I built a blog that wrote itself. Then I turned it off.

The setup was simple. A nightly cron pulled the day's news into a Postgres table. A small Laravel job picked the top three stories, asked an AI to draft a one-page brief, then passed the brief to a newer-generation model that wrote the article with proper meta tags and published it to a blog. Twice a day, a new post. No human in the loop.

For two weeks it worked. The posts were grammatical, on-topic, properly tagged. I felt like a wizard. I tweeted a screenshot.

Then the posts started rhyming

By week three the articles were echoing each other. Same opening hooks. Same metaphors. The model was drawing from a shrinking well — the news samples were similar enough that the briefs were similar, and the briefs being similar made the outputs converge.

By week five I could predict the structure of any post by reading the title.

What I should have caught earlier

Automation without fresh, varied input always regresses to the mean. There is no exception. If the input distribution narrows, the output distribution narrows faster, because the model amplifies whatever signal it can find. The fix is not "better prompt." The fix is "more interesting input."

I turned it off. The blog is a graveyard with around 60 posts I can't tell apart. Useful experiment, useless artifact.

The lesson rhymes with one I keep relearning: automate the boring part, leave a human at the interesting part. The interesting part on a blog is the angle. The angle is exactly what the cron couldn't generate.