Why AI-Generated Writing Sounds Generic (And What to Do About It)

You’ve probably done this. You reach for ChatGPT or Claude or Gemini to draft something personal — a board update, a customer reply, a message to a friend you haven’t talked to in a while — and what comes back sounds fine. Technically correct. Grammatically clean. And also somehow… not you.

There’s a specific feeling that goes with reading AI-generated writing where the voice is supposed to be yours. A slight uncanny. A sense that the sentences would work in anyone’s mouth. That the words could have come from any competent person, which is a strange thing to say about a message that’s supposed to sound like you specifically.

Sam Kececi, founder and CEO of The Sentience Company, has a reframe of this problem that I haven’t been able to stop thinking about.

The problem isn’t that AI writes. It’s that everyone’s AI writes from the same base.

Here’s the common frame: AI-generated writing sounds generic because AI is not creative, or because AI can’t capture human nuance, or because AI is trained on the average of the internet.

Sam pushes back on all three. His argument is simpler and, I think, more useful.

The issue isn’t that AI writes for you. It’s that everyone’s AI writes from the same three or four base models. When 90 percent of the drafts flowing through the world get filtered through ChatGPT, Claude, or Gemini, the outputs converge on a median voice. Not because any one of those models is bad — they’re extraordinary — but because they’re statistically optimized for likely responses. Which means the median. Which means everyone starts to sound the same.

“Imagine if 90 percent of the tweets you saw were just written by the same guy,” Sam said. “You would still get annoyed. If I went on Twitter and my whole Twitter was just one guy, I would quit Twitter.”

That’s the actual dynamic. The complaint isn’t that the writing is machine-produced. It’s that the writing is undifferentiated. Every third product launch on LinkedIn now uses the same three-word opener, the same em-dash cadence, the same “it’s not [X] — it’s [Y]” flip. Every founder update has the same “let’s talk about…” start. The AI didn’t invent those patterns. It amplified them.

What voice actually is, from the inside

Here’s what’s interesting about the diagnosis: it points at the solution.

If the problem is that the base model is the same for everyone, then the solution isn’t “write it yourself” or “learn to prompt better.” The solution is a base model that’s not the same for everyone.

Sam’s whole company is built on this thesis. Sentience creates what he calls a “personal simulation” — a per-user model that captures the specific way you write, the decisions you’ve made, the arguments you find compelling, the words you actually use. Not a persona prompt layered on top of GPT. An actual model shaped by your data through a combination of retrieval and fine-tuning, one that outputs sentences that sound like yours because it’s been trained on yours.

Whether or not you buy the specific technical approach (per-user LoRa fine-tuning + a proprietary knowledge synthesis system he calls the “sleep phase”), the diagnosis holds. And the practical implication for anyone writing today — regardless of what tool they use — is that generic prompts produce generic voice. Specificity is the input that survives.

What to do if you can’t build your own model

Most people won’t have a per-user AI model tomorrow. So here’s the practical layer.

Feed the model your actual writing. Not “here’s my brand tone.” Actual writing. Real emails you’ve sent, real posts you’ve published, real drafts you’re proud of. Attach them, paste them, keep a running “my voice” doc that you feed to the model at the start of every session. Generic voice-guidelines produce generic voice-imitation.

Edit the model’s draft, don’t accept it. Every generic sentence is a slot for a specific one. If the model writes “we’re excited to announce,” you write “here’s what we shipped.” If it writes “leverage cutting-edge technology,” you write the specific thing you actually did. Your voice is what survives the edit, not what shows up in the first draft.

Watch for the tells. AI-generated writing has patterns. Fake contrasts (“not X — Y”). Absurd certainty (“always,” “never,” “everyone”). Manifesto language (“this changes everything”). Triple-parallel structures (“no vendor. no black box. no negotiation.”). Once you see the patterns, you can’t unsee them. And once you can’t unsee them, you’ll cut them from your own drafts.

Own the voice that survives. The uncomfortable truth Sam’s diagnosis surfaces is that voice was always something you had to build. AI just made the cost of not building one immediately visible. If you don’t have a distinctive way of writing, the median voice will fill the vacuum. If you do, it survives — and gets more valuable, not less, because it’s now scarce.

The uncomfortable version of the argument

There’s a version of Sam’s thesis that goes further than most people want to sit with.

“No one can compete with you at being you,” he told me, borrowing a Naval Ravikant line. And he pointed out that a year ago, some people could argue they were faster at coding than AI — and now that argument is basically gone. Speed at coding is a race that’s ending. But being you — with your specific experiences, your specific hot takes, your specific way of seeing the world — is a race no one else can enter.

Which means the writing problem isn’t a writing problem. It’s an identity problem, forced into visibility by tools that flatten everything else.

You can’t outsource your voice to a system trained on the median. Not because the system is bad. Because your voice is by definition not the median. It’s whatever’s left when you strip away everything that could have come from anyone else.

Which, it turns out, is more valuable now than it was a year ago.

---

Frequently Asked Questions

Why does AI-generated writing sound generic? AI writing sounds generic because 90 percent of drafts get filtered through the same three or four base language models (ChatGPT, Claude, Gemini), which are statistically optimized for likely responses. The outputs converge on a median voice. The problem isn’t that AI writes — it’s that everyone’s AI writes from the same statistical base. Sam Kececi of The Sentience Company argues the fix is personal models trained on the individual, not shared models trained on everyone.

Is AI writing killing personal voice online? The visible effect is that online writing is becoming more uniform. Product launches, LinkedIn posts, and founder updates increasingly share the same rhythmic patterns and vocabulary because they share the same underlying model. The counter-move is either building a personal AI model that captures your specific voice, or heavily editing generic drafts against your own writing style before publishing.

How can I make ChatGPT sound more like me? Feed it your actual writing at the start of every session, not brand-tone guidelines. Paste in real emails, posts, or drafts you’re proud of and ask the model to match the specific voice. Then edit every generic sentence into a specific one. Watch for AI writing tells: fake contrasts, absurd certainty, manifesto language, triple-parallel structures. Cut them ruthlessly.

What’s the difference between ChatGPT memory and a personal AI model? ChatGPT memory retains facts from your past conversations with ChatGPT, but the underlying model is still shared with everyone. A personal AI model is trained on your writing, decisions, and context — the model itself is unique to you. Sentience uses per-user LoRa fine-tuning on top of open-source base models, plus a nightly synthesis process the founder calls the sleep phase.

Will personal AI models replace ChatGPT? Not directly. ChatGPT and Claude remain useful for tasks that don’t require personalization — coding, data analysis, general information retrieval. Personal AI models complement them for tasks that require your specific voice, judgment, or context: personal writing, decision-making, interview prep, running your business. Sam Kececi frames it as: use ChatGPT for code, use Sentience for taste.

Why do all AI-generated LinkedIn posts sound the same? They share the same base model, the same prompt patterns, and the same “brand voice” instructions. When thousands of founders prompt ChatGPT with “write a LinkedIn post about my product launch,” the model returns statistically likely outputs, which converge on a shared style. Distinctive voice requires specific inputs — actual past writing, personal experiences, specific arguments — not generic guidelines.

What are the “tells” of AI-generated writing? Common tells include: fake contrasts (“this isn’t X, it’s Y”), absurd certainty (“always,” “never,” “everyone”), manifesto language (“this changes everything”), triple-parallel structures (“no vendor. no black box. no negotiation.”), too-perfect rhythm (hook → claim → three bullets → dramatic close), and smug conclusions (“and that’s the real lesson here”). Once you see them, you can’t unsee them.

Is human writing better than AI writing? Human writing isn’t automatically better. AI writing that captures a specific voice can be excellent. The real question is whether the writing feels differentiated — whether it could have come from any competent writer, or only from this specific person. Distinctive voice is the metric, not human vs AI origin.

Should founders use AI to write their newsletters or LinkedIn posts? Depends on how you use it. Using AI to draft and then heavily editing to your voice is defensible and common. Using AI to draft and publishing without editing tends to produce content that sounds like everyone else’s. The differentiator is the editing pass — that’s where your voice enters. Some founders build personal AI models specifically to skip this friction.

How does The Sentience Company solve the voice-flattening problem? Sentience builds a per-user personal AI model trained on your emails, calendar, Slack, notes, and meetings, updated nightly through a knowledge synthesis process. The output voice matches yours because the model is shaped by your data, not by the median of the internet. Users own their model as a legal asset — it’s structured so it can be transferred or inherited. Watch Sam Kececi explain the architecture on the TwoSetAI podcast.