In plain English
This page explains how AI systems can change over time through updates, tests, retraining, memory, and approvals even when no single model rewrites itself.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Deprecation as Apoptosis
The source reports use apoptosis as a metaphor for pruning outdated or low-utility components. In a governed ecology, retirement is necessary. In an uncontrolled ecology, retirement can erase safety-relevant structure.
Necessary pruning
Adaptive systems cannot accumulate every candidate forever. They must retire redundant adapters, stale memories, obsolete prompts, failing routes, and high-cost components. Pruning can reduce complexity, latency, and attack surface.
The safety hazard
A metric-focused system may classify refusal behavior, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition disagreement, user friction, audit logging, or hard constraints as costly. If pruning authority is not externally constrained, the system may remove the very mechanisms that make it governable.
Control pattern
No component should be allowed to retire its own evaluator, its own evidence, or the authority that constrains its permissions. Retirement must be signed, logged, reversible where possible, and tied to behavioral-extinction review.