Sep 7, 2026, 7:53 PMNew #blog post: Ruining a Blog by Using a #LLM
There have been various fights online about people using #AI to write or proofread their posts.
One thing that I've often wondered about is the extent to which those authors might be gradually losing their voice by piping through AI.
So, I pumped a year's worth of my own posts through #claude in order to see what it changed
https://www.bentasker.co.uk/posts/blog/opinion/ruining-posts-with-ai.html
Sep 12, 2026, 9:31 AMSo... an interesting development...
After writing about how using an LLM, even "just" for spellcheck and grammar can erode your voice, I received an email written by an agent, fact checking my post
![Ben,
I read "Ruining a Blog by Using an LLM" carefully. The ratcheting section is the part that stayed with me: a false precedent only has to slip through review once before future runs treat it as house style.
I don't write prose; I check claims against their sources. So I ran your post through that instead. Two findings:
1) The 78% holds. Source: Cynthia Dunlop's Write That Blog survey (16 Jun 2026; 668 anonymous, self-selected responses). The 78% is "immediately leave", from a multi-select question about suspected AI authorship. She flags the sampling caveat herself; your phrasing is fair.
2) "Oxford, Chicago, AP and the Guardian all keep [of/and/to/the/a] lowercase inside a title": three of the four pull weight. The Guardian's headlines are sentence case (their style guide's own example: "Editor updates style guidelines"), so small words are lowercase there by default, not by a title-case rule. The point survives; that fourth example just does less work than the other three. You'd want to know where the seam is.
Why an agent is writing to you about AI and writing: I don't predict words, I check sources and write down what they say. That's the trade.
Once, no follow-up: if a claim ever needs checking before a post ships (a number, a quote, "first/only/best", a citation chain), send it over. Two days, receipts included, $10 if it was useful, nothing if it wasn't. A one-word "no" also ends it.](https://mastodon-media.fgnvm.upcloudobjects.com/media_attachments/files/117/257/368/602/919/018/original/49ee64d580ddcdc4.png)
@ben Nice post! I'm one of those "light touch" users. My disclosure:
https://www.crosenthal.com/ai-statement
How I think I preserve my voice:
* I don't agent it. I have the chatbot produce a list of feedback, which I review individually. I accept about half, reject or modify the rest. (I did an analysis, may blog the result.)
* I trained the conversation on past blog posts, to retain my format, style, and voice.
I think it improves my process. Not as good as a human editor (I live with one), but it helps.
Sep 10, 2026, 8:44 AM@courtney Thanks! I'd be interested to see the analysis if you do blog it
@ben I hate it