Tenex Staff Change Their Writing to Avoid AI Suspicion

Tenex’s October 21 employee survey gives marketing managers a reason to be specific when criticizing AI-assisted writing. Of about 50 staff who answered, half said someone had suspected AI use in work they said they’d written themselves.

Nearly two-thirds reported changing how they naturally write to avoid looking machine-made. For a team producing customer emails or campaign copy, that introduces another demand on the writer: anticipating what a reviewer thinks AI sounds like.

The report has an institutional Tenex byline and makes its own limit explicit:

Our survey can’t tell us how often that feeling is right.

A manager’s suspicion leaves the employee with an accusation to answer. A concrete objection gives them an edit they can make.

A writing problem with an uncertain author

The sample comes from one AI-services company. These are employees’ recollections, without an independent check of who wrote each disputed text. The report doesn’t publish survey collection dates or response-level data. It can’t establish how common false suspicion is among marketers, much less whether a particular colleague used a model.

The employee response is still worth a manager’s attention. A direction to make a paragraph sound less like AI leaves considerable room for interpretation. The writer can change the rhythm or remove a phrase while preserving a weak argument. The reviewer can get a more familiar surface without receiving a more useful draft.

Consider a hypothetical customer email that promises a product will save time. A reviewer might dislike its polished tone. The substantive question is which customer, doing which task, has evidence for that claim. Rewriting the opening in a more casual voice doesn’t supply the missing evidence.

We’ve covered Tenex’s move to sell reusable AI agents. This finding concerns the people working around those tools. Its relevance is to the standards colleagues apply to one another, even when a company encourages AI adoption.

Why familiar prose raises suspicion

There is research behind the impression that models produce similar work. In Artificial Hivemind, researchers used a dataset of 26,000 open-ended prompts and reported repeated answers within models and similar answers across different models. The paper was first submitted in October 2025 and revised this month.

That gives a reason for readers to encounter familiar phrasing across generated work. It doesn’t establish the authorship of an individual document. The experiment studies model output; Tenex’s employees are describing judgments made about people.

A separate style-imitation study, submitted in September 2025, evaluated more than 40,000 generations per model using samples from over 400 authors. Its abstract reports that models approximate personal styles more closely in structured formats such as news and email than in informal blogs and forums.

For a marketing editor, that complicates a simple human-versus-machine reading. A tool can approach someone’s familiar style in one format while struggling in another. Neither study validates a colleague’s instinct that a sentence proves AI involvement. Both are historical research context, separate from the new employee survey.

Editing for appearances has limits

A fair objection is that cutting repetitive phrasing can improve the work. A campaign draft that makes the same point several times deserves a tighter edit. An unsupported claim deserves a source, whatever produced its first version.

The distinction matters when feedback becomes personal. An employee can explain why a claim belongs in the piece. They’re in a much harder position when asked to disprove a feeling about its tone. A manager who keeps returning to authorship without identifying the defect leaves the standard unclear.

Clay’s writing policy offers a related way to frame the review around the reader’s time. Marketing teams can ask whether a document makes a decision easier before debating the process that produced its sentences.

Tenex’s accompanying field guide asks a contributor:

Can I show how I improved the initial outputs?

That question has a useful application when someone has used AI. The writer should be able to explain a meaningful change, such as correcting a claim or choosing a different customer example. Merely making an output less recognizable doesn’t establish its value.

Tenex also sells help with AI adoption, giving it a commercial interest in the discussion and its proposed guide. The internal survey doesn’t measure whether the guide improves writing.

For managers navigating AI’s effect on marketing jobs, the practical obligation is to make a judgment employees can respond to. For the next customer email, mark the unsupported promise and ask for evidence of that specific benefit. The writer should leave the review knowing why the claim isn’t ready.