Revising to avoid sounding like AI in fear of being AI-shamed is one of those things that should not be a problem but is. Society has to stop AI shaming blindly as the fact that something was generated by AI does not mean it is bad or even incorrect. The creator could have done lots of research and proofreading, just that rewriting of not their . But on the other hand, we have no control over our critics and as long as society continues to AI shame, creators have no choice but to invest time in rewriting.
Yes, now whenever I naturally write an em dash, I hesitate (but usually just go for it anyway). When most people see a hint of AI, they're out instantly. In most cases, yeah, it's pure slop with zero information gain. But in some cases, it's someone who has never written before but -- with LLM assistance -- decided to share their experiences for the first time. Or it's someone who worked super hard to write the article from scratch -- even if it's not their native language -- then asked an LLM to "polish it" or "correct the English." As much as I hate the "style" of AI writing, there are nuances to consider.
Thanks for this super-interesting write up. I know devs who only use LLMs to write, leading to AI-sounding articles, and those who don't touch it with a 10 foot stick (and still use LLM for research).
It's rare for authors to disclose how they use LLM in their writing. I am curious about that as a reader and fellow writer.
Can you share how you generated these charts? Are these custom built or is it part of existing software?
What struck me is that AI-scented technical writing is not one failure mode. Sometimes it is retrieval, sometimes incentives, sometimes review, sometimes a missing owner for the claim. That is why I find it useful to ask where in the stack the failure actually entered. I wrote about that failure-map approach here: https://petermccannstrain.substack.com/p/the-nine-layers-where-agents-break
Very interesting data! I wonder whether the bias against "AI-scented" writing reduces LLM use. But then again, it's not too difficult to write with an LLM and then edit to make it sound less AI. Removing em dashes, planting spelling errors, and such.
Revising to avoid sounding like AI in fear of being AI-shamed is one of those things that should not be a problem but is. Society has to stop AI shaming blindly as the fact that something was generated by AI does not mean it is bad or even incorrect. The creator could have done lots of research and proofreading, just that rewriting of not their . But on the other hand, we have no control over our critics and as long as society continues to AI shame, creators have no choice but to invest time in rewriting.
Also, Vincent's (super recent) article is a good read: https://vincent.bernat.ch/en/blog/2026-blogging-llm
Yes, now whenever I naturally write an em dash, I hesitate (but usually just go for it anyway). When most people see a hint of AI, they're out instantly. In most cases, yeah, it's pure slop with zero information gain. But in some cases, it's someone who has never written before but -- with LLM assistance -- decided to share their experiences for the first time. Or it's someone who worked super hard to write the article from scratch -- even if it's not their native language -- then asked an LLM to "polish it" or "correct the English." As much as I hate the "style" of AI writing, there are nuances to consider.
Thanks for this super-interesting write up. I know devs who only use LLMs to write, leading to AI-sounding articles, and those who don't touch it with a 10 foot stick (and still use LLM for research).
It's rare for authors to disclose how they use LLM in their writing. I am curious about that as a reader and fellow writer.
Can you share how you generated these charts? Are these custom built or is it part of existing software?
Wrestling with Claude ;) I did analysis in Gemini and Claude and ChatGPT (to triple check it), then created charts with Claude.
What struck me is that AI-scented technical writing is not one failure mode. Sometimes it is retrieval, sometimes incentives, sometimes review, sometimes a missing owner for the claim. That is why I find it useful to ask where in the stack the failure actually entered. I wrote about that failure-map approach here: https://petermccannstrain.substack.com/p/the-nine-layers-where-agents-break
Very interesting data! I wonder whether the bias against "AI-scented" writing reduces LLM use. But then again, it's not too difficult to write with an LLM and then edit to make it sound less AI. Removing em dashes, planting spelling errors, and such.