How I use AI on this blog

In September 2026, I reworked my complete blog:

• Design: This blog has a new theme. It has a dark mode, works better on phones, has a better search page, and images open in a lightbox where you can page through galleries.
• Old articles: I imported my articles from Medium. Then I went through all of the roughly 900 articles: I fixed typos, grammar, broken formulas and markup, added alt texts to images, removed tracking parameters from links and cleaned up the tags.

I could not have done this in a few weeks without AI. I mainly used Claude. That's why I want to be transparent about what AI does on this blog and what it doesn't.

Images: "AI generated" and "AI modified"

Images that an AI tool made have a small badge in the top right corner:

• AI generated: The whole image comes from an AI tool. So far, I've used Claude, ChatGPT and Gemini. I want the caption to name the tool as well. I'm still adding that to older images.
• AI modified: A real photo that an AI tool changed. Cropping a photo or adjusting its brightness in a normal image editor doesn't count.

Images that don't get a badge:

• My own photos and screenshots, unless an AI tool changed them.
• Diagrams: charts, flow charts, network diagrams, pyramids and maps with data. They don't pretend to be a photo, so nobody takes them for real. If an AI tool made the diagram, the caption still says so.
• Charts from data: When Claude writes the Python code for a chart, the numbers come from the cited data, not from the AI. The caption says "created with Claude AI in Python/matplotlib" and names the data source.

The EU AI Act only requires a label for deep fakes: images that look like real people, places or events and would falsely appear authentic^[1]. I label more than that, because I think you should know when a picture wasn't taken by a camera.

Copy editing

AI proofreads my articles. It's a bit more than a spell checker, more like copy editing:

• Fixed directly: typos, grammar, broken formulas and markup, obvious calculation errors and dead links.
• Only flagged: outdated facts, missing sources, unclear structure, a weird tone. AI adds those to a list, and I decide what to do with them.
• Research: AI looks up missing numbers and checks facts, but only with a source that I can follow.

I wrote down the rules for this blog in a file that the AI reads before it changes anything: how to format sources, which tags to use, when an image gets a badge.

I check everything

Nothing goes online that I haven't checked. Every change is a diff in Git, so I can see exactly what the AI changed before I publish it. If an article is wrong, it's my mistake, not the AI's.

No AI slop

I will not generate AI slop. I don't see a point in doing so: there is enough generic text on the internet, and I write because I want to share what I think and what I found out.

But I might use AI to rephrase a clumsy sentence, fix issues or find a better formulation. The ideas, opinions and experiences are mine.

Spell checker, search engine, talented intern

I use AI like three tools I already knew:

• A spell checker: It finds typos and grammar mistakes, and it suggests better formulations.
• A search engine: It finds sources and data much faster than I could. But just like with a search engine, I read the source before I trust it.
• A very talented intern: It is fast, knows a lot and does tedious work without complaining. But it also makes mistakes with full confidence. So I give it clear instructions, and I review its work.

Footnotes