Replacing Myself with AI

How I learned to stop worrying and love the GPT.

I knew the AI draft had gone wrong when I agreed with every sentence and trusted none of it.

It had the right shape: confident intro, neat workflow, optimistic conclusion. It used all the respectable words. Productivity. Creativity. Collaboration. Pipeline. It sounded like someone had replaced my actual experience with a brochure about my actual experience.

Which, inconveniently, was the whole problem.

The danger of AI writing is not that it is always bad. The danger is that it can be plausibly smooth before it has earned a point of view.

This post is an artifact from that phase. I was using GPT for lesson outlines, banner concepts, quiz prompts, rewrites, summaries, and translation experiments. Some of it worked. Some of it saved hours. Some of it produced the kind of bland professional paste that makes every person in a paragraph sound like a stakeholder.

I did not need less AI. I needed more friction.

The First Scar

The first real mistake was treating “looks finished” as evidence of thinking.

I would ask for a lesson outline and get something coherent in ten seconds. That coherence was intoxicating. A bad blank page makes its emptiness obvious. A decent AI draft hides the missing decisions under good grammar.

Then I would review it and find the usual problems:

• the examples were technically correct but forgettable
• the quiz questions tested vocabulary instead of judgment
• the explanations avoided the one hard distinction the learner needed
• the tone had drifted into corporate daylight

Nothing was catastrophically wrong. That was the scarier part. It was all a little too fine.

What AI Became Good For

Once I stopped accepting first drafts as artifacts, AI became much more useful.

I use it best as a pressure tool:

• Find the vague part: “What claims in this lesson are unsupported?”
• Generate the wrong answers: “What would a smart beginner pick for the wrong reason?”
• Change the audience: “Explain this to a React developer who has never used Postgres.”
• Create contrast: “Give me three examples where this advice fails.”
• Compress the mess: “Turn these notes into an outline without adding new claims.”

Those prompts do not replace judgment. They make judgment easier to apply. The model is very good at producing surfaces. My job is to decide which surfaces deserve to become structure.

The Pipeline, With Guardrails

The workflow that survived is less glamorous than the old version:

1. Collect raw material I start with notes, code, links, screenshots, and the specific learner confusion I am trying to address.

2. Ask AI to organize, not decide It can group ideas, reveal missing steps, and suggest sections. I do not let it choose the thesis without a fight.

3. Write the sharp parts myself The opening, the decision rule, the examples, and the ending need human taste. Those are where the article earns trust.

4. Use AI to interrogate the draft I ask what sounds generic, what assumes too much, what a skeptical reader would object to, and where the examples fail.

5. Keep the artifact honest If a paragraph could live on any SaaS blog, it gets rewritten or deleted.

That last rule is doing a lot of work.

Quizzes Changed My Mind

Quiz writing is where AI became genuinely useful for me.

Not because it can write perfect questions. It cannot. It loves obvious distractors, accidental ambiguity, and explanations that glide past the misconception.

But it can generate a field of possible wrong answers quickly. Then I can look at the list and ask: which of these represents a real learner mistake?

That is the useful collaboration. The model produces clay. I decide whether it is a bowl, a brick, or landfill.

Replacing Myself Was The Wrong Frame

The title is a joke, but the frame is wrong.

I am not replacing myself with AI. I am replacing the parts of my process that were already mechanical: first-pass grouping, alternate phrasings, translation scaffolds, draft quiz options, image concepts, summary passes.

The parts I cannot replace are the parts readers actually notice when they are missing: taste, scars, priority, skepticism, and the willingness to say “this sounded good, but it was fake.”

AI made me faster. More importantly, it made some of my weak spots easier to see. The cost is that I now have to be more deliberate about not publishing prose that merely behaves like prose.

That is the bargain I can live with.