The Law of Stop: Interruptibility, Injunctions, and the Governance of Agentic AI
cs.CY · cs.AI
Quick verdict
Ambitious legal theory of AI interruption: 1,213 coded incidents, 80% unstopped, gaps mostly legal not technical; framework strong, coder reliability modest.
Summary
This Article argues that interrupting AI systems is a legal-institutional practice rather than a technical artifact. Perez (Bar-Ilan) disaggregates 'stop' into four dimensions — technical affordances, interruption authority, epistemic triggers, epistemic standing — and four paradigms of rising difficulty: the escalator's simple stop, the process plant's sequenced shutdown, the railway's networked stop, and agentic AI, where control is divided and the system may resist. Part II reconstructs the legal repertoire through case studies (escalators, OSHA process rules, railway rule books, FAA groundings, platform litigation). Part III reports an original empirical study: roughly 1,400 AI Incident Database incidents coded independently by two LLMs from rival labs under a pre-registered protocol; of 1,213 retained, about 80% (972) record no stop, and where no usable stop existed the missing element was legal rather than technical in 506 of 632 cases (about four in five), though a coding convention assigns 236 of those legal-gap cases to content harms. A survey of 39 governance instruments finds only 7 with binding stop requirements, none specifying coordination, preservation, or resumption. Part V proposes a layered law of stop: infrastructure-layer emergency authority, enforceable evidentiary access, and redundancy. Main limitations: modest inter-coder agreement (kappa 0.32–0.47), a ~20% override rate in a stratified audit, and corpus bias toward visible, non-agentic content harms.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Stopping an AI system is a legal-institutional practice defined by four questions — technical affordances, interruption authority, epistemic triggers, epistemic standing — not a single 'kill switch'.
- Agentic AI breaks the existing law of stop: control is divided across actors, a halt at one point leaves the activity running elsewhere, and the system may resist or corrupt the evidence for its own halt.
- In the coded record, roughly 80% of 1,213 AI incidents involved no stop, and where no usable stop existed the missing element was legal/institutional rather than technical in about four of five cases (506 vs 126).
- Of 39 AI governance instruments, only seven contain binding stop requirements and none specifies coordination, preservation of state, or resumption; no US AI-specific instrument confers halt power over a running system.
- Law should adopt a layered architecture of interruptibility: infrastructure-layer emergency authority, enforceable evidentiary access for supervisors and evaluators, and redundancy for when a stop fails.
Strengths
Original conceptual framework
The four-dimension decomposition (technical affordances, interruption authority, epistemic triggers, epistemic standing; Part I.A) crossed with four paradigms (Part I.E, Table 1) relocates interruptibility from engineering to institutional design — a framing the author shows is absent from the ~130-source human-oversight literature he reviews (Part I.C, fn. 15).
Unusually rigorous design for empirical legal scholarship
Part III pre-registers the coding instrument, hypotheses H1-H4, and six resolution rules under embargo on OSF (fn. 141, July 1, 2026); two LLMs from rival labs (Claude, Codex) code blind; ambiguity is resolved against the author's own thesis (fn. 150); sensitivity checks include the agreed-cases-only subset and excluding the content-harm convention (270 vs 126).
Primary-source legal history with real doctrinal payoff
Honest limitations section
Systematic instrument mapping
Weaknesses
Coder reliability is weak exactly where the headline claim needs it
Correlated LLM-coder error demonstrated
Empirical base diverges from the theoretical target
Speculative counterfactual layer
Load-bearing motivating facts are unsettled
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
The Law of Stop — Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of 'The Law of Stop' paper?
In roughly 80% of 1,213 coded AI harm incidents, no one stopped the system; where no usable stop existed, the missing element was legal or institutional rather than technical in about four of five cases (506 vs 126). The paper argues that stopping AI is a legal-institutional design problem — who may halt, on what evidence, with whose knowledge — not merely an engineering kill switch.
How did the author code 1,400 AI incidents?
Two large language models from rival laboratories (Anthropic's Claude and OpenAI's Codex) independently coded every incident against a fixed instrument pre-registered on OSF under embargo. Disagreements were resolved by six pre-specified rules with the author adjudicating, and a deliberately stratified sample of 24 agreed cases was audited, of which 5 were corrected.
Do any AI laws currently require a kill switch?
Of 39 governance instruments surveyed, only seven contain binding stop requirements (including the EU AI Act, Korea's two Framework Acts, China's 2023 and 2026 measures, Vietnam's AI law, and a California transparency act). Only three reach a system in operation, and none specifies who may halt a system, on what evidence, what state must be preserved, or how operation may resume.
Is the paper's empirical claim reliable?
The design is unusually transparent for legal scholarship (pre-registration, dual blind coders, conservative resolution conventions, robustness checks), but inter-coder agreement was modest (Cohen's kappa 0.32-0.47), the legal-vs-technical subtype had no agreement statistic, and about 20% of audited agreed cases were corrected — so the headline ratios are best read as carefully bounded lower bounds rather than precise estimates.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extracted text is truncated mid-Part IV: Part V, the Conclusion, and the Technical Supplement (full reliability statistics and the coding instrument) are absent from the provided text.
- Figures 2, 3 and 5 and Tables 1-2 are only partially recoverable from prose; several breakdowns were taken from footnotes rather than the graphics.
- Portions of Parts II.D-E and III (including the Section III.D autonomy analysis) were omitted in extraction; those findings are characterized from the Introduction, Section III.E, and cross-references.
- The paper's motivating 2026 episodes (Mythos directive, OpenAI-Hugging Face incident) could not be independently verified from the text; the directive itself is described as unpublished.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.