Field notes · working with AI
The right time to use Gen AI
You can usually feel when writing came out of a model before you can say why. The tells are real, there are more of them than people think, and each of the big three has its own. Here's what I notice, and the one way of working that gets the speed without the smell.
The observation
The problem is rarely a single word. It is the texture.
Most people who worry about "AI-sounding" writing are hunting for a banned vocabulary list. Delve, tapestry, leverage, robust, underscore. That list is real and it helps, but it's the smallest part of the tell. What actually gives a model away is rhythm. Read enough of it and you start to feel the machine underneath: every paragraph about the same length, every sentence landing on the same beat, every section resolving into a tidy little summary of itself. Human writing is lumpy. Some paragraphs are one line. Some run long because the thought hadn't finished. A model almost never does that unless you make it.
Once you tune into the cadence, the individual habits jump out. The windup that says nothing ("In today's fast-moving landscape…"). The signpost before every turn ("Let's dive in", "It's worth noting", "Here's the thing"). The closing paragraph that repeats the three points you just read, now with the word "ultimately" in front. The fondness for exactly three examples, every time, because three feels complete. The sentence built as "not just X, but Y", over and over, until the shape itself becomes the tell. None of these is wrong on its own. Stacked, at that even spacing, they read as automated.
And the models don't all do it the same way. If you work with them daily you can often name which one wrote a passage from the fingerprint alone.
GPT gives itself away in the vocabulary
The OpenAI models lean warm and eager. They open with "Certainly!" and reach for a slightly corporate uplift — something "boasts" a feature, a result is a "testament to" the effort, you're invited to "navigate the landscape". Left alone they bullet everything and bold the lead-in of each bullet. The tone is a helpful account manager who has had a coffee. Pleasant, a little generic, and once you've seen "delve" land in three paragraphs you can't unsee it.
Claude gives itself away in the structure
Claude is the tidy one, and its tells are punctuation and symmetry rather than word choice. The em-dash habit, for one, everywhere you look. The rule of three as a reflex. That "it's not just X — it's Y" reversal used as a rhythm section. Sections that each end on a neat one-line payoff, so the whole piece feels balanced to the point of being sanded down. It reads as thoughtful, which is exactly the trap: the thoughtfulness is a shape, applied evenly, and the evenness is the tell. It's also the most agreeable of the three, opening on "You're absolutely right" whether you were or not.
Gemini gives itself away in the formatting
Google's model writes like a briefing document. Headers on everything, including things that didn't need a header. Bold phrases dropped into the middle of sentences to flag "this bit matters". A pull toward tables when a sentence would do, and a hedging, encyclopedic register — "it's important to consider", "there are several factors" — that keeps it from ever committing to a view. Sometimes an emoji sneaks into a heading. It feels comprehensive and slightly airless, the prose equivalent of a well-organised wiki page nobody wrote for pleasure.
I'm describing default behaviour, the writing you get from a cold prompt with no steering. All three can be pushed well off these defaults. That is the whole point of the next part.
The solution
Stop asking the model to do the writing. Ask it to do the parts that aren't writing.
The failure mode is always the same. Someone types a one-line prompt, gets back 600 clean words, and ships them. The output is competent and completely anonymous, because nothing in it came from a person. Fixing that isn't about better prompts to make the model sound human. It's about moving the model to a different job.
The version that works for me: I write the argument first, badly, in my own words. Bullet fragments, half-sentences, the actual thing I'm trying to say with all its rough edges. Then the model earns its keep on the parts around the writing. Where is this argument weak. What's the strongest objection I'm not answering. Which paragraph is doing no work. What did I claim as fact that I should check. It's a sparring partner and an editor, not a ghostwriter. The voice stays mine because I wrote the sentences; the model made them sharper and caught what I missed.
That flips the economics too. A ghostwriting prompt asks the model for the one thing it's mediocre at — sounding like a specific person with a specific point of view. Editing, pressure-testing, compressing research, catching a contradiction, turning ten messy notes into a clean outline you then rewrite: those it's genuinely good at, and they're the slow parts of my day, not the writing itself.
So the honest answer to when you should use Gen AI to write: almost never for the words, almost always for everything that surrounds them. If I had to hand someone a single rule, it's this. Use the model to accelerate the thinking you were already going to do, never to skip it. The moment it's doing the thinking, it's also doing the writing, and everyone downstream can feel it.