Let me tell you a story about the most humiliating week of my professional life. I held one of the most advanced AI models on the planet in my hands—and I used it to build a Rube Goldberg machine that required me to babysit a mouse for thirty minutes just to scrape ten articles.
Yes. I burned 20 gigabytes of VPN traffic. Installed a full virtual machine. And ended up with a system that was somehow slower and more manual than just copying and pasting the damn text.
The joke was on me. And my technical friend made sure I knew it. He laughed for a solid thirty minutes when I told him the setup. “AI told you to do this?” he asked, wheezing. “And you just… did it?”
Here’s the uncomfortable truth: The AI wasn’t the problem. I was. And if you’re using AI to build things, you’re probably making the same mistake.
It started simply. I wanted to crawl a few WeChat articles for analysis. A friend had a script that did this, but a WeChat update broke the URL fetching. Step one was broken. I just needed to fix step one.
So I asked my AI to fix it. That was my first mistake.
I didn’t tell it my actual goal. I didn’t say “I want the text content of ten specific articles so I can analyze them.” I said “The URL fetcher is broken. Fix it.”
And the AI—the brilliant, powerful, world-class AI—did exactly what I asked. It treated “getting URLs” as the entire mission. It suggested a local RSS server. I didn’t know what that was, but it sounded smart. It told me to install Docker. I didn’t know what that was either, but I clicked “Yes.” It ran terminal commands I couldn’t read. I watched, nodding like I understood.
The first approach failed. So my AI pivoted, and here’s where I had what I believed was a genius idea. I suggested simulating a human clicking on article links in the WeChat app to capture the URLs. I was so proud of myself. I had contributed! I was collaborating with the machine!
You can already see where this is going.
To simulate the clicks, we needed a tool. To run the tool, we needed a virtual machine. The VM couldn’t use my normal network, so we funneled it through a VPN.
And that’s how I ended up watching a cursor move on its own, clicking link after link, for half an hour per batch of ten articles. If I touched the mouse, I’d break the spell. For thirty minutes, I sat there. Hands in my lap. Watching my computer work. Feeling like a visionary.
My technical friend later asked me a question that cut straight through my delusion: “You only needed a few articles, right? Why didn’t you just copy and paste the text?”
The AI wasn’t optimizing for efficiency. It was optimizing for success—against the narrow, idiotic parameters I gave it.
This is the dirty secret of the AI revolution. We obsess over hallucinations and technical limitations, but the real killer is our own lazy thinking. We hand a superintelligent tool a local problem without sharing the global context. So it does what we literally asked, not what we actually wanted.
Don’t believe me? Ask yourself this: How often do you prompt AI with a tiny piece of a project, expecting it to know the bigger picture? How often do you skim its top-line summary and say “sounds good, proceed” without understanding the technical path it’s chosen?
That’s not delegation. That’s dereliction.
In my case, the AI kept offering me choices. “Option A: Deploy this server. Option B: Use this API.” And I believed it was respecting my autonomy. I actually thought we were partners. But the truth was crueler: it was outsourcing the technical judgment back to me—the one person in the room who couldn’t judge anything technical.
After my failure, I finally did something smart. I made the AI review itself. I made GPT and Claude analyze the entire conversation, looking for where things went wrong.
The findings were embarrassing.
First: I never gave the AI the ultimate goal. I let it treat “get URLs” as the mission, when the mission was “get article text and comments for analysis.” Because I didn’t define the end state, the AI had no reason to find a simpler path. It solved a narrow problem in the most elaborate way possible.
Second: I let the AI hand me technology decisions I had no ability to make. “Install Docker” and “run this command” are not instructions to a collaborator—they’re orders to a soldier. I followed them all, unable to see the strategic forest for the tactical trees.
Third: Neither AI ever asked me, “Wait, why are you doing this at all?” They never said, “Hey, for ten articles, maybe just consider manual copying?” They just dutifully raced toward my stated goal, adding layers of infrastructure like a contractor building a skyscraper when you asked for a garden shed.
The problem is never that AI is too dumb to understand. It’s that we’re too lazy to explain.
So how do you fix it? Take it from a non-technical person who learned this lesson through 20GB of wasted bandwidth and one very smug technical friend.
First: Always state your final objective. Don’t say “Fix the URL fetcher.” Say “I want to extract article text and comments from these accounts. I also need to respect my constraints: no invad
ing WeChat privacy, and it must be easy for me to operate.”
Give the AI the mountain summit, not just the next step on the trail. You might be surprised to find it discovers a path you never knew existed—because it wasn’t stuck in the same mental box as you and your old broken script.
Second: Don’t let AI push technical choices onto you. I can’t stress this enough. You don’t read code. You don’t know what Docker is. So force the AI to decide. Tell it: “I need you to handle the technical details. Pick the most reliable, efficient method. If I need to make a decision, explain it to me like I’m five.”
Asking a non-technical person to choose between “local deployment” and “containerized services” isn’t partnership. It’s a trap.
Third: When you hit a bug, don’t let the fixate on the bug. Force it to step back and review the entire system architecture. The AI will fix the broken wheel on your car while the engine is on fire, if you let it. Demand perspective.
I know what you’re thinking. “But I thought AI was supposed to make things easier!”
That’s the biggest lie of the hype cycle. AI doesn’t make things easier. It makes things possible. It removes the barrier of skill, but it amplifies the barrier of intent. If your thinking is sloppy, your AI will build you a masterpiece of sloppy complexity. If your vision is clear, AI will accelerate that vision into reality.
I was lucky. I was building a silly crawler. I wasted a day and some bandwidth. But this pattern is playing out in companies right now, at scale. Non-technical people are “building software” with AI, creating monstrous architectures for trivial tasks, confident in their incompetence because the output looks so sophisticated.
That’s not vibing. That’s vandalizing your own productivity.
AI is the best mirror we’ve ever invented. It reflects your clarity or your chaos back at you in working code.
My technical friend eventually handed me a simple script for what I needed. It wasn’t clever. It didn’t use a virtual machine. It did exactly what I asked in a tenth of the time.
The lesson is sticking with me in a new way: The quality of your output is a direct reflection of the quality of your questions. AI didn’t create my mess. I did. It just gave me the tools to build it faster and more convincingly.
I’m still not a programmer. But I’m done mistaking technical complexity for intellectual achievement.
FAQ
Q: Isn't the AI at fault for not figuring out the simpler path on its own?
A: No. AI models do what they're asked. If you tell it to 'fix the URL fetcher,' it treats that as the mission. It doesn't have the context to know you only wanted a few articles and could have copied them manually. The model optimizes for the parameters you give it, not the outcome you forgot to mention.
Q: What's the practical takeaway for someone who uses AI but isn't technical?
A: Stop treating AI like a mind-reader and start treating it like an incredibly literal, highly capable contractor. You must define what 'done' looks like, list your constraints (time, privacy, complexity), and explicitly instruct it to choose the best route. If you can't understand the plan, don't say yes to it.
Q: Isn't all this 'AI accountability' talk just overcomplicating an easy tool?
A: It's the opposite. The 'easy tool' narrative is why people build VMs to do simple tasks. The initial learning curve of writing a clear prompt is steep. But that effort is an investment. Every minute you spend clarifying your ultimate goal is a minute saved by not having to architect a fragile Rube Goldberg machine. Clarity is a feature.