The Work I Couldn't Defend

I was impressed with what AI had done with my ideas. I hadn’t yet checked how many of them were still mine.

In July, I gave Claude Code and Codex the outline of a book I was exploring called The Abundance Trap. Along with it went my notes, articles I’d saved, and instructions for turning the whole thing into a website. Months of thinking, handed over with a request to make it look good and make sense.

The looking-good part was handled rather efficiently.

On my 52-inch Dell monitor, I could see a project that looked considerably further along than it had when I’d submitted the prompt. The homepage pulled the ideas together. A page I called Rewind took visitors through six decades of innovation, with roughly 120 examples of everyday difficulties we’d managed to eliminate. There was a lot to like. Some connections I’d been circling for months were suddenly right there on the screen.

I’d also left quite a bit in my head. Thoughts I hadn’t explained, distinctions that mattered to me, beliefs that weren’t spelled out in the material I’d supplied. I’d been thorough about the colours, though.

The AI filled the gaps. Some of its additions were spot on. Others came close enough to pass a quick inspection. A few expressed views I didn’t hold, with exactly the same confidence and lovely typography as everything else.

That was harder to notice while I was busy being impressed.

Once the initial excitement settled, I went back through the pages more carefully. I needed to work out how much of what I was admiring actually said what I meant.

The suggestion about friction was where things started to look off.

I’d been exploring what we lose when technology makes things easier. Some difficulties are a waste of everyone’s time. Others help us develop judgment or learn something we’ll need later. Remove those, and the convenience may come with a cost we haven’t noticed yet.

The AI had picked up on that distinction. It suggested that bringing back some useful friction could help protect the capabilities we were at risk of losing. There was a logic to it. If removing a difficulty takes away something valuable, restoring a little of that difficulty might help.

Then you have to persuade an actual human being to do it. They can be quite difficult about these things.

I’ve spent twenty-five years in marketing and digital products. “Would you like this to be a bit more of a pain in the ass?” is a difficult pitch. Especially to someone who’s just found an easier way to do it. People aren’t generally shopping for new inconveniences. There’s quite a reliable supply already.

The suggestion sounded sensible on the page. I couldn’t see it working as an answer people would willingly adopt. I wanted to explore other solutions, and I hadn’t supplied one yet. The AI had helpfully closed that discussion for me. My agreement hadn’t been necessary to complete the page.

I removed the recommendation. The homepage could help people notice what they might be giving up. It could raise questions about what their children were learning, or what a smoother process no longer asked them to do. Those were questions I wanted to pursue. I didn’t need a finished-looking answer occupying the space where I was still working.

The obvious diagnosis is bad prompting, which is convenient because the cure is more prompting.

There’s some truth to it. I could have explained more. I could have been clearer about what the AI was allowed to decide. I’d handed over an incomplete account of my thinking and then admired how complete it looked when it came back. I was in an excellent position to identify the person responsible.

But I can’t rely on writing a prompt so comprehensive that there’s nothing left for the model to assume. I need to inspect what comes back, especially when it looks convincing. Once it looks finished, questioning it starts to feel like being fussy. That’s a shit standard for deciding whether I agree with something.

I’m certainly not giving up the tool. I use AI every day, sometimes hundreds of times. It lets me build things I’d otherwise struggle to attempt. Declaring myself above all that would be an impressive position to maintain for the twelve minutes before I wanted its help again.

I do want to be much clearer about the help I’m asking for.

I want a sounding board that’ll challenge an argument, show me another angle, and help me find research I can actually check. It can improve my grammar and help a reader follow what I’m saying. If I’ve skipped a step in the reasoning, I want it to point that out. A compliment about how insightful I am won’t repair the hole, although I’m sure I’d enjoy receiving it.

The missing step is then mine to work through. It doesn’t get to finish the argument on my behalf and leave me to discover my new position during proofreading.

On later pages, I worked out the connections and the order of the argument before asking AI to help express them. I wanted to know what the reader needed to understand and why. There was still plenty for the tool to do once I’d bothered to explain what the hell I was trying to say.

I also kept things the AI had helped produce. The two-wall comparison was one of them.

A prison wall stops you going anywhere. A climbing wall gives you something difficult to do, and doing it develops a capability. Demolishing both in the interests of convenience would rather spoil the second activity.

The comparison helped me explain a distinction I believed. I understood why it worked, and I could see its limits. It didn’t tell me which difficulties a parent should preserve, or how to get someone to choose a useful challenge over an easier option. Those questions still needed answers. Getting two walls to illustrate a point hadn’t qualified me to start telling people how to live.

That’s why counting how many words I typed doesn’t tell me much about authorship. By that standard, dictation is cheating and every presidential speechwriter is an accomplice. I can use help with the sentences. I can have AI write the code and arrange the page. I still need to understand the reasoning and make the decisions about what I’m asking someone else to believe.

If somebody asks why an example belongs, or what follows from the argument, I need an answer I can explain myself. “Hang on while I ask the thing that wrote it” would clear up any remaining questions about how thoroughly I’d thought it through.

The first version of that website made it easy to feel further along than I was. I had a mixture of my own thinking, useful help, and conclusions I hadn’t chosen. It all looked equally ready.

I still want the help. But if I’m going to put my name under an argument, I’d like to have done more than admire the fucking layout.