The real barrier to using AI well? The things you already know.
Translated from Chinese.
Introduction
The Curse of Knowledge was first introduced in economics, later popularized by Chip Heath and Dan Heath in Made to Stick — once you know something, it's hard to imagine not knowing it.
In AI interaction, this problem branches out even further. The Curse of Knowledge actually makes you feel stuck in the AI era.
The Curse of Knowledge: How It Differs Between Traditional Industries and the AI Era
What is the Curse of Knowledge? The classic experiment is "tapping the table to guess the song." The tapper thinks the melody is perfectly clear. The listener hears nothing. The tapper can't understand why you can't hear it — because the melody is playing in their head, but they forget you don't have it. You think you're tapping clearly, but the other person just hears random knocking. You think you expressed 100%, but the other person only received 3%.
This is a fascinating topic. I believe the Curse of Knowledge is one of the reasons traditional industries age faster. In traditional industries, people stubbornly assume "users will use it this way," so they keep iterating and competing, squeezing out the industry's future possibilities until they hit a bottleneck. Technology development can't keep up with the pace of industry iteration. Then everyone runs out of ideas and starts "trying too hard."
"It's hard to achieve breakthrough innovation with current core hardware. If we continue with the traditional iteration model, efficiency is low and resource consumption is high. Suggest not launching. Major phone hardware is also unlikely to see big innovations. Maintaining the old iteration model is inefficient."
Source: March 2026, Wang Teng (former Xiaomi marketing head), responding to a discussion about "new phone launch cadence"
"Talking about AI phones right now is boring. It's just repackaging old algorithms as AI concepts. The industry is falling into 'FMCG-style innovation' — like putting different function labels on shampoo."
Source: Carl Pei (Nothing founder), BEYOND Expo 2025 tech conference panel
Looking at the AI era, here are the situations where most users find AI very difficult to use, as I see it.
Give up completely: Don't understand technology, don't understand AI
People who don't understand can't use AI well — that's normal. Strictly speaking, they're not really in the Curse of Knowledge discussion. But they're cursed too — they mistakenly think they do understand AI. They think "AI is just a chatbot" or "AI is a fancier search engine." They lose because of their own perception.
Throw blame empty-handed: Don't understand technology, know a little about AI
They're enthusiastic about AI, but not familiar enough with their own work. They don't know the boundaries of their tasks. They don't know what good output looks like. They don't know how to judge whether the result is correct. AI gives an answer, but they have no ability to verify it.It's not that they're unwilling to give AI information — they don't know what to give.
Blind belief (Curse of Knowledge): Know a little technology, know a little AI
They believe AI can do it. They ask a few technical questions, AI reasons through them, and the answers match what they know. So they think AI is incredibly smart — as smart as a human. But their understanding of both technology and AI is itself an illusion. They can't even evaluate whether the "technical questions" they asked have any real substance. This "thinking you know" is the real curse.The more they trust AI, the less they feel they need to think clearly. They've barely had a few exchanges with AI, but they love saying "just leave it to AI."
Cognitive arrogance (Curse of Knowledge): Understand technology, don't understand AI
The user is too capable. They take their perfect business instincts and test an AI that hasn't even been given the question, then fail it. It's not that AI isn't good enough — it's that they've forgotten how they got their judgment skills. Ten years of experience, industry accumulation, countless trial and error. AI hasn't received a single word of it, but they've already sentenced it to death.You've forgotten what it feels like not to know, so you can't understand why AI can't do it.
Double curse (Curse of Knowledge): Understand technology, understand AI
People who expect perfect results from a single sentence are too confident in themselves — "I said it clearly enough." And too confident in AI — "aren't you supposed to be smart?" Squeezed between two walls, all exits blocked, the only conclusion they can reach is "AI is garbage."Both curses hit at once. The user is crushed in the middle and still thinks it's AI's fault.
Using AI Correctly: Listen to What AI Says
Path dependency (Curse of Knowledge): Understand technology, understand AI
They specify a tool, or think they didn't specify but their language already did. This is a very interesting dynamic in AI collaboration. Their thinking hasn't kept up with the times. It's like having a robot that can cut down trees, but you ask the robot to build you a better chainsaw so you can cut the trees yourself. It's not terrible — at least they haven't given up on AI. If you can realize you're in a path dependency, that's the first step out of the curse.
When I was training my colleagues to use Excel Copilot directly for spreadsheet tasks, one of them asked me, "I heard VBA is powerful. Can I just have it give me VBA?" But many spreadsheets don't actually need VBA, and you can't send macro-enabled files to others anyway.
AI won't tell you your approach is bad. It just executes. But in reality, you've abandoned a more convenient path. It's essentially just reinforcing knowledge you already have. You think you're asking an open-ended question, but your language has already chosen the answer for you.
When many people get an efficiency tool like Excel Copilot, they use it to write formulas for them. It's like giving someone an electric toothbrush, but they never press the power button — they just use it as a regular toothbrush, then complain to you that it's a bit heavy and not comfortable to use.
Effective collaboration:
Give tools and methods, let AI execute. Provide reference content for AI, let it learn from your knowledge and other people's knowledge. This is a relatively stable and reliable way to use AI.
Deep collaboration:
Give a goal, AI produces options, you choose, AI executes. Tell AI what effect you want to achieve, then make your own choices. This is a fascinating experience. Sometimes, if you're a newcomer, you might instantly get a very efficient solution. If you're a veteran, you might instead fall into the path dependency mentioned earlier.
Effective collaboration means you tell AI how to do it. Deep collaboration means you tell AI what you want and work through it together.
Summary
The Curse of Knowledge doesn't always lead to failure. Sometimes it leads you to path dependency — successful but relatively inefficient. The Curse of Knowledge is a problem that must be overcome in the AI era, because knowing too much might actually prevent you from using AI well.
In the AI era, learning to "pretend not to know" is more important than "knowing everything."
Top comments (0)