AI coding tools generate larger changesets (batch sizes) than human developers, and DORA research confirms that larger batch sizes correlate with higher failure rates. The industry-wide decline in software reliability is a direct, measurable consequence of AI-assisted development patterns. ## The Mechanism 1. **AI generates more code per change** -- agents produce larger diffs, touching more files per commit 2. **Larger batch size → higher failure rate** -- established DORA research finding, not speculation 3. **Review capacity doesn't scale** -- humans can't meaningfully review 3,000-line AI diffs the same way they review 300-line human diffs 4. **Reliability declines industry-wide** -- aggregate effect is measurable: failure rates are up ## Why This Matters This is empirical evidence for what the [[Joy Theft Paradox]] describes qualitatively: AI takes over the building, but the *verification burden* on humans increases faster than the productivity gain. The system produces more output but less reliable output. ## Connection to Existing Concepts - Validates [[Agency Preservation Standard]] -- understanding every line matters more when batch sizes are large - Explains why [[Constraint-Based Agent Governance]] is necessary -- you can't review large batches manually, so you need automated guardrails - Inverts the [[AI Productivity Paradox]] -- productivity rises while *quality* falls, creating a deferred cost - Reinforces [[Developer as Builder Not Orchestrator]] -- Mateus Morais's concern about unmaintainable AI code is now backed by DORA data ## Boundary Conditions - Teams with strong CI/CD and test coverage may absorb larger batch sizes without reliability loss - The tradeoff is not inevitable -- it's a governance failure, not a physics constraint - Smaller, well-prompted AI changes (constraining batch size) could preserve reliability ## Related Concepts - [[Joy Theft Paradox]] -- Qualitative description of the same structural shift - [[Constraint-Based Agent Governance]] -- The solution: guardrails over review - [[Vibe Coding Paralysis]] -- Another pathology of unconstrained AI-assisted development - [[Agency Preservation Standard]] -- Never commit code without full comprehension ## Source - [["Software reliability is down across the industry; failure rate is....md]] -- Gergely Orosz citing @rhein_wein + Nathan Harvey, February 2026 --- *Atomic concept extracted February 2026*