The growing use of artificial intelligence in site reliability engineering is shifting attention from whether software will replace engineers to whether it could weaken the skills they need during unusual failures. An October 30 analysis published by the SigNoz newsletter argues that the greatest risk may emerge when automation handles routine incident work but humans remain responsible for the hardest exceptions.
The argument draws on a longstanding problem in human-factors research: systems can make operators more important at the moment they also give them less practice. The article cites psychologist Lisanne Bainbridge's 1983 work on the ironies of automation, which examined how automated industrial processes left people monitoring stable operations for long periods and then intervening during rare, demanding breakdowns.
Applied to reliability engineering, the concern is straightforward. Tools already help group alerts, detect patterns in logs, execute known runbooks, suggest causes from previous incidents and remediate familiar failures such as restarting a crashed service. More advanced systems aim to correlate metrics, logs and traces, assess change risk, predict incidents and initiate rollbacks. Each capability can reduce repetitive work and shorten response times.
But the remaining human work becomes disproportionately novel. Engineers may be called in for cascading failures, ambiguous signals or incidents that do not match historical patterns. Those situations depend on mental models developed by reading raw telemetry, forming hypotheses, tracing dependencies and repeatedly resolving smaller failures. If AI routinely summarizes evidence and proposes causes, engineers may exercise those skills less often.
The analysis distinguishes deskilling from what it calls never-skilling. Experienced staff may lose fluency after delegating tasks, while junior engineers may never acquire it if automated assistance is present throughout their training. That creates an organizational challenge because visible productivity can improve even as unaided capability becomes harder to measure. A green dashboard or rapid resolution of known incidents does not prove that a team can manage an unprecedented outage.
The piece is an analysis rather than a controlled study of SRE teams, and it does not establish that current AI tools have already caused widespread reliability failures. Its comparisons with medicine and aviation are cautionary analogies, not direct measurements of software operations. The practical implication is therefore about risk management rather than abandoning automation.
Teams can preserve human capability by designing deliberate practice into operations: rotating engineers through diagnosis before revealing automated recommendations, reviewing raw evidence during post-incident analysis, running failure simulations and ensuring junior staff understand system architecture. Automation can still handle repetitive work, but organizations need ways to test whether their people can operate when the recommendations are incomplete or wrong.
