Governance of Medium replaces an unobservable authorship binary with an inspectable practice axis.
The paper argues that “form mediation versus content mediation” cannot reliably diagnose AI-assisted production. Translation, tone, and reference selection all alter what a work means, while any attempt to exhibit content prior to expression produces another expression.
Governance can be inspected. Five dimensions enumerate it: direction, selection, iteration, constraint, and review. High-governance practice specifies goals that outputs can fail, rejects and revises against stated reasons, imposes nondefault constraints, and gives the bearer authority to reject the final work. Low-governance practice accepts default output with little shaping.
Governance is not bearing. A highly directed practice may be closed to correction, and a low-governance output may carry upstream bearing laundered through the model. Governance is instead a practical proxy at the production-record layer.
The vocabulary also serves as an entry point for practitioners whose institutions discourage explicit AI-provenance language. “High-governance LLM assistance” can describe actual practice more truthfully than either purity or total-generation categories, while still leaving fuller provenance work possible.
The paper proposes testing whether governance correlates with distributional distance from a model’s default output. Agreement would produce a two-layer instrument; disagreement would identify forms of direction that leave little stylistic trace.