In 1983, cognitive psychologist Lisanne Bainbridge published "Ironies of Automation," one of the most cited papers in human factors research. Its core finding was paradoxical: the more you automate a process, the more critical the human operator becomes during the rare moments automation fails — and the less prepared they are to intervene. Four decades later, the same dynamic is playing out in site reliability engineering, where AI agents now handle alert grouping, runbook execution, log anomaly detection, and auto-remediation of known failure modes. The industry has settled into a tiered model. PagerDuty frames it explicitly: Tier 1 incidents (known issues, known fixes) get fully automated. Tier 2 (partially understood problems) receive AI recommendations with human sign-off. Tier 3 (novel, cascading failures) stay human-led. The logic sounds clean. The problem is that Tier 3 expertise doesn't materialize from thin air — it's built through years of hands-on Tier 1 and Tier 2 work, the repetitive pattern recognition that develops gut-level system intuition. The evidence from adjacent fields is consistent and uncomfortable. A study of endoscopists found that AI-assisted polyp detection caused unassisted detection rates to drop from 28% to 22% after a period of AI use. In aviation, autopilot dependence degraded situational awareness enough that the FAA now mandates manual flying time as a countermeasure. The mechanism is identical in each case: offloading routine cognitive tasks atrophies the skills those tasks were quietly maintaining. Beyond deskilling — losing capabilities you once had — researchers have identified a compounding problem they call "never-skilling." Junior SREs entering the field today inherit environments where AI handles most incident response workflows. The opportunities to build foundational intuition through hands-on practice are vanishing before they ever materialize. The training pipeline itself is broken, and the degradation is self-reinforcing: SREs notice their skills eroding, lean harder on AI to compensate, which accelerates the erosion. The most insidious feature of this cycle is its invisibility. Cognitive research suggests that because AI tools make tasks feel easier and enhance visible performance metrics, users cannot accurately judge their own skill degradation. Dashboards stay green. Response times improve. Then a novel incident hits that doesn't match any pattern the AI has seen, and the human in the loop discovers the muscle has atrophied. You stop reading raw log streams because the AI summarizes them, stop forming hypotheses because the AI suggests root causes, stop building mental models of architecture because the AI maps dependencies. Each looks like a productivity win. Collectively, they hollow out the expertise that matters most. The piece, authored by SigNoz (an open-source observability platform), proposes two countermeasures: deliberate inefficiency — designating certain incidents as human-practice opportunities even when AI could handle them, mirroring the FAA's manual flying mandates — and designing for human-in-the-loop rather than human-on-the-side, keeping engineers cognitively engaged rather than passively supervisory. Both are sound in principle. Neither addresses the structural incentive problem: companies optimizing for mean-time-to-resolve have no mechanism to value latent human capability that only manifests during black swan events. This is ultimately a piece of thought leadership from a vendor with skin in the game — SigNoz is building AI features into its observability platform and has a commercial interest in the "AI that supercharges SREs" framing rather than the "AI that replaces SREs" one. That doesn't invalidate the argument, which is well-grounded in Bainbridge's research and supported by cross-domain evidence. But the reader should note that the proposed solution conveniently lands on "buy an observability platform that keeps humans in the loop" — which is exactly what SigNoz sells.
