You know the draft I mean. You asked ChatGPT for a blog post, and thirty seconds later you had one: fluent, grammatical, correctly structured, five tidy H2s, a conclusion that "ultimately" depends on your specific needs. Nothing in it is wrong. And nothing in it is yours. Swap your logo for a competitor's and the post still works — which is the whole problem, because content that any company in your category could have published is content that does nothing for yours.
If you've felt that flatness and wondered whether you're prompting badly, this post is for you. You're not prompting badly. You're using a tool exactly as designed — and the design explains both the flatness and the fix.
You've read this post before
Generic AI content isn't randomly bad; it fails in a recognizable pattern. Once you see the pattern, you can spot a blank-box draft from across the room — and so, increasingly, can your readers:
The five tells of blank-box content
- Hedged claims. Every assertion arrives pre-softened: "can be a great option," "it's important to consider," "results may vary." Nothing is staked, so nothing can be wrong — or memorable.
- Symmetrical listicles. Five reasons, each exactly one paragraph, each given equal weight, as if the writer had no view on which one actually matters. Real experts are lopsided.
- The throat-clearing opener. "In today's fast-paced digital landscape…" A sentence that could preface any topic is a sentence that says nothing about this one.
- No enemy. A real point of view is against something — a common practice, a competitor's approach, a piece of conventional wisdom. Blank-box drafts offend no one and therefore persuade no one.
- No stakes. Nothing in the draft suggests the writer would lose anything by being wrong. There's no skin in it — no "we tried this and it failed," no number the author is accountable for.
None of these are grammar problems, which is why "make it punchier" edits never fix them. They're all the same problem wearing five outfits: the draft has no author.
Why does ChatGPT content sound generic?
Because a language model predicts the likeliest next words given its input, and given a generic prompt, the likeliest output is the average of everything ever written on the topic.
That's the honest, mechanical answer, and it's worth sitting with, because it contains no AI-bashing. A large language model is trained on an enormous share of the written internet. When you type "write a blog post about employee onboarding" into an empty chat, you've given it a topic and nothing else — so the statistically safest completion is the consensus of every onboarding post in its training data. Not the best one. Not the most opinionated one. The central one, with the sharp edges of every individual writer sanded off by the averaging.
Here's the reframe that changed how I think about this: the median is the design goal of a blank box, not a bug. When the model knows nothing about you, writing the most broadly probable version of the post is exactly the right behavior — it's what "predict the likely text" means. The old programmer's complaint was garbage-in, garbage-out. That's not what's happening here. ChatGPT's output isn't garbage; it's polished, competent, and empty. The problem isn't garbage-in. It's nothing-in.
This is also why the model is genuinely brilliant the moment you give it something real to work with. Paste in a messy transcript of you ranting about your industry and ask it to structure the argument, and the output is often excellent — because now it's predicting the likeliest continuation of your thinking, not the internet's average. The tool didn't change. The input did.
What the model never knew about you
Look again at the five tells, and notice they're all absences. The hedging exists because the model doesn't know which claims you'd defend. The missing enemy exists because it doesn't know who you're positioned against. Specifically, a blank chat has never been told:
- Your positioning — who you serve, what you refuse to be, and why a buyer should pick you over the two alternatives they're already considering.
- Your opinions — the takes you'd argue for on a sales call, the industry consensus you think is wrong, the advice you give even though it's unpopular.
- Your recent reality — what you shipped last month, what a customer said last week, the objection you keep hearing. The specifics that make content sound alive.
- Your voice — how you actually talk: your sentence rhythm, the words you'd never use, whether you're blunt or wry or careful.
In theory you could type all of that into the prompt. In practice, nobody does — it's pages of context, it doesn't fit in a prompt you're retyping at 11pm, and even when you paste a chunk of it, the chat forgets everything the moment it ends. Tomorrow's session starts from zero. So every draft is written by a stranger meeting you for the first time, and it reads like one.
So how do you make AI sound like you?
Capture your voice, positioning, and opinions once in a persistent layer, feed every draft from that layer, and keep a human editing pass for judgment.
Notice what's not in that sentence: prompting tricks. "Act as a world-class copywriter" doesn't add information; it rephrases the request. The fix is not a cleverer instruction — it's context, supplied durably instead of retyped per session. Concretely, three moving parts:
First, capture once. Write down the things the model can never guess: your one-sentence positioning, who you're against, the five opinions you'd defend under pressure, three examples of writing that sounds like you and one that doesn't. This is an afternoon of work, and it's the highest-leverage afternoon in your content operation.
Second, feed every draft from it. The manual version is a style-guide document you paste into every new chat before you ask for anything — clunky, but it works, and if you do nothing else after reading this post, do that. The systematic version is a layer that holds the context permanently so no one has to remember to paste it. That's exactly what we built FirstOrg around: a structured onboarding interview captures your positioning, voice, and opinions, and Deep Lattice keeps that memory underneath every piece the engine produces, updating as your company changes — so draft forty knows you as well as draft one. If you're weighing that against raw ChatGPT, we've written an honest side-by-side comparison.
Third, keep the human pass. Context gets a draft to 90% you; the last 10% — is this claim actually true, would I really say this, is this the hill I want to die on — is judgment, and judgment stays human. That editing pass is also most of what search engines and readers are really responding to when they reward "human" content; as we've covered elsewhere, Google doesn't penalize AI content — it penalizes unhelpful content, which is what contextless drafting produces at scale.
The median is free. Your opinions aren't.
Here's the strategic picture. Everyone in your category has the same blank box you do. The median take on every topic in your industry now costs $20 a month and thirty seconds — which means the median is about to be the most crowded, least valuable real estate in content. Publishing it doesn't hurt you, exactly. It just spends your hours making you invisible.
What's scarce is what was always scarce: a specific person, with specific experience, willing to say what they actually think. Your positioning, your losses, your unpopular advice — no model has them, no competitor can copy them, and no reader confuses them with filler. The test I'd hold every draft to, whoever or whatever wrote it: would you sign it? Not "is it accurate" — would you put your name under it on a sales call? The only content worth publishing is the kind only you could have signed.
That standard is reachable without a writer on staff — the system for it is the same one we laid out in running content marketing without a marketing team: you own the point of view, and the production runs from it. The blank box was never the enemy. Blankness was.