The Judgment Reserve

A senior engineer posted something last week that a few thousand people apparently recognized in themselves. He used to be able to hold a solution in his head, trace it line by line, see how the pieces connected before he'd typed a word of it. Lately that ability feels like it's degrading. Not because he stopped being smart. Because he stopped doing the thing that kept the ability sharp.

That's not a complaint about AI tools. It's a report from someone watching a skill atrophy in real time and being honest enough to say so in public, to two hundred replies' worth of people who'd noticed the same thing.

He's not wrong to be worried, and it isn't just a feeling. MIT Media Lab researchers wired up 54 people writing essays with an LLM, with a search engine, and with nothing but their own heads, then measured brain connectivity with EEG. The LLM group showed weaker neural coupling, worse recall of their own work, and — this is the part that should land in an engineering org — the deficit didn't reverse when the tool was taken away in a later session. The researchers call it cognitive debt: you borrow the effort now, and the interest is a mind that's worse at doing the thing unassisted, on a timeline that outlasts the assistance.

Every conversation about AI and engineering skill so far has been about juniors — whether the pipeline of future seniors is drying up because the reps that used to build judgment now get skipped. That's real, and worth its own worry. But the MIT data wasn't about students. It was about people who already knew how to write. The debt accrues on skill you already have, not just skill you're building. A senior engineer's judgment isn't a credential he earned once. It's a muscle he's currently either using or losing, and the losing part doesn't announce itself until the one week it matters.

You can see the shape of where it matters most in a completely different thread, this one from an engineering manager describing a year of technical interviews run with AI tools fully allowed. The finding that stuck: candidates split cleanly on one thing, and it wasn't how fast they produced code. It was how they treated what the AI handed back — as a finished answer to merge, or as a junior's first draft to interrogate. The generation step had stopped differentiating anyone. The judgment step was the entire signal left.

Put those two threads together and you get the actual risk, and it isn't "engineers will forget how to code." It's that the specific capability your org needs most from its senior people — catching the AI-generated fix that's subtly wrong, recognizing the novel failure mode no training data covers, making the call at 2 a.m. when the playbook runs out — is drawn from the exact reserve that atrophies fastest when it goes unused. Call it the judgment reserve: the accumulated capacity, built by having actually wrestled with hard and ambiguous problems, that you draw down precisely when there's no lookup table left to consult. It doesn't show up on a velocity chart. It shows up, or doesn't, the day the incident isn't the kind you've seen before.

Most engineering orgs are optimizing to spend that reserve down to zero without noticing, because every visible metric points the other way. Tickets close faster. PRs merge faster. The dashboard looks like health. What it's actually measuring is throughput on the problems that were already legible enough to hand to a model — which is a real and valuable thing to automate. The quiet cost is that the people you're counting on to handle what isn't legible are getting fewer and fewer chances to practice handling it, and the org has no line item for that until the outage that doesn't match any runbook, on the night the person who could've solved it in fifteen minutes two years ago now needs an hour and a half.

We wrote last quarter about friction that got removed from product decisions and revealed an idea problem underneath. This is the same mechanism aimed at a different target: the friction wasn't just testing whether an idea was worth building. Some of it was the rep that kept a specific person's judgment sharp. Removing all friction indiscriminately doesn't just expose what you didn't have — it can also quietly dissolve what you did.

A technical leader's job now includes an inventory that didn't used to exist: which friction is waste, and which friction is training. The mundane migration, the boilerplate CRUD endpoint, the fix with twenty Stack Overflow answers already — hand all of it to the model, gladly. But the architectural call with no clean precedent, the incident that doesn't pattern-match to last quarter's, the review of a system nobody currently on the team designed — those are exactly the reps you protect on purpose, even when the AI could technically produce an answer faster. Not because slower is virtuous. Because the fifteen minutes it costs today is what keeps the hour and a half from becoming eight when it counts.

That's a harder inventory to run than most roadmap decisions, because the payoff is invisible until the day it isn't, and by then the org has already decided which muscles it let go slack. It's also exactly the kind of call a fractional CTO exists to make: someone whose own judgment reserve stays stocked because every engagement forces a genuinely new problem, no playbook supplied, and who can look at your team's real risk profile and tell you which friction to protect before the incident does it for you. If you're not sure which of your team's hard problems are still building judgment and which have quietly become just another prompt, that's exactly the conversation worth having.