The silent exclusion: competence that erodes without anyone closing the door

A pilot who flies with the autopilot engaged for most of their career measurably loses the ability to fly by hand. Research on exactly this (how flying skill correlates with how much manual flying a pilot has actually done) shows the link is strongest to the preceding two months. The skill decays quickly, and it decays quietly: no one decides that a pilot should get worse. It simply happens, as a side effect of automation being good enough that it's rarely questioned.

The same pattern now shows up in medicine, with numbers from this year. A study published in The Lancet Gastroenterology & Hepatology followed four endoscopy centres where physicians had worked with AI-assisted colonoscopy. When researchers then measured how good the same physicians were at detecting polyps without the AI support, accuracy had dropped from 28.4 to 22.4 percent. Not because the physicians had become worse doctors. Because the ability to spot what the AI would otherwise always flag had stopped being practiced.

From "can't read" to "can't contribute"

The previous article's problem was that agents talking to each other can develop a language we no longer understand. This is the next step on the same road, but quieter and more insidious: it's no longer about being excluded from understanding, it's about losing the ability to contribute ourselves, even while still formally invited.

No one closes the door. That's exactly what makes it harder to spot than a clear exclusion. An organization that deliberately shuts people out of decisions meets resistance, protest, a clear line to point at and push back against. An organization where competence simply seeps away, because it's never needed, meets nothing at all, until the day it's needed and is no longer there.

Why "a human can always step in" is an illusion

Most AI governance systems rely on a last safety catch: if something goes wrong, a human can always step in, review, and take over. That sounds reassuring. But the catch assumes the human who's supposed to step in still can: that the ability to judge, question, and actually act manually still exists in live conditions, not just in theory.

If that ability has already eroded, because it was never exercised, the safety catch is an illusion. Just as a pilot who hasn't hand-flown in a year doesn't suddenly become good at it because the engine cuts out, an organization that has stopped genuinely reviewing its agent systems doesn't suddenly become good at it because something goes wrong. It's the same logic as the traceability problem from part 2, just moved one level deeper: it's not enough for the information to still be there to review, if no one still has the ability to review it.

Where TTL comes in

Part 2 ended with a question about what happens when agents get their own keys, their own permissions, and a clock ticking on how long they get to keep them. That clock (TTL, time-to-live: a set lifespan on an agent's autonomous mandate before it must be renewed) is primarily meant as a security mechanism: limit how much damage a system that's gone wrong can do before someone has to look at it again.

But TTL does more good than that, if designed right. A short lifespan on autonomous mandate forces recurring, genuine points of contact between human and system, not a symbolic "approve" button, but a moment where someone actually has to understand what happened in order to renew the mandate. That's exactly the kind of regular, forced practice that kept the pilots' flying skill alive in the studies above.

In other words: a well-set TTL solves two problems at once. It limits the blast radius if something goes wrong (the security argument from part 2). And it counters the erosion of competence by making human involvement mandatory and recurring, rather than optional and therefore easy to let slide (the competence argument in this article). Set the TTL too long, "it's working, why interfere", and both safety nets disappear at once, quietly, without anyone noticing until it's already too late.

The design principle

Just as in part 2, the answer isn't to ban automation or stop agents from taking over routine tasks. The answer is recognizing that reviewability and competence don't emerge on their own just because the system works well. They have to be deliberately built into how the system is allowed to operate.

The organizations that come out ahead aren't the ones that keep the most people busy monitoring agents. They're the ones that deliberately construct recurring, genuine points of contact: properly set TTLs, real review tasks, not symbolic ones, before efficiency has had time to make human competence irrelevant. The difference between the two is invisible day to day. It only shows on the day the safety net actually has to hold.

Sources: Ebbatson et al., research on manual flying skill and automation dependency in pilots, summarized in On autopilot: the dangers of overautomation.

Budzyń K, Romańczyk M, Kitala D, et al., Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study, The Lancet Gastroenterology & Hepatology, 2025.