The ChatGPT voice problem

Transmission 03 — Anchor In Orbit

Read three blog posts back to back. Don't look at the byline. Just read.

You can't tell whose are whose.

The rhythm is the same in all three. There is a short opener. A long second sentence with a parenthetical. A two-word sentence somewhere in the middle. A tricolon. An italic line at the bottom doing the work of a conclusion. The vocabulary is the same: compounding, leverage, moat, asymmetric. Words that meant something specific in the mouths of specific people, now used by everyone because the same machine handed them out.

This is the ChatGPT voice problem, and it isn't really a ChatGPT problem. It's what happens when everyone uses the same tool to draft. It happened to MySpace layouts in 2006. It happened to Squarespace landing pages in 2014. It happened to design portfolios on Webflow in 2020. Tools level the floor. They also level the ceiling.

The floor is fine. AI has made writing accessible to founders who were staring at a blank document on a Tuesday night with nothing to ship. Something beats nothing.

The ceiling is the problem, and the ceiling is where brand lives.

What every founder who uses the same machine ends up with is the same writing. Their content stops being attributable. The customer's job becomes harder, not easier. Every brand sounds like every other brand, so the customer has no way to remember any of them.

Voice is the antidote. Voice is not what an AI assistant gives you for free.

A brand voice is the pattern of words you choose and the words you refuse to print. The machine will hand you the choosing on a plate. It cannot do the refusing. Refusing requires a fingerprint: a written list of vocabulary you've banned, sentence shapes you've decided aren't yours, rhetorical moves you've already used too many times. Without the fingerprint, the model defaults to the mean of what it's read, which is the mean of what everyone else has already written. Mean reads as competent. Mean also reads as nobody.

Four levers run a brand voice: diction, rhythm, stance, restraint. The machine does well on diction and rhythm. It does poorly on stance. It is worse at restraint.

Stance is whether you take a side. The default LLM voice is studiously neutral. It hedges. It surveys. It "considers." A brand that hedges is invisible, because the customer has nothing to agree with and nothing to push against. The first move past the default is deciding what you are against. You can write a post entirely about what you reject and still be more memorable than the brand that wrote about what it believes.

Restraint is what you refuse. Five words you will not print. Three sentence shapes that aren't yours. The closing move you've already used four times and will not use again. Restraint is invisible to outsiders, and it is the reason brands you can identify by their writing sound the way they do. Patagonia refuses certain words. Liquid Death refuses certain tones. Mercury refuses certain rhythms. They sound like themselves because they decided what they would not be.

There is a test. Print three of your last posts. Print three of a competitor's. Mix the stack up and hand it to a friend who reads in your category. Ask them to sort. If they cannot, the machine wrote both stacks, and your brand is hiding behind tooling that does not belong to you.

You don't have to stop using the assistant. The assistant is a draft. Every draft needs a pass that is not "make this sound more natural." It needs the pass where the banned word goes. Where the polite hedge gets replaced with a position. Where the four-time-used closing move gets cut.

That pass is where the brand lives. Not in what the machine wrote. In what you refused to publish.

The brands that will be recognized in 2027 are the ones writing down five banned words today. Everyone else will be sounding like the machine.