AI writing sounds like AI because of what you left out, not because of how the model writes. The tell people notice is not a style the model prefers, it is the sound of a machine filling gaps you did not fill for it. If you have not said who is reading, what they already believe, and what you are not willing to claim, it will produce the average of everything ever written on your topic. The average is what “sounds like AI” means.
That answer is easy to assert. Below is the part we can actually show, and the part we cannot.
We went and counted
We run a tool that turns a rough goal into a structured brief, so we hold a record of how people describe what they want before any AI touches it. In August 2026 we read all of it: 1,678 briefs from 1,102 different people, after stripping out our own accounts, which were 72% of all usage and would otherwise have described us rather than anybody else.
Then we searched every one of those briefs for the request this article is about. Writing that sounds human. Writing in my voice. Not robotic. Not obviously AI.
Seven. Seven briefs out of 1,678. Two of those were the same brand brief matching on the word “voice”, and two more were one prompt someone pasted in twice. Genuinely on topic: three, maybe four. Call it a quarter of one percent.
That number is worth sitting with, because search behaviour points the opposite way. People look this problem up constantly. They just do not raise it at the moment they are actually briefing an AI. Whatever “sounds like AI” is, it is a complaint people make afterwards about work already done, and almost never an instruction they give beforehand.
What people send instead
The same 1,678 briefs, measured on length:
| Measure | Value |
|---|---|
| Median brief | 23 words |
| Briefs under 10 words | 390 |
| Briefs under 5 words | 149 |
One hundred and forty nine people opened a tool built for describing what they want, and described it in under five words.
That is not a criticism of them. It is the actual shape of the problem. Twenty three words cannot carry a reader, a purpose, a constraint and a voice all at once. There is not room. So the model supplies all four from the average of its training data, which is precisely the thing that reads as synthetic.
When we tagged what people were visibly struggling with, the top two were these:
- “No idea what to ask for.” Named in 65 briefs, the single most common obstacle in the dataset.
- “Has the facts but not the wording.” Named in 33.
Those are different people with the same output problem. The first has nothing to put in. The second has everything to put in and no idea it was supposed to go in.
What this data cannot tell you, and why we are saying so
Here is where an article like this would normally produce a table of magic instructions with percentages attached. We are not going to, because we cannot, and the reason is worth stating plainly.
Our pipeline never scores output quality. We record what people typed and whether they finished. We do not measure, anywhere, whether the writing that came out the other end was any good. So there is no honest version of the sentence “our data shows that adding instruction X makes writing sound 27% more human.” Nobody has that number. Anyone printing it either measured something far narrower than the claim, or made it up.
There is a related figure we could have misused and did not. Briefs that arrive specific finish our questionnaire 94.6% of the time. Briefs that arrive vague finish 1.4% of the time. Both are real and exact. Neither means what it would be convenient for them to mean: they measure who completes a form, not whose writing improved. Presenting that as evidence about writing quality would be the same sleight of hand this article is complaining about.
So what follows is labelled. One first-party finding we can stand behind, which is that almost nobody asks for this and almost nobody supplies enough for it. Then a mechanism, which is reasoning rather than measurement.
The mechanism, labelled as reasoning
A language model asked for a blog post with no further detail does not fail. It succeeds at a different task than the one you meant. It produces the most probable blog post, and the most probable version of anything belongs to nobody in particular. That is the entire phenomenon. “Sounds like AI” is what the middle of a distribution sounds like when you were expecting one specific person.
Which means the fix is not a tone instruction laid on top. Adding “write in a natural, human, conversational voice” narrows the distribution a little and moves its centre somewhere equally generic. It is still an average. It is now an average of writing that was told to sound casual, which is its own recognisable flavour.
What moves it is information only you hold. Not “professional but approachable”, which every writer alive reads differently. Instead: the sentence your company would never print. The objection this particular reader arrives with. The claim you are deliberately not making. Constraints do more work here than descriptions, because a constraint removes probable output, and probable output is the problem.
What I got wrong building this
I built Briefing Fox on the assumption that people cannot say what they want from an AI. The data backs that half of it up hard: 65 briefs naming “no idea what to ask for” as the obstacle, a median goal length of 23 words, 149 people who opened a tool built specifically for describing intent and used fewer than five words to do it. That part of my premise was right, maybe understated.
What I got wrong is what I assumed the shape of that gap was. I expected a good chunk of the confusion to be about voice. Sounding robotic, sounding like a template, wanting to sound “like them.” Instead the gap is almost entirely about content, not tone: people show up without the facts, the structure, or the angle, and the 33 who had “the facts but not the wording” still weren’t asking for a voice, they were asking for help organizing what they already had. Seven out of 1,678 went anywhere near “sounds human.” I’d have guessed the number was in the dozens at least.
In hindsight the reason is obvious once you say it: nobody drafting a request thinks about how the output will sound. They think about whether it will be right. Voice is the complaint you make after you’ve read a bad result, not a spec you write beforehand, and a product built to capture intent up front was never going to see it, because it isn’t there yet at that point in the process. The tell people call “sounds like AI” is downstream of the same gap that produces a 23-word goal, it’s just noticed later and blamed on the wrong stage.
What to do instead
Before the next thing you ask an AI to write, put three things in writing that were not going to be there otherwise. Who is reading it and what they already think. One sentence you would never publish. What the piece has to achieve that a competent generic version would miss.
It takes about ninety seconds, and it is the whole difference. Not because AI needs a ritual, but because those three things are the ones you were always going to leave out, and they are exactly the ones the average cannot supply.