Accidental Friction vs. Protective Friction: What AI Should and Shouldn't Remove
Updated
Knowledge on this page was mainly distilled from AI Didn’t Automate the Grind, It’s Doing Something More Interesting.
Two Kinds of Friction
Not all friction is the same. Understanding the difference is essential to using AI effectively rather than destructively.
Accidental Friction
Accidental friction is like a squeaky door that makes it annoying to go outside. It exists due to tooling limitations, missing context, setup complexity, or knowledge gaps. Examples:
- Spending an hour on boilerplate before you can test an idea
- Searching for the right syntax or API call
- Not knowing the right vocabulary to ask a question
- Manual data reformatting between tools
This friction adds cost without adding learning. AI removes it well, and removing it is almost always a net positive.
Protective Friction
Protective friction is like a seatbelt—it forces a pause and makes a choice feel expensive enough to consider carefully. Examples:
- Customer conversations before building — Paul Graham's "Do Things that Don't Scale" argues unscalable work teaches what customers actually need before teams automate the wrong thing
- Manual processes that reveal edge cases — handling exceptions by hand exposes failure modes that automated systems would hide
- Slow iteration that forces commitment — when each attempt is expensive, you think harder about what to try next
- Writing from scratch to test understanding — generating text with AI can mask gaps in your own comprehension
This friction carries information. Removing it blindly removes the feedback loop.
AI Removes Friction Indiscriminately
AI doesn't distinguish between a squeaky door and a seatbelt. It lowers all costs of starting, iterating, and producing. This is why it can simultaneously feel like a superpower and a trap.
The superpower: questions that would have died in setup become working artifacts. The trap: validation steps that would have caught bad assumptions get skipped because the output looks done.
A Practical Filter
Before removing friction with AI, ask: is this friction adding information or just adding cost?
- If removing it means you try more things against reality, remove it
- If removing it means you skip the step where reality pushes back, keep it
The goal isn't to maximize speed. It's to maximize the rate of honest feedback.
Q&A
What is accidental friction?
Friction that exists due to tooling limitations, setup complexity, or knowledge gaps. It adds cost without adding learning—like boilerplate code, syntax lookup, or manual data reformatting. AI removes this well, and doing so is almost always beneficial.
What is protective friction?
Friction that carries information and forces useful pauses—like customer conversations before building, manual processes that reveal edge cases, or writing from scratch to test comprehension. Removing it blindly removes the feedback loop that prevents costly mistakes.
How can you tell which kind of friction AI is removing?
Ask whether removing the friction means you try more things against reality (accidental—remove it) or skip the step where reality pushes back (protective—keep it). The goal is maximizing honest feedback, not maximizing speed.