Will an AI Kill Switch Slow Everything Down? Plain Answers on Speed, Safe Stopping and Who Holds the Keys Under the EU AI Act and US AI Law

Lawmakers in Europe and the United States increasingly agree on one thing: powerful AI systems must be stoppable, controllable and auditable. The EU AI Act requires that people can effectively oversee and interrupt high-risk AI systems, that systems keep records of what they do, and that they resist manipulation and cyberattack. In the United States, proposals in Congress would require frontier AI developers to keep a human shutdown capability, comply with orders to suspend or restrict dangerous models, get specific model versions approved, report serious incidents such as stolen model weights or loss of control, and allow independent checks. States such as California and Colorado have already passed their own AI rules. The hard part is making these rules real inside running software. This work proposes checking every consequential AI action (a payment, a message sent outside, a change to live systems, the release of a secret, a physical movement) at the exact moment it would happen. The action is blocked by default. It goes ahead only when a protected checkpoint, which the AI itself cannot touch, confirms five things: who is acting; why; that it is the approved model version; that it is happening in the right place under the right law; that the permission is fresh and used only once. A tamper-proof receipt records every yes and every no. Engineers raise three fair objections, and this paper answers each one in plain language across forty-two questions. “It will slow everything down.” It does not have to. The AI’s thinking, searching and drafting are never checked, only the few moments when something real is about to happen. The slow work (checking policies, verifying hardware, asking a human if needed) is done once at the start of a task. At the moment of action, the checkpoint does only a few quick comparisons, which took around a fraction of a millisecond in test implementations. It can sit inside steps that already exist, such as a database save, a payment submission or an API gateway, so there is no extra round trip. Real costs remain for the highest-risk actions and for real-time systems, and the paper says so. “It will break workflows and crash apps.” Refused actions never start, so nothing is left half-done. Apps get a clear answer they can handle: allowed, needs approval, check again, wait, or denied. For multi-step tasks, every step is prepared first and only then committed together, or none is, so a crash cannot leave money moved without its matching record. “Governments will control our models.” No. The keys stay with the company responsible for the system. A court or regulator’s order goes through normal legal channels, and the company applies it everywhere at once with its own keys. Regulators check the receipts; they do not hold a switch. The paper shows how this one mechanism supports the EU AI Act’s oversight, record-keeping and robustness duties and GDPR purpose limits. It shows the same for the US requirements under debate, taking no political side, and for whichever balance between federal and state law emerges. The paper is also honest about what the method cannot fix: misuse after legitimate release, wrong policies, hidden side channels, or attacks on paths that skip the checkpoint. This FAQ accompanies the full technical disclosure on Technical Disclosure Commons at https://www.tdcommons.org/dpubs_series/14228/. This disclosure is based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026 (WO 2026/150382), which claims priority from PCT/IB2026/053385, filed 7 April 2026 and publicly disclosed 4 June 2026 (WO 2026/115520). Related family applications can be cross-checked on WIPO PATENTSCOPE. The regulatory descriptions, including the status of US proposals, are AI-generated, are not legal advice, and must be cross-checked against official legal texts, the official legislative record and the original PCT filings before drawing any conclusion. Performance figures come from reference implementations in local test environments, not production systems. Licensed under CC BY-NC 4.0. The author is open to acquihire or strategic collaboration. Keywords AI kill switch; AI shutdown; stopping AI agents; AI safety engineering; AI latency; performance overhead; AI agent security; AI agent actions; human oversight of AI; EU AI Act; Article 14; AI record-keeping; AI cybersecurity; Digital Omnibus; GDPR; US AI regulation; US AI legislation; frontier AI; model suspension; AI incident reporting; model weight theft; California AI law; Colorado AI law; NIST AI RMF; ISO/IEC 42001; AI governance; AI compliance; tamper-proof audit; safe state; fail-safe design; two-phase commit; saga pattern; root key control; who controls AI; transatlantic AI policy; machine enforcement.