The SRE Deskilling Argument Leans on One Colonoscopy Study, Where Detection Fell From 28.4% to 22.4%
Software / explainer
The SRE Deskilling Argument Leans on One Colonoscopy Study, Where Detection Fell From 28.4% to 22.4%
A SigNoz essay circulating again on Hacker News borrows its evidence from medicine and aviation. Here is what that evidence shows, and what it does not.
The claim is that AI will not replace site reliability engineers but will make them worse at the part of the job that matters. It comes from an essay on SigNoz's newsletter, dated Feb. 28, 2026, which was back on Hacker News on Oct. 30 with 509 points and 380 comments. The evidence is real. It is not about SREs.
What the essay argues
The essay is written by an author who signs as Elizabeth, for SigNoz, an observability vendor. It leans on Lisanne Bainbridge's 1983 paper "Ironies of Automation": the more you automate, the more critical the human becomes in the rare moments automation fails.
Its framing question is "When AI handles 95% of your incident response, do you get worse at handling the 5% that actually matters?" The essay itself treats the 95% and 5% as illustrative. They are not measured.
It cites three outside sources: Bainbridge, a Lancet gastroenterology study, and a 2019 aviation paper by Behrend and Lafargue on pilots and autopilot. The aviation paper is cited without figures. The claim that the FAA mandates more manual flying has no link. The PagerDuty three-tier incident model has none either.
The one study with numbers
The medical evidence is the Budzyń et al. study in The Lancet Gastroenterology & Hepatology, published online Aug. 12, 2025. It is retrospective and observational, at four Polish endoscopy centres in the ACCEPT trial. AI polyp detection arrived at the end of 2021. The authors compared non-AI colonoscopies in the three months before and after.
| Measure | Before AI | After AI |
|---|---|---|
| Patients | 795 | 648 |
| Adenoma detection rate, non-AI exams | 28.4% (226) | 22.4% (145) |
| Adenomas per colonoscopy | 1.91 | 1.92 |
Patients had a median age of 61 and 58.7% were female, enrolled between Sept. 8, 2021 and March 9, 2022. In the adjusted analysis, AI exposure carried an odds ratio of 0.69 (95% CI 0.53 to 0.89) for detection. The absolute difference was 6.0 points, with a 95% confidence interval of 1.6 to 10.5. The authors' conclusion is hedged: "Continuous exposure to AI might reduce the ADR of standard non-AI assisted colonoscopy." The essay rounds the figures to 28% and 22%.
- Before AI introduced28.4 %
- After AI introduced22.4 %
Source: Budzyń et al., The Lancet Gastroenterology & Hepatology, 2025, via Warwick WRAP, accessed 2026-10-31
Why it does not transfer cleanly
The study compares two three-month windows. Assignment to AI or non-AI exams was by examination date, and the authors found the drop varied greatly between centres. Reports on the paper also note that adenomas per colonoscopy did not change, and the paper's own interpretation stops at "might."
Nothing here measures incident response. No study cited in the essay tracks engineers' diagnosis time or error rates after months of AI-assisted on-call. The essay is an argument from analogy, and says it was inspired by a founder's LinkedIn post, not a dataset.
Where the analogy still bites
A skill that is exercised less decays. That is the Bainbridge point and it predates language models by forty years. The essay's two proposals are modest: handle some incidents manually on purpose, and design AI so humans stay "in the loop, not on the side."
SigNoz sells observability software and says its goal is "AI that supercharges SREs." That is a disclosed interest, and it makes the proposal about keeping humans engaged convenient. The same worry about practice shows up outside operations: Martin Fowler's case for writing your own prose rests on a measured reader study, which is the kind of evidence the SRE argument still lacks.
Teams adding agent tooling, such as the Superpowers skills framework now installable into 16 coding agents, are in the same position: the tool is measured on what it does, not on what the engineer loses.
What would settle it
A before-and-after on real on-call data, with mean time to resolve for incidents handled without AI. Until a team publishes one, the colonoscopy result is a reason to run drills, not a finding about SREs. Teams adopting AI incident tools can ask a cheaper question now: when did anyone last handle a page without it?
Sources
More in Software
- 01Python 3.15 Ships a Built-In sentinel After Five Years of PEP 661The new type replaces the object() idiom with a one-line constructor that survives copying, pickling and type checking, but a Python 3.15 floor is the price of using it.
- 02Lightpanda 1.0 Passes 80% of Chrome's Web Tests, by Its Own CountThe Zig-built headless browser left beta on Oct. 2, and its own benchmarks put it ahead of two Rust rivals on a 940-page crawl, with no hardware listed.
- 03Claude-Mem Hits 107,900 Stars, and Its Own Pages Disagree on LicenceThe Claude Code memory plugin has gained about 35,500 stars since May, but a May write-up says AGPL-3.0, the repository now says Apache-2.0, and the README describes an optional cloud sync that earlier coverage did not.
- 04Jepsen Found 96 Issues in ParticleDB 1.1.4 and Tells Users to Halt DeploymentKyle Kingsbury’s 31 October report, funded by vendor Orkes, finds lost writes and ignored constraints in healthy single-node runs.