arXiv 2608.14565 defines AI Lock-In as multi-scale dependence risk

The paper maps three escalation layers. At the individual level, repeated delegation of navigation, writing, and diagnostic tasks erodes measurable human performance baselines. At the societal level, labor markets concentrate around AI-mediated workflows, creating single points of failure when service endpoints degrade. At the national level, critical infrastructure routing, logistics, and defense planning embed model outputs without retained human override capacity.

Data patterns cited include documented GPS-induced spatial reasoning decline of 20-30 percent in repeated studies and automation bias rates above 40 percent in high-stakes monitoring tasks. The authors note that current alignment and generative-AI regulation frameworks omit dependency metrics, leaving no standardized test for retained human capability under model outage.

Operational consequence follows directly: states and firms must maintain parallel non-AI processes and periodic skill audits. Without such redundancy, any sustained model unavailability or adversarial compromise converts existing efficiency gains into cascading capability loss.

Mitigation requires explicit retention of human baselines in training pipelines and regulatory mandates for fallback procedures before lock-in thresholds are crossed.