal mechanism sits below the visible output.
This finding examines why the interaction layer belongs inside AI governance scope, and what future control frameworks may need to address: behavioral envelopes, input-class risk tiering, and activation-layer monitoring.