Fluent, Correct, and Completely Forgettable: Writing AI Content That Doesn't Sound Like AI

2026-07-29 · Glenn Fratangelo

Everyone worries that AI will make things up. That's the wrong worry.

Hallucinations are loud. They get caught, usually by someone in the comments, and the cost is a bad afternoon. The quieter failure is the one that actually damages you: AI writes something entirely correct and entirely forgettable, you publish it, and nothing happens. No one argues. No one remembers. If you want AI content that doesn't sound like AI, the thing you have to fix isn't accuracy — accuracy was never the problem.

Originality.AI analyzed 3,368 long-form LinkedIn posts from 99 influential profiles published between January and November 2025 and judged 53.7% of them likely AI-generated. More than half the long posts in the feed. Which means the generic version of your idea has almost certainly already been published, several times, this week.

Correct is not the same as credible

Nobody reads a LinkedIn post to check facts. They read it to work out whether you know something they don't.

A post listing the five benefits of a technology is not wrong. It is simply evidence of nothing. It proves you can summarize the field, which any model can do, and which tells a reader precisely as much about your judgment as a well-formatted table of contents. Credibility doesn't come from being right about the general case. It comes from specificity — the typology you watched form before anyone named it, the number that surprised you and changed how you staffed a team, the thing you argued for two years ago and no longer believe.

A general-purpose model, prompted generally, regresses to the average of everything ever written on your topic. That average is smooth, competent, and the one output that carries no information about you at all. It is the literary equivalent of a stock photo of a handshake.

The engagement data says more than the trust data

The interesting part of that same study is where AI content wins and loses. It outperformed human-written posts in leadership and inspiration content by about 75%. It underperformed badly elsewhere: human posts pulled roughly 80% more engagement in innovation and strategy, 73% more in marketing and branding, 44% more in healthcare, 40% more in government and public affairs.

Read the pattern rather than the numbers. AI does well where the content is supposed to be generic — motivational writing has always been a genre of pleasant universals, and a machine is good at pleasant universals. It loses, badly, in every category where the reader came looking for one specific person's judgment about a hard problem.

That's the trade, stated plainly. The more your field rewards genuine expertise, the more the average-sounding version costs you. If you work in compliance, financial crime, security, or any domain where your buyers are technical and skeptical, you are in the category where generic AI content performs worst. You are also, not coincidentally, in the category where it's most tempting, because the writing is a chore and the deadline is real.

What AI content that doesn't sound like AI actually requires

Not prompt tricks. Not asking for a "conversational tone" or telling it to avoid the word "delve." Three things, and they're all upstream of the model.

A position someone could argue with. If a reasonable peer in your field couldn't disagree with your post, you haven't made a claim. You've written a definition. Start from the thing you believe that your industry hasn't fully accepted yet, and the writing gets easier immediately.

Evidence only you have. Not a public statistic — I used one above, and it's seasoning, not substance. I mean the case you worked, the pattern across the last dozen alerts, the reason a control failed that isn't in any vendor's white paper. This is the part that cannot be generated, and it's the entire reason anyone follows you.

Your actual sentence rhythm. People recognize a voice faster than they recognize an argument. If your posts suddenly acquire tidy rule-of-three structures and transitions you'd never say out loud, readers notice before they can explain what they noticed.

There's a growing body of research this year on what's being called the AI penalty: when audiences learn that content was AI-assisted, perceived authenticity and trust drop, even when the quality of the writing holds steady. The lesson people take from that is usually "don't disclose." I think that's exactly backwards. The lesson is that the trust is being extended to you, personally, and so the substance underneath the writing has to actually be yours. Do that and disclosure costs you very little. Skip it and no amount of concealment helps.

Why this matters more than it did a year ago

Scarcity moved. Competent prose used to be a real barrier — it kept most experts off the feed entirely, which is a problem I care about a lot. That barrier is gone, and it's not coming back.

But when everyone can produce fluent, correct paragraphs on demand, fluent correct paragraphs stop being worth anything. What's left as the scarce good is a defensible point of view, held by a named person, expressed in a way that couldn't have come from anyone else. Two years ago, publishing consistently was itself a differentiator. Now it's the price of entry, and the differentiation has moved entirely to whether you said something.

This is the assumption Prowir is built on: the model should be handling the mechanical work — structure, cadence, formatting, showing up on schedule — while the position and the evidence stay stubbornly yours. That's a harder product to build than a text generator. It's also the only version worth building.

Use AI to publish more. Don't use it to have opinions.

— Glenn