The author says testing frontier AI models (OpenAI’s call for U.S. rules, EU cybersecurity tests) only covers the first risk. The bigger gap is decisions that stay in force after the facts behind them go stale—credit, borders, sanctions, aid, supplier bans.

Shared systems spread that problem: one agency’s provisional flag becomes another’s input without an expiry date, so a correction at the source may never catch the copies.

Proposed fix: a “decision-validity horizon”—record the key premises, give each decision a time or condition limit, revalidate or lapse when it expires, and send machine-readable expiry/correction with any cross-border share. Existing AI laws log and explain; they still do not say until when a decision remains authoritative.

Governments are learning to test advanced AI systems. They also need a rule for retiring the decisions those systems leave behind.

This week, the politics of artificial intelligence moved closer to a familiar regulatory question: what should a system prove before it is trusted? OpenAI urged the US Congress to adopt mandatory national safety requirements, including testing, independent assessment, cybersecurity protections and incident reporting. Across the Atlantic, the European Union’s cybersecurity agency began testing two advanced frontier models. The debate is moving, correctly, from voluntary assurances toward evidence and oversight.

Yet model testing addresses only the first moment of risk. A second arrives after an output has been converted into an operational decision, copied into institutional records and left to govern a person, company, transaction or border crossing. The model may not change. The evidence may. A decision that was reasonable on Monday can become indefensible by Friday while every database still reports it as valid.

This is not simply an accuracy problem. It is a problem of time and authority. Global AI governance needs a rule that consequential decisions cannot remain in force longer than the premises that justified them.

Every consequential automated decision runs on two clocks. One measures how long the decision remains active. The other measures how long its supporting facts can reasonably be expected to remain true. When the first clock outlasts the second, yesterday’s evidence continues to exercise today’s authority.

Call the second clock the decision-validity horizon. It is the time or condition window in which the premises behind a decision can still support its continuing effect. A credit decision may depend on income and interest rates. A travel-risk assessment may depend on a changing identity match or threat picture. A supplier exclusion may depend on ownership, control or sanctions status. An aid-eligibility determination may depend on displacement, income or access conditions.........

© Eurasia Review