The tells, and how to strip them

There is a default register, and you can hear it

Model outputs converge on a recognisable style: measured, slightly formal, fond of lists, fond of tricolons, heavy on transitional phrases, ending with a summary that restates the beginning. It is not bad writing. It is average writing, which is what a system trained to produce the most acceptable continuation will produce.

You cannot remove it by asking for “human-sounding” prose. That is an adjective and adjectives do not carry style. You remove it by naming the mechanics.

The list that actually works

Keep this and paste it when the writing matters:

- No opening line that restates my question.
- No closing paragraph that summarises what you just said.
- Do not use: delve, tapestry, landscape, realm, testament,
  seamless, robust, leverage, unlock, navigate (figurative),
  crucial, pivotal, "it is important to note", "in today's",
  "not only... but also", "that said".
- No em dashes. No rhetorical questions.
- Vary sentence length. At least one sentence under six words
  per paragraph.
- No lists unless I asked for a list.
- Do not begin consecutive paragraphs with the same word.
- Say "is" rather than "serves as", "shows" rather than
  "highlights", "uses" rather than "utilises".

Every line is checkable. That is why it works, and why “sound more natural” does not.

The word list will need updating. The specific vocabulary that reads as machine-written shifts as models change and as human writers absorb it, which is itself worth noticing, because the borrowing goes both ways now.

The structural tells, which matter more

Vocabulary is the surface. Three deeper habits give it away.

Every point gets equal weight. Real writing is lopsided: one idea gets three paragraphs, another gets a clause. Machine prose distributes evenly, because nothing in it decided what matters most. The fix is to say so: “one of these points is more important than the others. Give it most of the space and say why.”

Nothing is at stake. A model does not have a position, so it produces balanced coverage of a question where a person would have committed. If you want an argument, ask for one: “take a position and defend it. Do not present both sides.”

No specific detail that was not supplied. Human writing is full of things that could only have been observed. Machine writing generalises, because inventing a specific is a fabrication risk. This is the tell that cannot be prompted away, and it points at the fix, which is that the specifics have to come from you.

The position on “AI detectors”

Tools claiming to detect machine-written text are unreliable in both directions. They flag human writing (with documented and repeated bias against people writing in a second language) and they miss machine writing that has been edited. Universities and publishers that treat their output as evidence are making a mistake, and if you are ever accused on that basis, the detector’s own published error rate is the thing to point at.

What follows for you: do not use one to check your own work, and do not trust one used on you.

Where this stops being a style question

If a document goes out under your name, the fix is not a better prompt. It is that you write the parts that require a position or a specific, and use the model for the parts that do not.

Anything that carries obligation (a reference, a condolence, an apology, a promise, an assessment of a person) should be your words. This is not a rule about detection. It is that those documents are only worth anything because a person wrote them, and a perfectly styled version of one is worth nothing at all.

The practical sequence

1. Give it the material and the reader.
2. Add the mechanics list.
3. Read the draft and delete the two sentences that say nothing.
4. Put back the one specific detail only you know.

Step four is where the draft stops sounding machine-written, and it is the only step no prompt can do for you.

Why the register exists at all

It helps to know that this style is not a quirk somebody could remove. Preference training rewards responses that a wide range of raters found acceptable, and the safest way to be acceptable to everyone is to be measured, balanced, structured and slightly formal. The register is the shape of an average of opinions.

That is also why it resists instruction. You are asking for output away from the centre of a distribution the training deliberately pulled towards. It can be done, with mechanics and specifics, and it takes ongoing effort rather than one clever line at the top of the prompt.