Emergency Intervention and Verification in AI-Dependent Infrastructure
Year: 2026
Type: preprint
Abstract: Stopping an AI deployment does not necessarily revoke the authority it has acquired through other systems. This provisional preprint proposes a bounded assessment procedure for evaluating whether two emergency intervention paths remain effective under specified failures or compromises, distinguishing shared components from shared defeat conditions and separating successful action from independently verifiable evidence of the protective outcome. It yields a scoped pass, failure or inconclusive result rather than a claim of statistical independence or universal controllability. A short assessment of electromagnetic-pulse proposals and a U.S.-only legal survey explain the conditions under which the test can be used. Status: this is a provisional proposed method. It has not been peer reviewed, has not been tested by independent assessors, and has not been validated in real facilities. The worked example (a greenhouse-control laboratory, with all timings, observations and assessment records) is entirely fictional and illustrative; it is not evidence from an actual facility. The paper builds on the intervention requirement discussed in Li (2026), Unkillable (https://doi.org/10.5281/zenodo.23223075), and on the observation-window concept in Li (2026), Which Failure Do You Fear More? (https://doi.org/10.5281/zenodo.22821194). AI assistants helped with drafting, editing, citation checking and formatting; the author directed and reviewed all content and is responsible for it.
Author: Li, Ana
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235147494
