Title: Can the A.I. experts who seek to improve A.I.’s reliability paradoxically decrease system-level reliability?

Body:  We often read about how AI causes experts to deskill and/or novices to never-skill. Recently, Nature Medicine says the issue is novices are "mis-skilling" (learning the wrong way). The solution to deskill, never-skill, and mis-skill is often presented as re-skilling humans (without AI) or improving the AI itself. These may both be good solutions, but I wonder if scholar are over-simplifying the triangle relationship between novices, experts, and AI. Is the very act of trying to improving the triangular relationships paradoxically worsening the problem?

 Interactions between technology and humans is complex, but can be counter-intuitive. Notably, Bainbridge’s “Ironies of Automation” (1983) showed that automation leaves humans responsible for monitoring work that they, ironically, no longer perform. Trim (2026) extends this to A.I. coding: reduced hands-on coding decreases the expertise needed to monitor A.I.’s code. The repercussion is that novices lacking skills reduce code security (Perry et al., 2023; Prather et al., 2024; Gardella, Bolton & Riggs, 2026). However, beyond these prior studies, I think the issue is deeper: Is reskilling novices as experts and improving AI all that is needed? I question some methodological assumptions made regardless of whether studies argue for improving either 1)  A.I.’s own technology or 2) human interactions with it. Statistically, when A.I. scholars test for positive or negative results for using A.I., many compare novices against experts in a cross comparison, when longitudinal studies are required. With a cross comparison, the A.I.’s benefits and threats to humans are linearly related, while a longitudinal study may reveal the larger catastrophes may lie in imbalances between the two rates-of-change of dynamic interactions.

Therefore, I suggest scholars should look at how the rates-of-change in interactions between novices, experts, and A.I. create imbalances. I hypothesize that if we test whether and how novices are trained to become experts longitudinally, we would find novices cannot aim to match expert skills, but must supersede them. Why? Because today’s experts are not necessarily a model for the experts needed tomorrow. Yesterday’s novices became experts through tacit knowledge: writing, troubleshooting, testing, and diagnosing code. However, as A.I.’s coding abilities become more reliable, both novices and experts may increasingly let A.I. perform much of that work. This does not mean that all skill disappears, or that no new skills emerge. Rather, the concern is that independent debugging and error-detection skill may decline faster than organizations can replace it with new forms of oversight. Adding to the difficulty, A.I.’s rate-of-change continually improves. For novices to maintain skepticism about A.I.-generated code requires consistent cognitive effort, while the volume and complexity of A.I.-generated code may grow faster than human review capacity. Even the maintenance of existing guardrails may therefore be insufficient as coders must continually strengthen the guardrails that assist A.I against those who misuse A.I. Otherwise, A.I. may become the last gatekeeper against a cyber threat. If that threat eventually exceeds the A.I.’s reliability, neither the technology nor skilled human monitors will be adequate.

The following graph visualizes this theory as one possible trajectory: skills initially improve, but later decline because skepticism fails faster than A.I. reliability rises. Even if cyber threats remain stable initially, they may accelerate because a bug that remains undetected inside A.I.-generated code can compound as the volume of A.I.-generated code grows exponentially. Eventually, declining human skill may not compensate for threats that A.I. fails to handle.

The marked points describe stages in that progression:

Expert Detection: Human skill and A.I. reliability intersect. This is a period when human expertise can still independently detect A.I. errors.

Automation Bias Tipping Point: A.I. reliability overtakes human skepticism. Trust in automation becomes stronger than the inclination to question it.

Novice Failed Detection: Deskilled coders cannot identify problems that escape the automated system.

Cyber Threat Exceeds A.I. Reliability: Threats eventually outpace the A.I. itself. At this point, neither the technology nor deskilled humans monitoring it are adequate.

Implication: After a longitudinal study on the triage between novices, experts, and AI to identify how the dynamic rate-of-change causes an imbalance, a second study longitudinal study about how the balance can be restored by testing the dynamics of tacit knowledge of experts, AI, and how novice can supersede experts.