The Hidden Cost of Convenience in AI Tools
When AI tools write automatically, learn silently, act without asking — you trade legibility for speed. Some of that friction was doing something.
A 2024 study in the journal Society found a measurable negative correlation between frequent AI tool use and critical thinking. Not a dramatic effect, but real and replicated — and in the direction you'd expect if someone asked: what happens to a skill you stop exercising? The researchers called the mechanism cognitive offloading. When a system reliably handles something, the brain gradually reallocates attention away from it. That trade seems fine, most of the time. You can't recall phone numbers when your contacts app has them, and nobody thinks that's a crisis.
But there's a specific version of this that matters more than the general case, playing out quietly in the AI tools being built right now. When an AI silently writes to your data — modifying preferences, adding contact notes, marking tasks resolved — without surfacing that it happened, you lose something that isn't really about cognitive capacity. You lose the ability to know what your tool currently believes to be true about you.
The frictionless ideal
The dominant design goal in AI products right now is to remove friction. Auto-apply. One-click. Learn silently. The logic is defensible: friction is overhead. Every review step, every confirmation dialog, every "are you sure?" is a tax on attention. If an AI is right 95% of the time, requiring approval on every action means you're paying the cognitive cost of review for the 95 interactions that didn't need it.
That math looks fine until you realize it says nothing about how you find the other 5%.
The problem isn't error rate. It's that when a system acts silently, you have no mechanism for knowing when it erred. You can't locate a mistake you didn't see being made. When silent writes accumulate, each one looks small. Slightly wrong contact note. Half-resolved task. Preference updated on the basis of misread context. Individually, nothing. Together, a knowledge base that's partially wrong in ways you can't trace to their origin — and a growing gap between what you think the system knows about you and what it actually recorded.
What gets lost
Aviation safety researchers have a name for what happens when pilots fly highly automated aircraft: automation complacency. The documented drift toward passive monitoring of a system that usually handles things correctly. In 2009, the autopilot on Air France Flight 447 disengaged at altitude after speed sensors iced over. The pilots, trained on an Airbus A330 designed specifically to minimize manual workload, couldn't reconstruct the plane's state in time. They had been passengers in their own cockpit long enough that flying it felt foreign. All 228 people on board died.
This is not a proportionate analogy for an AI that silently updates your contact notes. The stakes are completely different. But the mechanism is the same: when a system reliably acts in the background, your mental model of what it's doing degrades. You stop tracking it because tracking it hasn't been necessary.
The 2024 cognitive offloading research quantified part of this. Frequent AI use correlated with reduced critical thinking, mediated specifically by reduced scrutiny of AI outputs. People were accepting conclusions without checking them. Not from incuriosity. Because checking had stopped being rewarded. The system was usually right, and reviewing it was overhead.
"Usually right" and "inspectable" are different properties. A system can have both. Most aren't designed for both.
There's also a compounding problem that's easy to miss. Errors written silently to memory continue influencing future reasoning — an AI asked about your preferences next month will read the incorrect version and act on it. The mistake propagates. And because it was never surfaced, you have no way to know it happened or when it started.
Where useful friction lives
The case for friction in AI writes isn't an argument for making things slower. It's about the difference between a system that acts and a system that shows you what it intends to do before it does it.
A diff is fast. Reading "changed: prefers async comms → prefers direct Slack messages" takes three seconds. Approving it takes one. The overhead is trivial. What you get in return is a running picture of the state of your knowledge, because you were present for each update. The AI is still doing the work. You're just keeping track of what it did.
This is what makes certain friction productive. Not slowness for its own sake. The maintenance of your own internal model of a system making changes on your behalf. Without it, you're operating on trust plus hope. With it, you're operating on something closer to actual knowledge.
The goal isn't to slow the AI down. It's to remain a participant in what it does.
The interesting thing about reviewing AI writes is that it becomes quietly boring. A list of proposed changes, accept or reject, over and over. But after a few months of it, you notice something: you actually know what your knowledge base contains. Not vaguely, not approximately. You know, because you watched each piece arrive. That turns out to be rarer than it sounds. And rarer than it should be.
Asgeir Albretsen is the founder of Harbor.