I noticed that Hank Green paused three of his channels this week. Someone watching an Ask Hank Anything episode noticed him say "I appreciate the pushback," which is a phrase ChatGPT reaches for when you challenge it, and the clip went around the social media sphere. He says that particular line was an ad lib, and that is quite interesting. What he admitted to was bigger than one sentence: "I did use ChatGPT for research on this script and watching it I definitely get an AI feel, so I think it's fair to say I was relying too much on generated notes." Then he went further, and I remember this the most from when I checked it out. "I need to come to terms with the fact that the level of dopamine I've been getting from interacting with LLMs is not healthy for me or good for the world."

In reality, people did not find AI in his script but heard him say it out loud.

The internet then did the internet thing. Half of it decided he had revealed himself as a fraud, the other half decided the first half were Luddites. The question sitting underneath is more useful in my mind. Here you go: where is the line? The line where you pause and go: "Hmm, I should probably handle this part even if it's difficult."

It took shelving an entire novel for me to find mine.

A novel

Around February or March this year I went deep into learning the tooling. Claude, Codex, Antigravity, Kimi, a few open models running locally through Ollama. I wanted to test them on something I knew how to do well, writing books. I decided to foray into historical romance fiction.

I did it properly, or at least I thought so at the time. I downloaded a lot of romance and historical fiction, read through the conventions, and built a rules file: how this kind of book is structured, what the beats are, what has to be true by which chapter. Then Claude drafted it chapter by chapter against those rules. After the draft came an editing agent, then a specialist editor, then a proof editor, then a final check agent. Five stages, all of them automated, each one handing off to the next.

It worked, in the narrow sense that the machinery did exactly what I built it to do. I had finished the book eventually, reading it and going back and editing, commenting and understanding how LLMs write and work.

I did not notice the problem until I converted it to EPUB and sent it to my Kindle. Reading it as a book, rather than as output in a terminal, every chapter came back soulless. It did not flow. It was not consistent. The prose had that unfeeling generic quality that I can only describe as nobody being home. So I did what you would do: I edited. Then I built more agents to fix the flow, and ran it through again. Two months of that, and it never got better. Eventually I shelved the whole thing, because it was unreadable.

The part I still find interesting is that I could not edit my way out. Normally a bad draft has something underneath it that you are editing toward. This one had nothing underneath. There was no version of the book that sounded like me, because I had not specifically written it from the beginning consistently.

So the rule

I realised that it's most important for the early stages of any written work to be drafted by me.

The reason is in how the thing works. An LLM predicts the next token based on what it was trained to predict, and a lot of that training material is people talking on the internet. Ask it to produce sentences and it gives you the average of everything it has read, which is exactly what soulless means in this context. Nothing is malfunctioning when that happens; the model is doing the job it was built for.

Diagnosis is where these tools are genuinely good, and it is the part people skip straight past on the way to asking for a draft. I transcribe my ideas and ask a model to make sense of them, which is honestly the single most useful thing I do with any of this. It catches grammar problems I would otherwise be paying Grammarly for. On the novel it caught inconsistencies between characters and built me a sheet of every one of them, which I could then work through myself. All of that is super helpful, and I can rest assured that I checked every word before publishing.

I am writing The Republic of Birds from scratch now. I can feel the difference already. Reading chapter one into chapter two, the same person is clearly writing both, because the same person is.

Where the line sits day to day

The exception is code, and working out why took me a while.

I built a PowerPoint generator for lessons using agents: one to find images, one to define the lesson plan, one to build the deck, one to check the spelling. That process is predictable and it has a defined output, and the agents handle it well. NudgeCue, my classroom tool, is a bigger case. I do not have a lot of coding experience, and the agent genuinely taught me while we built it. I spent most of my time specifying rules, asking it to run security checks, and cross-checking what it gave me.

So why is generated code fine when generated prose is not?

Code has a spec. Prose has a voice. When I hand NudgeCue's authentication to a model, I can test what comes back. Does it run, does it pass, does it leak anything, does it do the thing I described. The standard sits outside the code and outside me, and it does not care who wrote it. If it holds up, it holds up.

Prose has no test that runs. The only standard is whether it sounds like the person whose name is on it, and the only instrument that can measure that is the person. Hand the sentences over and you have given away the measuring device along with the work, which is why I could not repair the LLM-generated novel. It had so many different voices that I couldn't bring my own into it without redrafting it from scratch.

That is where the line falls for me. Not because code is less important, but because code can be checked by something other than me, and my sentences cannot.

The honest bit

I depend on these tools a lot. I have used them every day for about a year, across the novel, the worksheets, the app, and half the things on this site, and I am not going to pretend that dependence is nothing.

The pull is the same one Hank described. Having a sparring partner available at any hour, one that hands back a fully formed outline in twenty seconds, is genuinely addictive. It is tempting every single time. And I think the cost is slower and harder to see than people assume: you do not notice losing your own thinking, you notice much later that you have stopped reaching for it. Some people have started talking like the model on camera, and you can hear it, that flat sense that you are listening to a script rather than a person.

This post is not exempt. I recorded myself talking through it for ten minutes, a tool transcribed the audio, and a model checked my facts, including one claim about a Google executive that I had misremembered badly enough that it came out of the draft. What it did not do was decide what I think. I love transcribing my thoughts as a first step, and it was helpful to get this draft out. Right now, I'm going through the draft line by line and cross-checking to see if my transcribed video is within this.

The test

I read a lot, and I have now read a lot of books that were generated. It is easy to spot once you have built the thing yourself, because the pattern recognition just sits there and will not switch off. Every AI tell is obvious in the text.

So the test I use is one question. Does this output have to sound like me? If it does, I write it. If it does not, I use the tool and get on with my day. Coding doesn't necessarily need my voice, so I make use of the tools for that. Grammar checks can also be used.

Your line will not be in the same place as mine, and it should not be. What matters is that you put it somewhere on purpose, rather than finding out where it was two months and one unreadable book too late.

References