A Threshold on Responsibility in Human–AI Interaction

Explicit Human Ownership

At every point where generated language influences a decision, a specific human individual must be identifiable as responsible for the outcome. Responsibility cannot be collective without attribution, procedural without signature, deferred to policy, or absorbed by the system itself. If no one can be named, responsibility has already been displaced.

Visible System Limits at the Point of Use

System limitations must be present during interaction, not relegated to documentation, onboarding, or legal disclaimers. Limits that appear only after the fact do not function as limits. They function as insulation.

Traceable Acceptance

There must be a durable trace showing what the system produced, under which conditions, and by whom the output was accepted, modified, or rejected. This trace is not for surveillance or optimisation. It exists so that responsibility does not vanish into fluency.

No Substitution of Judgment

Generated language may inform, clarify, or provoke reflection. It must not substitute judgment. When the system's phrasing becomes the final rationale, or its structure becomes the default reasoning, the system is no longer assisting. It is governing.

Deliberate Friction

Responsible use requires friction. Not usability friction. Cognitive friction. Systems that remove all friction do not empower users. They anaesthetise responsibility.