If you scroll to the bottom of each post on the site, you may notice a curious badge:
It's a badge from the Not By AI group, a project that looks to explicitly tag creative content—writing, visual art, music—as created by humans, not machines. Whereas other mechanisms like SynthID relies on watermarks from the GenAI provider and apps like Pangram look for telltale signs of generated text, this effort relies on simple disclosure to draw an explicit line.
For this blog, the badge is to signify that the words on the page were not predicted by stochastic parrots. But that's not to say that AI isn't involved in its production; in fact, I've been using this site as an excuse to play around with the latest AI models, from my Wordpress migration to testing the image generation capabilities of Midjourney and Google Flow. It's also useful as a research partner, grammar checker, and tireless brainstorming zealot. The goal is to achieve the Not By AI 90% rule, where the content cannot exceed 10% of the final output.
Here's how my usage evolved over months of trial-and-error:
- Early on, using LLMs for simple grammar and spell checking, merely replicating the functionality of Grammarly;
- With the release of Claude Writing Styles, using my posts here as a corpus of post-training data and creating a writing profile;
- Generating and revising post outlines;
- Asking the AI to critique works in progress, from paragraphs to sections to drafts;
- Further asking for a concrete numerical score and areas of improvement.
Within each step, there's an opportunity to run an optimization loop. That is, I can ask the AI to run through a step repeatedly, to hone in on grammatical nits and refine the writing flow, with incremental tweaks and feedback along the way to nudge the prose forward. Eventually, we will arrive at the local maximum, and Claude concludes with:
No notes left to add. Publish it.
Endearing. Albeit also patronizing.
But now I find myself staring at these pieces of writing that—while the words and structures and examples and insights are still my own—seem to have lost something in the AI-assisted refinement process.
Some of this feeling comes from AI's propensity to follow best practices to a fault. It's akin to running the prose by my high school English teacher, who happens to be a disciple of the 9th grade essay template that goes introduction → thesis → three pieces of supporting evidence → conclusion. Every time we go for another round of critiques and suggestions, my posts hone closer to this platonic ideal.
The bigger piece I've been missing, though, by outsourcing segments of the writing process to AI is the struggle to find the right words.
Here, I'm talking about an ability beyond finding technically correct words and sentences. The rhythm and texture of a sentence is something individual to the author; rhetorical devices like similes and metaphors, sharp examples and weird inclusions are deeply personal and form organic connections from the human author to the human reader. This is still something that current generation LLMs cannot produce, though Claude is often self-aware enough to point out these inclusions and urges me to keep the references.
The book On Writing Well hones in on precisely this element. The author, William Zinsser, wrote all of his articles and books well before any form of AI assistance[1]. He was adamant that the work of writing well required real work, the iteration and repetition and refinement of singular sentences until they passed his stylistic bar. This was the craft of writing, of using the totality of the language to entertain and delight—and to not shy away from the toil and frustration and difficulty of the task.
Another way to frame Zinsser's perspective is that he's advocating for a depth of thinking at the most granular levels of prose. Years ago, I lamented the loss of thinking because it became trivial to create pages of text with a single prompt. There was so much low-effort content that we added the term "slop" to the vernicular, and it received the dubious distinction of Merriam-Webster's word of the year. But I'm finding that even when I restrict AI usage to only certain aspects or pieces of the writing process, risks the same outcome: technical competence, stylistically dull prose.
There are certainly elements of the process that have been either sped up or tackled more thoroughly with the help of Claude and ChatGPT. But every invocation increases the chance that the stochastic editor steps too far, makes it too easy to accept an adjacent logical suggestion. Ultimately, there is no detection app or recognition algorithm to catch when the underlying structures and thought processes were produced with artificial intelligence. We're measuring what we can access, the sentences and paragraphs and comparing them against known AI-generated tics. Even Not By AI is laying its claim on the final output, of individual words written by human hands.
But there's perhaps a bigger, more audacious claim: that the scaffolding to the prose has also been driven with human effort.
Which means I'm not done developing an earnest, complementary, AI-augmented writing workflow quite yet.
His concession to cutting edge technology was moving from a mechanical typewriter to a word processor. ↩︎
