AI is deskilling your SREs before the next major outage
Automating routine incident response is destroying the muscle memory engineers need to fix novel, catastrophic system failures.
Tools · Source: Hacker News
What happened
AI is rapidly taking over routine site reliability engineering. Tools now handle alert noise reduction, log pattern detection, and basic auto-remediation. PagerDuty envisions a future where AI fully resolves known issues and assists with partially understood problems. Humans are left to handle only the most complex, novel outages.
This creates a dangerous paradox first identified in a 1983 cognitive psychology paper called Ironies of Automation. By removing humans from routine tasks, automation strips them of the practice needed to handle rare crises. Engineers lose their system intuition. They stop reading raw logs and stop building mental models of system architecture.
The problem extends beyond losing existing skills. Junior engineers face a crisis of never-skilling. Because AI handles the foundational triage work, new hires never develop the gut-level understanding required to fix unprecedented cascading failures. The training pipeline is fundamentally broken.
Key facts
- 1983 — Year Lisanne Bainbridge published Ironies of Automation
- 28% to 22% — Drop in unassisted polyp detection by endoscopists after using AI
- Tier 3 — PagerDuty classification for novel failures that remain human-led
- 95% — Estimated percentage of incident response AI could handle
Why it matters
Founders building AI tools for developers need to rethink their product interfaces. If your tool simply gives the user an answer to approve, you are turning them into a passive supervisor. This leads to vigilance decay. Builders must design systems where humans actively engage with the problem alongside the AI, keeping their cognitive skills sharp.
The second-order effect is a fragile engineering culture that collapses under pressure. When a novel incident bypasses AI pattern matching, your team will freeze. Dashboards will look green until a catastrophic Wednesday outage reveals that your senior engineers have rusted and your junior engineers never learned how to debug from scratch.
For builders
Design for deliberate inefficiency
Force engineers to manually resolve a percentage of routine incidents. Companies that optimize purely for speed will pay the price during catastrophic downtime. You lose short-term velocity but buy long-term resilience.
Build collaborative AI workflows
Stop building AI that just hands users a final answer to rubber-stamp. Build tools that require the engineer to dissect the problem alongside the AI. Tool makers who get this right will win enterprise contracts from risk-averse buyers.
Own the underlying data layer
API-based AI wrappers lack deep system context. Products that sit directly on the metrics, traces, and logs can build more accurate capabilities. Founders who control the data layer will crush thin AI wrappers.
My take
I see founders bragging about how their AI agents completely remove humans from the loop. They are building a ticking time bomb. If you abstract away the struggle of debugging, you destroy the engineer. I build AI to accelerate builders, not to lobotomize them.
Original reporting: Hacker News. This is my rewrite and opinion.