How the Sausage Is Made (on writing articles with LLM and some thoughts on AI use)
This article was written with the use of ChatGPT.
Not just spellchecked or lightly edited. The entire conversation that eventually became this article happened inside ChatGPT over several hours. However, If you find something in one of my articles that is wrong, misleading or simply poorly argued, I don't get to point at the language model. My name is signed at the bottom of the article. The complete, unabridged transcript is available here. Read it if you're curious and judge for yourself.
I don't find the process particularly strange. It feels remarkably similar to conversations I've had with colleagues for years. I start with an idea that has been bouncing around my head for some time. Months, sometimes years. I will try to explain it. I get challenged. I go down a rabbit hole, abandon it, come back, discover that two ideas I thought were connected actually aren't. Outline gets formed, first draft generated. I edit, rewrite, paste it back and forth. A paragraph disappears because it sounded clever but didn't survive scrutiny. Another one stays because I couldn't find a better way to express it after twenty attempts. Eventually an article appears.
I have been asked whether these articles are still mine. I don't have a universal definition of authorship and I'm not interested in inventing one. I only know how I think about my own work. The tool does not understand my experiences better than I do. It does not know what I believe. It cannot decide which compromises are acceptable or whether an argument actually represents my thinking. I don't use ChatGPT for any of that, I use it because it helps me surface my ideas and because currently it generates words better than I do.
You could ask why I don't simply work with an editor. The short answer is that I haven't found one who fits the particular problem I am trying to solve. I need someone who can keep up with a fairly specific technical niche, challenge my reasoning without requiring me to explain the entire domain from scratch, and help me turn the result into readable prose. ChatGPT is surprisingly good at combining those things. It is currently a better fit for this particular job than the alternatives I've tried. It also occasionally points out weaknesses in my thinking. I would be foolish to reject those simply because they came from a language model, just as I would be foolish to accept everything it produces uncritically.
I don't think generative AI is another passing fad. It is a genuinely transformative technology. The printing press changed how knowledge spread. The steam engine changed how we used physical labor. Electricity quietly found its way into almost every part of modern life. The internet fundamentally changed how information moves. AI will almost certainly earn a place on the very same list. What all of these technologies have in common is that they expanded what people were capable of doing. They changed the environment we operate in. They made some things trivial, others obsolete and created opportunities that previously didn't exist. None of them, however, relieved people from deciding what was worth doing in the first place. Human nature proved remarkably stubborn throughout all of those changes. We are still driven by incentives, still susceptible to fads, still very good at convincing ourselves that a powerful new capability is a solution before we've properly understood the problem.
Software engineering has its own version of this cycle. Every few years a new methodology, framework or technology arrives accompanied by promises that it will fundamentally change how we build software. Sometimes those promises are justified. Sometimes they are wildly optimistic. Most of them eventually settle into their appropriate place in the toolbox. The interesting part isn't which technologies succeeded and which didn't, but that the same mistake keeps repeating. We stop treating the new capability as a tool and quietly promote it into an objective. The conversation shifts. Instead of asking whether a particular approach is the best fit for the problem, we begin asking where else we can apply it. Adoption becomes a success metric in its own right. Once that happens, Goodhart's Law quietly takes over. If the organization rewards using the tool, people will find places to use the tool regardless of whether it improves the outcome. I've seen that pattern with Agile. I've seen it with microservices. I've seen it with blockchain. Today I see it with AI.
This is why I find many discussions about AI oddly unsatisfying. They often revolve around the technology itself, whether it is good or bad, whether people should embrace it or reject it, whether it will replace this profession or that one. Those questions make for lively debate, but they are not the questions I care about.
What problem are we trying to solve? What constraints are we operating under? What trade-offs are acceptable? Who owns the outcome? Who is accountable if we get it wrong?
Those questions existed long before anyone had heard of large language models. They will still exist when today's excitement inevitably gives way to whatever comes next. AI changes the available tools. It doesn't remove the need to exercise judgement about when, where and why those tools should be used. A hammer is a poor screwdriver. A spreadsheet is a terrible document editor. ChatGPT is an excellent conversation partner for some kinds of work and a terrible substitute for thinking. Using ChatGPT to help express ideas that I have already wrestled with makes sense to me. Using it to manufacture opinions I don't hold, or expertise I don't have, doesn't.
The interesting part, at least for me, isn't that AI helped produce this article. The interesting part is that you can inspect the process yourself. There were plenty of discussions that never made it into this article. We spent time talking about authorship, simulated reasoning, trust, faith, consciousness, taste, and whether any of those distinctions matter in practice. A lot of them were interesting, but they also made the article worse. They're still in the transcript if you feel like wandering off into the same rabbit holes I did.
You don't have to guess how much came from me and how much came from the model. You don't have to trust a vague disclosure saying "AI was used." You can watch me reject ideas, change my mind, argue over a single word, discover that I was making a claim I couldn't defend and remove it. Whether that earns your trust is entirely your decision. It feels more honest this way to me, rather than pretending the process doesn't exist.
