July 23, 2026

The assistant voice, and how it gives you away

There are two different problems in AI-written text, and most tools only address one.

The first is statistical: flat rhythm, predictable word choices, symmetry everywhere. The second is behavioural. Chatbots were trained to be helpful, and helpfulness has a style. It shows up in text that is otherwise fine, and it is the reason a support reply can pass every rhythm check and still read, unmistakably, as machine-written.

We found these by rewriting real support tickets and reading what survived. Six habits, in the order they annoy us.

1. The narrated method

"Rather than answer from memory, I ran the test on a clean install." "Instead of guessing, I checked the actual output." The construction is always the same: dismiss a lazy approach you never took, then announce the diligent one.

No human writes this. A person who ran the test just says what happened. "I ran the test on a clean install. Export took four seconds." The information is identical. The self-congratulation is gone.

2. The teaser heading

"What the logs actually revealed." "One file answers all of it." "The part nobody tells you." Headings written to make you want the next paragraph, in a document nobody is scrolling past.

A heading is a label. "Export timings" tells the reader where they are. The teaser makes a promise the section then has to keep, and it usually cannot.

3. The announced structure

"Let me walk you through this." "There are three things going on here." "First, some context." The text describing what the text is about to do.

Just do the thing. If there are three points, the reader will notice three points.

4. The framing preamble

"Great question." "That is a fair concern." "Honestly, this one surprised me." Warmth applied at the front, like primer.

It is not that a support reply should be cold. It is that the warmth is generic, and generic warmth reads as no warmth at all. Real warmth is specific: it refers to something the person actually said.

5. The closing encouragement

"Hope this helps." "Let me know if anything is unclear." "You have got this." The reflex to end on a supportive note, whatever the content was.

Sometimes a closing line is right, when there is genuinely a next step. Most of the time the last useful sentence was two sentences ago.

6. The escalating restatement

"This is not just a setting, it is the whole mechanism." "That is exactly the data I needed." The device that takes something ordinary and insists on its importance.

It is a tic with a tell: the sentence adds no information, only emphasis. Remove it and nothing about the situation changes.

The rule that makes this fixable

For a long time our rewriting rules said "delete the assistant voice", and that conflicted with the rules telling the model to preserve everything. Facts got dropped along with the framing.

The rule that works is narrower. Keep the information, cut the framing. "Rather than answer from memory, I ran the test and export took four seconds" contains one fact and one performance. Keep the fact. Drop the performance.

Youmanize applies this in every tone, including the ones where other rules relax. It also refuses to introduce these patterns, which turns out to matter: an early version of our pipeline removed one narrated method and cheerfully invented another two paragraphs later.

If you write support replies, sales emails or documentation with a chatbot, this is the layer your readers are reacting to. Not the vocabulary. The eagerness.

Back to the blog