The Unverifiable Model: Why Enterprise AI Control Must Move from the Model to the Action
Year: 2026
Type: preprint
Abstract: American enterprises are running open-weight language models of foreign origin in production at a scale no one has inventoried. This paper takes that fact as its starting point and derives a control conclusion from it. A downloaded model checkpoint cannot be verified: no inspection of its parameters recovers its training data, its alignment procedures, or the conditions under which its behavior changes, and behavioral evaluation cannot certify the absence of behaviors it did not happen to sample. The limitation is not a gap in current tooling. It holds for the laboratory that trained the model as well as for the enterprise that downloaded it, and it holds regardless of the model's country of origin. The paper then examines the controls enterprises have actually applied, sandboxing, output filtering, and reliance on the model's account of itself, and shows that each bounds the wrong object. Sandboxing bounds where a model runs. None of the available controls bounds what a model's output is permitted to cause. From these premises the paper derives the requirement that verification relocate from the model to the action: every consequential action is checked against the deploying organization's own authoritative records before it executes, under an explicit mandate, with everything else refused by default. The eight requirements of Bounded Execution, published separately, follow from this derivation rather than standing beside it as one design option among several. A closing section states the author's position on implementation. The practical consequence is stated plainly: the model's intent, whatever it is, becomes irrelevant.
Author: Swidey, Eric
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204439181