Three things I delete from almost every AI draft

The fake origin story opener

The first thing I usually delete is the opening paragraph.

Here is a real one, from a draft about my CSS DevTools workflow:

"I have been fascinated by web development since I first opened a browser. Over the years, watching the web evolve from simple static pages to complex interactive experiences has only deepened my passion for front-end craftsmanship."

This sounds fine if you skim it. It is also wrong.

I am 57. I dropped out of art academy. I learned HTML by breaking things until they looked less broken. There was no childhood browser epiphany. No carefully curated path from GeoCities to senior front-end engineer. The paragraph is not just inaccurate; it establishes a voice that is 25 years old, freshly graduated, and writing a LinkedIn origin myth.

The model is not trying to lie about my life. It is doing pattern completion. It has seen thousands of blog posts that start with:

• "I have always been interested in..."
• "Ever since I was a kid..."
• "My journey into X began when..."

So when you ask it for an article, it gives you an article-shaped object. Articles, in its training data, often begin with a personal origin story. Whether that story matches the actual person is not part of the objective.

Why I delete it:

• It wastes the first 150 words. If the piece is about CSS DevTools, the first paragraph should contain the word "DevTools," not a childhood memory the model invented.
• It sets the wrong age and tone. You can feel the imaginary 25-year-old in a co-working space, sipping cold brew and discovering React. That is not me, and regular readers can tell.
• It makes everything that follows less trustworthy. Once the first paragraph is fake, you start wondering what else was made up for narrative convenience.

Most of the time I do not edit these openers. I select the whole paragraph and remove it. Then I start with the concrete thing the post is actually about, even if that means opening on a sentence that sounds slightly abrupt. Abrupt and true beats smooth and fabricated.

The forced AI angle

The second thing I hunt for is the word "AI" itself.

I wrote about this earlier in Fluent emptiness and the missing Tuesday morning.

At one point my Make.com setup was over-optimized for SEO goals. The system prompt kept nudging the model toward "AI" as a magic keyword. The result: every draft tried to smuggle in an AI angle, even when the topic had nothing to do with it.

Here is a paragraph from a draft about CSS Grid:

"AI can help visualize grid structures and suggest optimal layouts based on your content. By combining CSS Grid with AI-powered design tools, you can rapidly prototype responsive designs without getting bogged down in manual tweaks."

This is a post about a layout technology in the browser. I was talking about grid-template-columns, Firefox DevTools, and the little overlay that shows you which cell you are in. Dropping a paragraph about "AI-powered design tools" into the middle of that is a non-sequitur.

The model did not discover a deep connection between CSS Grid and generative models. It was trying to please the invisible algorithm it was primed to care about. Somewhere upstream, it had learned that mentioning AI is good for engagement, so it wedged it in.

Why I delete it:

• It breaks the focus. If you are learning CSS Grid, you do not want to detour into speculative tooling. You want to know where your columns went.
• It smells like keyword stuffing. Readers can tell when a paragraph is there for Google, not for them.
• It confuses the scope. Not every technical topic needs to be reframed as an AI story. Sometimes a layout system is just a layout system.

My editing rule is simple: if the post is not about AI, any paragraph that suddenly is about AI goes in the bin. I run a quick scan for the word "AI" and check every occurrence. Most of them do not survive.

The invented moral-of-the-story

The third cut usually happens near the end.

Language models like a tidy conclusion. They want to send you home with a lesson, often in three parts. Even when there is no lesson.

In a draft about my Withings scale, the model added this:

More on this in my article Most U12 Baseball Practices Are Broken: Drill Reps vs Game Reps.

"Over time, daily data taught me to trust the process over the number. Watching the small fluctuations added up to a bigger picture: consistency matters more than any single weigh-in. This simple routine quietly reshaped my mindset about progress."

It sounds reasonable. It is also not what happened.

What I actually learned is that the number on the scale is mostly water and I should ignore it until Friday. That is not a mindset shift. It is a scheduling adjustment.

The model invented a moral because that is how essays in its training data often end. There is a personal struggle, some data, then a neat summary about trusting the process or embracing uncertainty. It is narrative muscle memory.

Why I delete it:

• It rewrites my experience. A small practical insight turns into a self-help parable I do not believe.
• It inflates the stakes. Not every gadget, habit, or script "quietly reshapes" anything. Sometimes it just saves you five minutes.
• It drags out the ending. The post should end where the point is made, not where the algorithm finds redemption.

There is a related pattern: the recap paragraph that restates the article in the past tense. It usually starts like this:

"In this post, we explored how daily weigh-ins with a smart scale can transform your relationship with data and progress. From understanding fluctuations to reframing your goals, these lessons can help you..."

I delete this without reading the rest. If the reader needs a recap, the post failed. If the post worked, the recap is insulting. It is there because academic and journalistic formats like tidy conclusions. On my blog, it just signals that a machine wrote this.

The title that belongs on BuzzFeed, not my site

One more cut happens before all of this: the title.

Read also Why my posts now start with five bullets.

Left to itself, the model happily generates things like:

• "My Exact Morning Routine for Peak Performance"
• "The Real Cost of Automating Everything"

These are not my titles. They are a pattern the model learned from ranking blogs and productivity threads. You can almost see the YouTube thumbnail.

My actual titles are usually six to eight words and understate the claim. They describe what the piece is about instead of promising a life upgrade.

So the first thing I do with an AI draft is select the title and press delete. Not because SEO is bad, but because I am writing for humans who already know roughly who I am. They do not need "The Real Cost of" anything. They just need to know what they are about to read.