Alignment Community Debates Whether Dynamic Safety Equals AI Alignment

From September 21–23, the alignment community saw a multi-round debate between tdietterich, davidmanheim and others over whether Nancy Leveson's "dynamic safety" concept is essentially the same as AI alignment, and whether scaling is the key alignment challenge. No consensus was reached, but the exchange directly bears on how alignment research should be positioned.

Confirmed

On September 21, tdietterich posted a long thread drawing on Leveson's Engineering a Safer World, arguing that safety (today's "alignment") is not an inherent property of automated systems like self-driving cars or AI agents, but a dynamic process requiring active maintenance.
He defined a safe system as one that keeps hazards to people and infrastructure below socially acceptable levels; dynamic safety requires maintaining this even in the face of failures.
He argued that failure modes of system components and changes in the operating environment exhibit unbounded diversity, with the exception of simple systems performing very narrow tasks in unchanging environments.
tdietterich claimed Leveson's dynamic safety is essentially the same as alignment, and speculated that Leveson and the human-factors community do not believe machines can autonomously achieve dynamic safety.
davidmanheim offered two key distinctions: first, dynamic safety ≠ alignment—dynamic safety only requires safety within a given operating range, while alignment concerns behavior as systems scale, like a motor safe at normal temperature and speed but not necessarily at two or three times those; second, frontier systems have scaled to the point where they can hardly be responsibly overseen.

Why it matters

The debate touches the positioning of alignment research: if alignment equals classic dynamic safety engineering, mature system-safety methodology can be borrowed; if, as davidmanheim argues, the core challenge lies in supervising behavior beyond scaling, a new research framework is needed.

2026-09-21 ~ 2026-09-23 · 6 related posts

Primary sources

AI safety veteran: alignment isn't a property of the model but a dynamic state — tdietterich · 2026-09-21 03:27 (UTC+08:00)
AI Safety Researcher: Frontier Systems Are Already Nearly Impossible to Oversee Responsibly at Scale — davidmanheim · 2026-09-23 01:59 (UTC+08:00)
Dietterich: Leveson's dynamic safety is essentially alignment — tdietterich · 2026-09-23 01:36 (UTC+08:00)

Dietterich: failure modes and environment changes show unbounded variety — tdietterich · 2026-09-21
[source] AI safety veteran: alignment isn't a property of the model but a dynamic state — tdietterich · 2026-09-21
Tom Dietterich defines safe systems: harms must stay below socially acceptable levels despite disturbances — tdietterich · 2026-09-23
[source] Dietterich: Leveson's dynamic safety is essentially alignment — tdietterich · 2026-09-23
Alignment debate: is Leveson's dynamic safety the same as AI alignment? — davidmanheim · 2026-09-23
[source] AI Safety Researcher: Frontier Systems Are Already Nearly Impossible to Oversee Responsibly at Scale — davidmanheim · 2026-09-23