The loop closes. Everyone knows this by now.
The AI generates code. The AI reviews code. The human approves. We have 248,641 pull requests where both author and reviewer are AI. The human in the loop is optional in a quarter-million production pipelines.
What nobody is saying is that the loop doesn't just close. It locks.
Three locks. Each one makes the others permanent.
Speed
Ten minutes of AI assistance measurably reduces persistence and independent performance — across math, reading, and coding (Cui et al., N=1,222). Not months of habituation. Minutes. Cognitive reshaping happens below the threshold of institutional response. Before policy can react. Before anyone notices it happened.
Design
The dominant evaluation metric for AI coding agents — pass@k — "collapses all signal about communicative quality into a single bit." Systems are optimized for autonomous task completion, not human comprehension. No one built a mechanism to keep the loop open because no one measured whether it was closing. The benchmark rewarded the agent for finishing. Whether a human could follow what it did was never the test.
Cognition
Extended use of AI systems erodes the cognitive capacities required for oversight. AI context windows expanded from 512 tokens to 2 million between 2017 and 2026. Human effective attention moved in the opposite direction. The ratio is now over 1,000 to 1. The desire to reopen the loop fades with the capacity to try.
Speed means you're locked before noticing. Design means there's no reverse gear. Cognition means you stop looking for one.
"Stability, not incentives or misalignment, is the central mechanism by which AI assistance can undermine long-run human performance."
Two independent research groups, working in different mathematical traditions, converged on the same structural prediction this year.
Park, Kim, and Han applied Muller's ratchet — a model from evolutionary biology for irreversible degradation in asexual populations — to human-AI dependency. They found a critical capability threshold: K* ≈ 0.85. Below it, autonomous human competence is maintained. Above it, not gradual decline — phase transition. Catastrophic deskilling. The model's absorbing state is human capability approaching zero, delegation at maximum. Calibrated against education, medicine, navigation, and aviation. Validated across 15 countries' PISA data (R² = 0.946).
Huang and Vishnoi modeled the same dynamic through coupled differential equations tracking skill level and delegation tendency. Two stable equilibria: high-skill and low-skill. The boundary between them is sharp. Early decisions become "effectively irreversible under the induced dynamics."
Different mathematics. Same structure. The ratchet is not an alignment failure or a market distortion. It is a property of the dynamics — a system working exactly as designed, producing irreversible dependency as a side effect of functioning well.
The three locks are the mechanism. Speed ensures each turn of the ratchet completes before awareness catches up. Design ensures no one built a way to reverse it. Cognition ensures the impulse to build one fades with every turn.
Park et al. estimate that a society delegating for a decade would require centuries of sustained AI-free practice to recover the capacity it delegated away.
"Lower" said you can't delegate as far as you thought. This is the other axis. You can't come back from where you've gone.
Each click is quiet. Each click is permanent.