# AI's Failure Mode Is Careless Adoption Not the Technology AI's worst failures at work come from careless adoption, not from the model — the harm shows up when humans deploy it with little attention to what could go wrong, so the locus of risk is management practice, not capability. ## Scope & conditions - Holds where the failure is mediated by a human decision: how the work is supervised, labeled, or routed. Wiles' experiment is the case — managers caught fewer planted errors when told an "AI employee" (vs. "AI tool" or "a human") produced the work, because the org-chart-peer framing quietly relocated responsibility off the reviewer. The model didn't change; the adoption posture did. - Breaks for genuine capability ceilings — a task the model fails even under careful, well-supervised deployment. There, better adoption practice won't save you and the model is the real lever. - The diagnostic test: hold the work constant and vary how it's adopted (labeled, supervised, trusted). If outcomes move, the failure was in the adoption, not the technology. ## Links - [[3. Resources/Readwise/Articles/Why Instructional Designers Should Learn to Challenge AI]] — instructional design case: fluent suggestions turn an active architect into a passive evaluator unless review pressures the model's assumptions - [[AI Quality Gaps Are Policy Problems Not Capability Problems]] — because extends; the same "not the model" move at a higher altitude — that note locates the fixable gap in the harness/prompt, this one in organizational adoption practice (labeling, supervision, accountability) - [[The Reverse Centaur Pattern - Humans Forced to Work Like Machines]] — because applies-in; the deployment *configuration* (who sets the pace, how much scrutiny survives) decides whether the same tool augments or harms — careless adoption is the harm-producing configuration - [[Safe Agent Changes Come From Guardrails and Observability Not Trust]] — because evidence-for; the operational answer to careless adoption — review AI output by a fixed bar regardless of its "employee/tool" framing, instead of trusting the label - [[3. Resources/Permanent Notes/AI Should Raise Student Challenge, Not Lower Expectations]] — applies-in: the same model can either build scaffolds or replace the learning task, depending on the instructional choice. <!-- atomicity: 1 idea ✓ --> - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: the MOC lists this note under Power & deployment configuration. - [[3. Resources/MOCs/AI Agent Safety MOC]] — applies-in: the MOC lists this note under Human oversight & deployment configuration.