This morning a customer sent over a 1,200-word draft about their new analytics dashboard. Second sentence: “In today’s rapidly evolving landscape of data-driven decision-making, organizations are constantly seeking innovative solutions.” I closed the tab, finished my coffee, and wrote back asking who actually wrote it. The reply was sheepish. ChatGPT had done the first pass, someone had “edited” it, and they were hoping we’d catch anything weird before it went live.
We caught a lot of things weird.
I review drafts like this most weeks. Marketing leads, founders, occasionally an agency pretending they didn’t outsource the outsourcing. Some of the work is genuinely good. A meaningful chunk of it is generated, mildly tidied, and pushed live. I have nothing against AI drafts in principle (our team uses models constantly for outlines, research, and first cuts), but there’s a gap between treating a model as a drafting partner and handing it your brand voice with the keys still in the ignition. The handed-off drafts have patterns. These are the five I flag most.
1. Generic openers that could belong to anyone
The throat-clearing first sentence is almost always the giveaway. Grammatically fine, factually inoffensive, and equally applicable to a CRM, a logistics startup, or a company that sells artisanal dog food. The model doesn’t know what’s specific about your business, so it reaches for the median opener it’s seen ten thousand times.
Generated: “In the competitive landscape of enterprise software, companies are constantly seeking innovative solutions to streamline their operations and drive growth.”
What a human would write: “We rebuilt our customer onboarding flow three times last year because the original design assumed users actually read instructions.”
The first one could open anything. The second one tells me what you do and that you’ve learned something painful doing it.
2. Lists where every item gets equal airtime
When I read a “7 ways to improve X” piece and every item gets the same two paragraphs, same definition-then-benefit structure, same confident tone, I assume a model wrote it. People who’ve actually done the work don’t write like that. They overweight the thing that worked. They dismiss the thing that didn’t. They spend three paragraphs on the channel that drove 80% of revenue and two sentences on the one that flopped.
Take marketing channels. If I spent eight months on SEO and two weeks on a podcast experiment that went nowhere, those shouldn’t get equal space in my writeup. The model gives them equal space because it has no idea which one mattered to me. Uneven attention is what expertise actually looks like on the page.
3. The same six transition phrases, over and over
“Let’s dive into.” “It’s important to note that.” “In today’s fast-paced.” “A comprehensive guide to.” “Leverage.” “Navigate the complexities of.”
I keep a running list. Three or more of these in one piece and I stop reading for content and start reading for provenance.
The phrases aren’t wrong exactly. They’re the connective tissue a model reaches for when it needs to bridge two ideas and doesn’t have a stronger one. A writer who’s lived with their topic builds weirder bridges: a callback to something five paragraphs earlier, a half-joke, an aside about a customer who did the opposite of what the article recommends. Models don’t have that material, so they reach for the safe connector. Once you’ve calibrated on it you can’t uncalibrate, which is mildly annoying because it ruins a fair amount of otherwise-decent writing.
4. Confident statistics with no source attached
This one actually worries me, and not just aesthetically. There’s a liability angle. A draft will say something like “Studies show that companies using AI-driven analytics see a 34% improvement in decision-making efficiency.” Which studies? Run by whom? Published where? The number is plausible enough that nobody questions it, which is the whole problem.
The model isn’t lying on purpose. It’s predicting what kind of number tends to sit in that sentence position based on its training data. That’s the generous reading. The less generous one is that it invented a stat because the sentence wanted one.
When we cite numbers in our own work, we link them. If your draft has a number and you can’t find where it came from in five minutes, cut the number and write a qualitative sentence instead. “This tends to improve decision-making” is a worse sentence than “34% improvement in decision-making efficiency,” but it has the advantage of being something you can actually defend.
5. The “not X, but Y” rhetorical trick, used as filler
I catch myself doing this one too. “It’s not just about speed, it’s about clarity.” Used once, it’s a sharp pivot. Used four times in 800 words, it’s a model trying to sound profound. The structure flatters the reader into thinking a real distinction is being drawn, when often the two halves are basically the same idea wearing different hats.
My guess is that models lean on it because it scores well on whatever pattern-matching they were tuned on. It sounds like insight. Insight is supposed to surprise you, though, and if the second half of the sentence is what you’d have guessed anyway, the rhetorical move is doing work the idea should be doing on its own.
So what do you actually do with this
If you’re running a content pipeline, build a checklist. Before anything ships: does the opener name something specific to the business? Are list items weighted unevenly, the way a person who lived the work would weight them? Are stats sourced and dated? Are the same transition phrases recurring? If three or more of those fail, send it back for a real rewrite rather than a surface edit.
I don’t have a clean number for how much of the open web is generated now, and I distrust the ones I’ve seen (see rule #4). Directionally, more of it is generated every month, and readers are getting faster at smelling it. The thing I’m working on next is a tighter version of the checklist above, baked into our review tool so the obvious tells get flagged before a human reviewer ever opens the doc. If that works, I’ll write it up. If it doesn’t, I’ll probably write that up too.