Your performance with AI is climbing; your performance without it is quietly falling — and you only ever see the first ledger. Aviation paid sixty years of tuition for this.
Swap the cockpit for an IDE and the autopilot for a coding agent, and you'll find aviation walked our exact path forty years ago. Skill decay isn't a question of "whether" — it's a dosing question of "how fast, and which skill first." Arthur et al.'s heavily-cited 1998 meta-analysis gave the curve: effect size is roughly 0 right after training, dropping to about −1.4 standard deviations after more than a year of nonuse — and cognitive, judgment, and accuracy-based skills decay faster than physical, speed-based ones. That's the worst possible combination: what AI takes over is exactly the judgment work that degrades fastest. This issue skips the "does AI make us dumber" hot-take and covers three things that actually bite in engineering: the dosing law of skill decay, the second ledger you never see, and a "dose-verification" prescription that even aviation regulators tripped over.
Arthur et al. meta-analyzed 50+ studies and nearly 200 data points, quantifying "use it or lose it" into a curve: skill loss deepens monotonically with the nonuse interval — about −0.01 right after training, dropping to roughly −1.4 SD after 365+ days. More importantly, their moderator analysis: cognitive, accuracy-based, artificial (non-natural) tasks decay far more than physical, speed-based, natural ones. Map that onto AI collaboration and it stings: what you hand AI was never the manual grunt work, it's should we design it this way, is this boundary condition right, where does this logic blow up — all the fastest-decaying judgment skills. So "I use AI to code every day, my chops won't rust" is an illusion: you're practicing directing and reviewing, not deriving by hand — and the latter is the ledger that empties fastest.
First run a skill audit, mapping your core skills onto a minefield by "outsourcing depth × decay speed":
# Skill-freshness audit: lower-right = danger (deeply outsourced + fast decay)
skill outsourced-to-AI Arthur class risk
architecture trade-offs high cognitive/acc → steep HIGH
boundary/failure-mode high cognitive/acc → steep HIGH
writing algorithms cold med-high cognitive → steep WATCH
looking up API/syntax high memory → relearnable low
naming/formatting high speed → shallow low
# Rule: fully outsource the low-risk; keep a "manual dose" for the high-riskThe point of the audit is triage: fully outsource syntax lookups and boilerplate — "memory + instantly relearnable" skills. Reserve the manual dose for the judgment skills that are both deeply outsourced and sitting in the steep decay zone.
Aviation human-factors gave these two ledgers a name: the out-of-the-loop performance problem. Endsley & Kiris (1995) ran a clean experiment: let an expert system take over navigation and operators do fine in the moment; but the instant the system fails and a human must take over, the fully-automated groups show significantly worse reaction time and situation awareness. The mechanism is a shift from active processing to passive monitoring — you no longer make decisions by hand, you just watch it decide, so you lose both the skill and the real-time grasp of "what's actually going on right now." That's the two ledgers: the first is "how capable you are with AI", climbing daily, the KPI you and your boss both watch; the second is "how much you have left once AI is gone", never on any dashboard — yet it decides whether you can catch it when AI goes down, hands you an out-of-scope task, or gives a plausible-but-wrong answer.
# Schedule an "AI outage day" to read the second ledger
Today: core task with zero AI, log just two things—
1. Stall points: which step used to be smooth, now needs a pause? # decay signal
2. Completion/time vs. usual with AI: how big the gap? # the out-of-loop gap
# Variant: take a task beyond AI's current ability, see if you can catch it solo
# Don't log a score, log "which judgments I no longer dare make myself"This isn't self-punishment, it's fitting a gauge to the second ledger. If you don't measure it, it stays 0 — until reality forces the account settled one day.
Aviation gave the prescription long ago — and stepped on its rake. FAA's 2013 SAFO 13002 says it plainly: continuous autopilot use does not reinforce manual-flight skills, and degrades the ability to "quickly recover from an undesired state," so operators must actively create hand-flying opportunities. Translated into a personal prescription, three ingredients — none optional: ① regularly complete core tasks without AI (the "hand-flying," reclaiming the steep-decay judgment skills); ② live through AI failure firsthand — not hearing "AI hallucinates," but actually getting burned by a confidently-wrong answer and debugging it out by hand. Aviation training repeatedly finds: verbal briefings and generic courses barely transfer, only experiential / startle scenarios stick; ③ personal accountability for the result — you sign off, not "the AI wrote it." Together these three counter Day 53's generate–verify inversion: production mode has AI generate and you verify; freshness must switch back to learning mode — you generate, AI verifies.
# Personal manual-dose prescription (pin to your weekly plan)
[ ] Manual dose: each week, pick 1 core task and do it fully by hand
→ choose a HIGH-risk skill (architecture/boundary reasoning), not syntax lookup
[ ] Failure drill: each month, deliberately dig into an AI wrong output,
reproduce by hand where it errs —— living it > being told
[ ] Accountability: every AI output ships under your name, your liability,
and you can walk through "why this is right" line by line
# Learning-mode switch: core skill = you generate / AI verifies (counter the inversion)Note the selection of ingredient one: spend the manual dose on the high-risk skills the point-1 audit surfaced. Wasting precious "AI-free time" on low-risk skills like syntax lookup is the most common fake diligence.
This is the most counterintuitive — and most valuable — lesson. Issuing SAFO 13002 didn't solve it: a DOT Office of Inspector General (OIG) 2016 flight-deck-automation audit found that operators did not ensure pilots actually got and used hand-flying opportunities — the prescription was written, the dose never landed, skills kept decaying. The reason is plain: hand-flying is more tiring, slower, more prone to small in-the-moment errors, and under first-ledger KPI pressure like "on-time rate," both pilots and airlines instinctively revert to autopilot. Relying on intent and willpower to sustain correct behavior is the shared failure mode of every freshness prescription. The fix is to move "verification" from the person to the process (Day 54's oracle inventory, Day 55's "bake it into the default rather than daily willpower"): fit the manual dose with a record and check independent of your will, so "did this week's dose actually happen" becomes a queryable fact, not a self-assessment.
# Fit the manual dose with "dose verification" — intent → queryable fact
[ ] Log the event: tag manual-dose tasks (git commit / journal),
count "core tasks done AI-free" at month end
[ ] Independent signal: use point-2's "post-removal time gap" as the objective gauge,
not the self-assessment "my chops feel fine lately"
[ ] Bake in the default: schedule it into a fixed slot / process hook,
don't rely on remembering—micro-dose + intermittent = the standard way it fails
# Verify "the dose actually happened," not "I intend to"The core is one line: writing the prescription is intent, verification is fact. Freshness rarely trips on "not knowing you should go manual" — it trips on "thinking you went manual when you didn't."
① Dosing law — decay deepens monotonically with nonuse time (Arthur 1998, over 1 SD after a year), and cognitive/judgment beats physical for speed of loss; audit first, spend the manual dose on "deeply outsourced + steep-decay" high-risk skills.
② Two ledgers — assisted gain (AI in hand, rising daily) masks post-removal decline (AI failed, unmeasured); schedule an "AI outage day" to read the second ledger — the out-of-loop gap is 0 until you measure it.
③ Three ingredients — regular AI-free core tasks + living through AI failure (living it > being told) + personal accountability; switch core skills back to "you generate / AI verifies."
④ Verify the dose — SAFO issued, skills decayed anyway, because landing wasn't verified (DOT OIG 2016). Intent isn't a dose — fit an independent record + bake in a process default.