AI isn't replacing SREs, it's deskilling them: where does the intuition for the critical 5% come from?
Oct 31, 2026Original SigNoz 뉴스레터 · AI Isn't Replacing SREs. It's Deskilling Them. ↗
In 3 lines
- The more is automated, the more humans only face rare big outages and lose the chance to build intuition.
- Endoscopists' unassisted detection fell from about 28% to 22% after using AI.
- The fix is keeping practice incidents for humans on purpose. A calculator-style rebuttal exists too.
"When AI handles 95% of your incident response, do you get worse at handling the 5% that actually matters?" That is the question SigNoz's newsletter posed on Oct. 30, reaching 549 points and 410 comments on Hacker News. This piece summarizes the post and the pushback from practitioners.
When AI handles 95%, who solves the other 5%?
The SigNoz newsletter asks not whether AI takes the job but what happens to human skill. Alert grouping, runbook execution, log anomaly detection and pod restarts are already automated or AI-assisted. The author notes that the more is automated, the more humans are called only for never-before-seen outages. The intuition to solve those comes from years of hands-on incident response, and those chances are shrinking.
A 1983 paper predicted this
The post invokes psychologist Lisanne Bainbridge's 1983 paper "Ironies of Automation." The more you automate, the more critical the operator becomes when automation fails, yet the less practice they get at exactly that intervention. People are left with long passive monitoring and rare crises. The author sees the same pattern repeating in knowledge work.
Lisanne Bainbridge · Automatica (1983) ↗The endoscopy study: 28% to 22%
The medical example is AI-assisted colonoscopy. After a period of AI use, endoscopists' unassisted polyp detection rate fell from about 28% to 22%. They had not forgotten theory; they stopped exercising the skill. The study appeared in The Lancet Gastroenterology & Hepatology.
The Lancet Gastroenterology & Hepatology · thelancet.com ↗The scariest part: you don't feel it
The author admits AI became the first thing they reach for and that they can feel the rust. Raw logs go unread, hypotheses go unformed during incidents, and architecture models go unbuilt. Each looks like a productivity win, but together they hollow out expertise. Because tasks feel easier, people struggle to judge their own skill accurately.
Juniors can't lose what they never built
The post separates deskilling (losing a skill you had) from never-skilling (never developing it). Junior SREs entering an AI-handled workflow get few repetitions to build intuition. And those who sense their skills slipping lean on AI more, which degrades them further.
The fix: keep some deliberate inefficiency
The proposal is not to drop AI. Like the FAA mandating manual flying time, teams can reserve some AI-solvable incidents as human practice and design for people working the problem alongside AI, not just approving its answers. Hacker News commenters added practices such as having the agent quiz you before merging, or a weekly hour of hand-coding.
Elizabeth (SigNoz) · Hacker News 댓글 · newsletter.signoz.io ↗The rebuttal: calculators deskilled long division too
Hacker News was split. One commenter said calculators killed long division yet the real task is taxes and reports, so AI lets us broaden. Another said Terraform deskilled hand-configured infra and nobody wants it back, so use AI to catch incidents faster. Deciding which skills to protect and which to hand off is the open homework.
Hacker News 이용자 cadamsdotcom, jedberg · news.ycombinator.com ↗Where do future seniors come from?
One comment stood out: spotting AI errors takes experience, seniors are made through junior-years training and mentoring, and that pipeline is narrowing. A startup developer who hands most AWS work to Claude said the efficiency is overwhelming, but skill atrophy is real and job interviews now feel scary. Designing time to do things yourself looks like a new career task.
Hacker News 이용자 jmyeet, sloob · news.ycombinator.com ↗
