Regulators weigh emergency halt powers against internal audit trails

Legal scholars Yonathan Arbel, Simon Goldstein, and Peter N. Salib propose a "halt right" for the US government to pause frontier AI training or deployment when catastrophic risks emerge (AI Frontiers). This mechanism offers a targeted emergency brake rather than full nationalization, avoiding equity stakes or operational control while redirecting investment toward safety evidence (AI Frontiers). Proponents argue this creates a necessary safety threshold without slowing broader American development, as the halt applies only to specific risky models.

Conversely, experts at Pacific AI argue that accountability requires rigorous, internal operational excellence that remains auditable in real-time. By linking each system version to its parent and recording every human or system decision, organizations can answer specific auditor questions about who decided what and when (AI in Healthcare). This approach treats AI risk as a version-controlled process, using deterministic code to enforce rules while the model handles judgment, which ensures that every change remains transparent and verifiable.

These positions highlight a fundamental tension between external emergency interventions and proactive, granular governance. While the halt right provides a final safeguard against systemic threats, the audit trail approach addresses the immediate challenge of accountability in live production systems where agents already operate with broad technical credentials.

• The UAE aims to shift half of its federal services and operations to AI agents within two years, necessitating new identity and accountability frameworks (Joe Hawkins from Prophecy Watchers).
• Pacific AI benchmarked four frontier models against 11 US laws and found none scored above 0.64 in handling adversarial cases correctly (AI in Healthcare).