Why AI-written content doesn't sound like you

1. Blog
2. Ops, Teams & AI

By Saroj Jha · September 21, 2026 · 5 min read

A language model favors likely text for the situation it is given. Without explicit rules, likely text means the average of your category, so AI drafts converge on the phrasing your competitors already use. Adjectives in a prompt do not fix that. Rules a sentence can fail do.

You gave the model your product, your audience, and a request for something “in our voice”. What came back was fluent, accurate enough, and could have been published by any of your competitors. Edit it and it gets better. Ask again and you get the same thing in a different order.

This isn’t a prompting problem, and more adjectives won’t fix it. It’s what the tool is built to do.

What the model is actually doing

A language model is trained to predict the next word, or part of a word, in a passage, and when it writes it favors the likely ones. Always taking the likeliest wording makes text “bland and strangely repetitive”, researchers found^[4]. Ask it for a paragraph about your product with nothing else to go on, and “likely” means the kind of paragraph it has seen most often in that setting: the way most companies in your category already write.

That’s why the drafts converge. The opening line about a fast-moving landscape. The platform that empowers teams. Three benefits in parallel structure. None of it is wrong. All of it is the middle of the distribution, and your competitors’ drafts are drawn from the same middle.

Your voice is, by definition, the part that isn’t average. So a tool whose default is the average will sand it off unless something stops it.

Controlled studies point the same way, though none tested marketing copy. In a 2024 experiment, 293 people wrote short stories; those given GPT-4’s ideas wrote stories rated more creative, and more alike^[1]. Essays written with InstructGPT, but not the older base GPT-3, repeated more of the same key points^[2]. And people brainstorming with ChatGPT produced ideas more alike as a group than with a deck of prompt cards^[3].

The same pull works on facts

I found out how far this goes on my own site.

Most of aajconsult.com was drafted with AI assistance. In September I audited every statistic on it, and the problems had a pattern. A sample of “939 B2B companies” appeared on two pages of the site, attached to two unrelated studies: one about retention by segment, one about pipeline coverage. A finding from Gartner’s buying-journey research, several years old, was dated 2024. Two performance figures were credited to Forrester, which had never published them; the only trace was a vendor’s blog.

Every one of those looked right. The sample was a believable number. The sources were the kind of firms that publish that kind of research. The dates were close. They were plausible, and plausibility is exactly what the tool produces: OpenAI researchers argued in 2025 that training and scoring reward a confident guess over admitting uncertainty^[5]. What a model doesn’t do by default is look anything up to check that the thing it’s describing exists.

I can’t tell you which of those errors a model introduced and which it copied from a page where a model had introduced them. That matters less than it sounds. Nobody checked, and every one of them read as true.

Voice drifts the same way. The generic sentence and the invented statistic are the same failure: the most likely output, with nothing in place to say otherwise. Every figure on this site now carries a named source or is labeled as a worked example or my own estimate; the benchmark library shows what that looks like.

Why “sound more like us” doesn’t work

The usual fix is to add adjectives to the prompt. Be conversational. Be bold. Sound human.

None of those can fail. There’s no sentence you can point to and say it broke “be conversational”, so the model can’t be held to it, and neither can the three people on your team who also write. An adjective describes a voice. It doesn’t define one.

A rule is something a sentence can fail. “Use contractions” can fail. “Never open with a question” can fail. “Don’t write leverage” can fail. That’s the whole difference, and it’s why so many brand voice documents I see get admired once and never applied.

What actually holds

In my practice, four things hold, in the order they matter to a tool.

Words you never use. A literal list: category filler that turns up in every competitor’s copy, claims you can’t defend, and the words that are wrong for you specifically. This is the one a check can enforce mechanically.

Phrases you do use. The part people skip. In my view, give a model a banned list and nothing else and it routes around the ban into a different flavor of generic. The positive list is what turns avoidance into a voice.

A reading-level ceiling. One number, such as a grade level. Leave it out and an automated check either never fires or fires on everything.

Before-and-after rewrites. Three real sentences from your own copy, wrong, then fixed. Abstract rules start arguments; a rewrite ends them.

Then give the tools the same rules you give the people. Put them in the context file your AI tools read, not in a PDF on a shared drive, so a check can quote back every violation with the paragraph it’s in rather than an editor catching them by eye late at night.

I’ve turned this into a free Brand Voice Guide template. It ends in a block you paste straight into an agent’s context file. If your team is adopting AI tools faster than anyone has written rules for them, the AI Tool Policy covers the rest, and I’ve written separately about practical ways to use AI without losing your voice.

The honest limit

Rule-checking is literal. It catches “leverage” and misses “harness the power of”. It scales the mechanical part of review so a person can spend their attention on the part that needs judgment. It doesn’t replace that person on anything that matters.

It has a blind spot, too, which I found by running a literal word check against my own voice guide. It flagged nine banned words, and every hit sat inside a list of words to avoid or a deliberately bad example. Anything written about banned words fails on the words it discusses. Read the findings, not the score.

What I’d do on Monday

Do this

Take the last five pieces your team published with AI help. Highlight every phrase you could imagine in a competitor’s copy. That list is the first draft of your words to avoid, and it’s usually longer than people expect.

Free tool

Check a draft against your own rules

The free Brand Voice Checker tests copy against your words to avoid, the phrases you use and a reading-level ceiling, and quotes each finding in its sentence.

If you want me to look at your own funnel, positioning and channels, my Growth Audit ends with three prioritized fixes and a 30-day action plan.