I drive an F150 every day. I couldn’t rebuild the engine if you paid me, and nobody thinks I should park it until I can. Still, a good driver knows a few things. He knows what the warning lights mean, when the GPS is wrong, and how the truck sounds when something is off. That’s the question I wanted to answer about AI before I wrote a single lesson about it. I didn’t need to know how to build it, but I did need to know whether I could read my own dashboard.
What I Tried
Before this series started, I took down a free guide I had published about AI. It told readers where AI belongs and what it would always struggle with. When I read it again, I realized I hadn’t tested most of what it said. So I sat down with a long list of honest questions and answered them without looking anything up. I wrote down what tools I use, what I think I understand, where things have gone wrong, and how I’d rate myself. Then I sorted my answers into three piles: what I know, what I half know, and what I don’t know yet.
What Happened
The “know” pile was bigger than I feared. I use Claude every day and ChatGPT about five days a week, and I pay for Claude, ChatGPT, and CapCut. I use AI to help draft blog posts, check my work, make images, decide on processes, and keep me honest. When AI gives me setup advice, I test it by making it work. If the page header shows up or the plugin runs, the advice was good. Claude once walked me through building page headers, and now I build those myself unless I get stuck. I’ve also built some complex, useful spreadsheets with it.
I know AI misses, too. ChatGPT regularly ignored my rules on word counts, punctuation, and bullet points. Claude follows my rules but still makes mistakes, and so do I. We catch most of them by going back and forth before too much work gets lost. I also know AI “memory” isn’t like mine. It comes from instructions and stored information that make it seem like it remembers me.
The “half know” pile is where things got humbling. I think AI comes up with answers from its training, some kind of logic, and limited internet access, but I can’t tell you what’s really happening inside. My guess for why it makes things up is that it takes in so much information that the quality of the sources gets lost. I think a model is something like a version. I know a token has to do with how AI use gets measured and charged, but not what a token actually is. When I tried to explain the difference between an AI agent and an automation, my two definitions sounded almost the same.
The “don’t know yet” pile was short but real. I don’t know how to check a fact that neither Claude nor I would question, like a law, a number, or how a tool works. My image prompts in Canva were weak, so the images were too. I tried building a journaling app and a donations plugin, and both stalled. I didn’t have the time to explain what I wanted, the patience to learn Xcode, or enough trust in ChatGPT to follow its code. A handful of other tools, like Grammarly and a podcast-style voice tool, never stuck either.
Looking back, a pattern showed up. My first attempt at planning my business with AI never narrowed down to anything I could actually use. One master plan for a business idea read like a technical manual with no feeling behind it. Another time, I asked a real question about a program I was applying for. The answer was accurate, but it came back in fragments, and I had to dig for the point. It wasn’t steering me wrong. It was taking the long way to a direct answer.
When I rated myself from 1 to 10, I ended up with two numbers. For how comfortable I am using AI, I gave myself a 7. For how much I understand what’s actually happening inside it, I gave myself about a 4. As someone who uses it daily, I’m further along than a lot of beginners, so the 7 is fair. On how it actually works, though, the 4 fits. Neither number is a complaint. The 4 is my starting line, and the whole point of this series is to watch it move.
What I Learned
Using it every day is not the same as understanding it.
I’d been treating comfort as knowledge. I can drive this thing well, and I use it for real work every single day. But when I tried to explain the words I hear all the time, half of them fell apart. The guide I took down was proof of that gap. I taught things I hadn’t tested.
I also learned something about my frustrations. The answers that bothered me most were not wrong. They were voiceless, scattered, and slow to get to the point. That matters, because “the tool gave me bad information” and “the tool gave me good information badly” need different fixes. I’m not sure yet which one I was really dealing with. Sorting the piles at least showed me where to aim first.
What I’m Trying Next
The next lesson is learning to ask AI a clear question. Here’s the open question I’m carrying into it: were those fragmented answers the tool’s fault, or the way I asked? I honestly don’t know. I’m going to take real questions from my own work and practice asking them more clearly, then compare what comes back. If the answers get sharper, that tells me something about me. If they don’t, that tells me something about the tool.
Before then, try something for yourself. Think of the AI words you use or hear every week. If you had to explain them to a friend, how many could you actually define? Don’t look anything up first. The honest number is the useful one.
R.L. Canty | Canty’s Consulting & Instructional Delivery
