The productivity numbers look great right now. The capability trajectory underneath them is a different conversation.
1. Your On-Call Team Is Losing Reps — And Not Noticing #
When AI starts handling routine incidents, something counterintuitive happens: the team gets slower at the unusual ones.
This is the Bainbridge paradox, documented in 1983 by psychologist Lisanne Bainbridge in the context of industrial automation. Her core finding: automation reduces the opportunities for operators to practice the skills they'll need in abnormal situations, precisely while leaving them responsible for those situations when they arise. Aviation discovered this the hard way. Nuclear operations too. Software engineering is next.
AI incident response tooling is genuinely effective on MTTR for known failure patterns. Engineers triage faster, mean time to resolution improves, dashboards look better. But engineers who aren't regularly working incidents without AI assist are accumulating what practitioners now call "comprehension debt" — a growing gap between system complexity and the team's actual ability to reason about it under pressure. The concerning part isn't that AI is handling incidents. It's that the practice time required to maintain sharp incident response instincts isn't being replaced by anything. Like a pilot who never practices a failed approach, the skills don't disappear overnight. They erode quietly.
The aviation parallel is instructive: pilots train in simulators for rare emergencies every six months whether they want to or not. Engineering organizations will need to build the equivalent — not as a nice-to-have, but as capability maintenance.
Why it matters:
For ICs: If you haven't manually walked a novel cascade in months, your instincts are stale — and you won't know it until you need them.
For leaders: MTTR trending down while system comprehension quietly erodes is a real risk scenario; those two metrics don't conflict until the incident AI can't pattern-match.
For founders: Incident simulation as a practice discipline is underinvested relative to where AI-assisted ops tooling is heading.
Routine automation always looks like pure gain until a novel failure arrives. That's the paradox.
2. The Revolt of the Reader Is Already Underway #
A quieter form of skill erosion is playing out in written communication — and it has measurable consequences.
Engineering blogs, RFCs, postmortems, design docs: these are the places where technical reasoning gets communicated, debated, and validated. They're also being flooded with AI-generated prose. Recent survey data found that 78% of technical readers stop reading when they detect AI involvement in a piece, and 71% subsequently avoid that author entirely. Detection isn't hard — structural tells like "and here's why that framing matters!" have become reliable signals. Once spotted, readers disengage not just from the piece but from the source.
This matters to engineering organizations because written communication is how technical credibility compounds. The engineer who writes clearly and specifically builds influence over time — influence that shapes how ideas get adopted, how architectural decisions land, how careers advance. If AI-generated writing triggers systematic avoidance among the people whose attention matters most, then leaning on AI for external communication means trading short-term effort savings for long-term credibility.
The paradox is clean: writing is a skill that improves only through practice, and off it to AI removes the practice while producing output that feels good enough. The feedback loop that would correct this — readers pushing back — is instead producing silent disengagement and author avoidance.
Why it matters:
For ICs: Your written voice is your professional surface area. AI-generating your design docs and blog posts is burning it slowly.
For leaders: AI-assisted communications policy needs to account for trust cost, not just time cost.
For founders: Authentic institutional voice is a moat. Once degraded, it's slow to rebuild.
3. At Population Scale, This Becomes Structural #
The Bainbridge paradox operates at the individual and team level. New research asks what happens when you zoom out further.
Researchers from the Santa Fe Institute and collaborators have modeled LLM adoption through an epidemiological lens — tracking how cognitive dependence spreads through populations and identifying tipping points where adoption dynamics become self-reinforcing. Their finding: under certain adoption-rate conditions, populations can shift toward persistent dependence with abrupt losses in cognitive competence. The model also identifies conditions for "cognitive immunization" — maintaining reversibility and limiting transmission rates before tipping points are crossed.
The epidemiological framing is a modeling choice, not a moral judgment. What makes it worth taking seriously is the mechanism it surfaces: that cognitive skills, like operational ones, require regular exercise to maintain, and that once a population has broadly outsourced a capability, the conditions for recovering it are worse than the ones that allowed it to erode. This isn't a speculative concern — it's the same dynamic that explains why early calculator adoption in education required rethinking how arithmetic fluency was maintained.
The engineering implication is concrete: organizations that systematically replace human judgment with AI outputs — in incident response, code review, written communication — are making bets about which skills they'll be able to rebuild when conditions change. Some of those bets will prove correct.
Why it matters:
For ICs: Choosing which skills to maintain independently is now a career decision, not just a productivity optimization.
For leaders: AI adoption policies need capability maintenance metrics alongside output metrics — you can't manage what you're not measuring.
For founders: The capabilities your team outsources today become the constraints on what you can build in five years.
The Verdict: Real or Hype? #
AI incident response automation → Real but unevenly distributed. The productivity gains are genuine; the comprehension debt is real and most orgs are only tracking the first number.
Reader backlash against AI writing → Real. The data is early, but the direction aligns with what technically engaged readers report — authenticity is detectable and increasingly valued.
Population-level cognitive dependence risk → Real but early. The epidemiological model is suggestive, not settled, but the mechanism points to dynamics that are well-established in other automation contexts.