Guides
AI-assisted business writing

House voice versus model voice

When a model writes your first drafts, your documents can gradually take on its voice instead of yours. We look at why that happens and how to keep your organization sounding like itself.

Every organization has a voice, whether or not anyone has described it. It’s the reason a regular reader can often tell who wrote a document before seeing the name on it. You can hear it in the length of the sentences, the words the organization uses and the ones it avoids, how directly it makes a claim, and how it speaks to the reader. In many organizations, that voice has never been written down. It lives in the habits of the few people who do most of the writing. When they leave, the voice can leave with them.

Once a model is writing the first drafts, the voice can start to drift. When each draft arrives in the model’s words, it takes a human to deliberately replace those words. If nobody does, much of that phrasing will survive in the final version. The voice shifts a little with each document, and over time, the organization’s reports and web pages can start to sound like those of other organizations using similar tools.

Several studies of writing with AI have found similar effects. In one, short stories written with ideas from a language model were rated as better written than stories written without them, but they were also more similar to one another (Doshi & Hauser, 2024). Another, a cross-cultural study, found that AI writing suggestions led Indian participants to adopt Western writing styles. The suggestions affected the content of what they wrote as well as its style (Agarwal et al., 2025). The largest, which analyzed more than 880,000 texts, found that when language models polished or rewrote people’s writing, the content survived but the range of writing styles narrowed (Sourati et al., 2026).

What model voice sounds like

Because the drift is hard to see from inside an organization, it helps to know what you’re listening for. Model voice tends to keep the same even tone from the first sentence to the last. Hedging is habitual, and firm commitments are rare. Paired phrases (“clear and concise”) and lists of three turn up often. Sections tend to open by announcing what’s coming and close by summarizing what’s been said. The vocabulary stays much the same whatever the subject. Our guide to the patterns that make readers stop trusting a text describes these habits in more detail, along with ways to address each.

A house voice is more specific because it reflects the organization’s own judgment about where to be blunt and where to be careful. A law firm and a design studio will draw those lines in very different places.

Why the voice drifts

The drift happens through ordinary editing. The draft produced by a model is often good enough that no one feels the need to rewrite it from scratch. Whoever edits it corrects the facts and the structure but leaves the phrasing alone because it isn’t wrong. When the same thing happens with every document, the model’s phrasing gradually becomes the organization’s. New staff then learn the voice from the most recent documents. By then, those documents already sound like the model. Because the change is gradual, it’s hard to say when it happened.

None of this is anyone’s fault, but a policy that tells staff to “keep our voice” won’t stop it on its own. Without a written description, the people applying the policy have nothing to check their drafts against.

Write the voice down

A practical defense is to write down how your organization sounds, in enough detail that someone can edit a draft against it. A phrase like “professional but approachable” won’t do because it could describe almost any organization. The description needs to record the actual decisions: how long the sentences run, whether the organization says “we” or uses its name, how it handles uncertainty, which words it never uses, whether it allows contractions or acronyms, and how it opens and concludes a document. Keep it to a page or two, with a few passages of your own best writing as examples.

Once it exists, the description tells anyone editing a model’s draft what “make it sound like us” means. New writers can learn the voice from it, too, instead of picking it up from recent documents that may already sound like the model. The voice description belongs in your style guide. If you don’t have one yet, the description is a good place to start. We can also draft the guide with you if your team doesn’t have the time.

Edit the voice back in

Once your voice is written down, you have something specific to edit a model’s draft against. Read each paragraph and ask whether your organization would have put it that way. If the answer is no, rewrite it. Expect to cut a lot: the announcements, the hedges, the paired phrases, and the summaries. You’ll also swap some of the model’s vocabulary for your own. Occasionally, you’ll need to add a sentence where your organization would have made a blunt point that a model is unlikely to make.

Editing the voice back in is what turns a competent document into one that’s recognizably yours. With a deadline looming, it’s an easy step to skip, but every time you do, the model’s phrasing prevails.

Keeping it over time

With your organization’s voice written down, the character of your writing will survive staff turnover and a switch to new tools. It still needs a designated owner: someone who checks that everything your organization sends out follows it, and who updates it as the organization grows. If you send us work throughout the year, a group account puts the same small team of editors on every document. Once they know your voice, they can make sure a report you publish in October still sounds like the one you published in March.

References

  • Agarwal, D., Naaman, M., & Vashistha, A. (2025). AI suggestions homogenize writing toward Western styles and diminish cultural nuances. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (pp. 1–21). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713564
  • Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), Article eadn5290. https://doi.org/10.1126/sciadv.adn5290
  • Sourati, Z., Karimi-Malekabadi, F., Ozcan, M., McDaniel, C., Ziabari, A., Trager, J., Tak, A. N., Chen, M., Morstatter, F., & Dehghani, M. (2026). The shrinking landscape of linguistic diversity in the age of large language models. Nature Human Behaviour, 10(10), 1925–1942. https://doi.org/10.1038/s41562-026-02550-0

Losing your voice?

Tell us how your organization sounds, or should, and we’ll agree the scope with you before quoting.

Get a quote