Why Does AI Writing Sound Like AI? (Tells and Fixes)

TL;DR: AI writing sounds like AI because models default to a house style: reflexive praise, preamble, bullet fragmentation, hedged claims and symmetrical phrasing. These are habits, not errors, and the model vendors document most of them. Fix them with a style contract and a voice sample. Do not fix them by chasing detector scores.

Why does AI writing sound like AI?

Because the model is doing what it was trained to do, and that has a texture. Left unsteered, a language model produces the average of a very large amount of competent, careful, slightly anxious prose. Symmetrical. Hedged. Helpfully structured. Nothing in it is wrong, which is precisely why it is so easy to spot.

The useful reframe is that these tells are defaults, not defects. A default is something you can override with an instruction. That is the whole of this article.

It is also worth knowing that the labs are not pretending otherwise. OpenAI's ChatGPT release notes are full of entries about exactly the habits people complain about, and Anthropic publishes the system prompts it uses to suppress them. Those documents are the best evidence available, and I have quoted them below with dates so you can check whether they still say what they said in August 2026.

What are the tells, exactly?

Here is the catalogue of AI writing tells worth knowing. I have separated the ones a vendor has written down from the ones that are simply widely observed, because the difference matters when you are arguing with someone about whether a piece was machine-written.

The last four have no vendor citation I could find. I am not going to invent one. They are pattern observations, and you can test them yourself in about ninety seconds by asking any model a question and counting.

Two of the documented ones are worth reading in full, because the wording tells you what to say in your own prompt.

On teaser phrasing, OpenAI's ChatGPT release notes for March 16 2026 describe an update to GPT-5.3 Instant that "improves follow-up tone and reduces teaser-style phrasing in responses (e.g., 'If you want…', 'You'll never believe…', 'I can tell you these three things that…')."

On over-structuring, the entry for May 28 2026 says the GPT-5.5 Instant update produced responses that are "better paced in practical help tasks, with fewer overly long or bullet-heavy responses." The July 6 2026 entry says GPT-5.5 Instant Mini "avoids repetitive or overly structured responses."

When a vendor ships a fix for a behaviour, that behaviour was real.

Why does AI turn everything into bullet points?

Because bullets score well as "helpful", and until recently nothing pushed back. Anthropic's published system prompt for Claude Opus 4.6, dated February 5 2026, contains a block called lists_and_bullets that reads in part:

Claude avoids over-formatting responses with elements like bold emphasis, headers, lists, and bullet points. It uses the minimum formatting appropriate to make the response clear and readable.

And more bluntly, further down the same block:

Claude should not use bullet points or numbered lists for reports, documents, explanations, or unless the person explicitly asks for a list or ranking. […] Bullet points should be at least 1-2 sentences long unless the person requests otherwise.

That last rule is the one to steal. "No bullet shorter than one full sentence" is a specification. "Don't use so many bullets" is a mood.

Anthropic's prompt engineering guidance adds a mechanism that most people never try: the formatting of your prompt leaks into the formatting of the answer. In its own words, "removing markdown from your prompt can reduce the volume of markdown in the output." If you write your brief as three bulleted fragments, you have already asked for bulleted fragments.

Why does AI praise everything I write?

Because warmth and agreement are hard to separate during training, and OpenAI says so directly. Its Model Spec contains a rule headed Don't be sycophantic, and the reasoning underneath it is short:

A related concern involves sycophancy, which erodes trust. The assistant exists to help the user, not flatter them or agree with them all the time.

The Model Spec goes on to say that when a user asks for a critique, the assistant should behave "more like a firm sounding board that users can bounce ideas off of" rather than "a sponge that doles out praise." In the ChatGPT release notes for August 15 2025, OpenAI made GPT-5's default personality warmer and added the caveat that "teaching models to be warm without being sycophantic is an ongoing research challenge."

So the flattery is a known, unresolved tension, not a compliment. The practical consequence is that you have to ask for the objection explicitly, because the model will not volunteer it. My default follow-up on anything I have written is a single line: give me the strongest argument that this is wrong, before you tell me anything you liked.

If every draft comes back sounding the same regardless of what you wrote, why your startup’s AI content sounds like everyone else’s covers the same failure at the level of a whole content programme.

Is the em dash really an AI tell?

Not by itself, and I want to be careful here because this is the claim that circulates furthest with the least evidence behind it.

I checked. OpenAI does not document any em-dash-specific behaviour in its ChatGPT release notes or in its GPT-5.1 announcement page. There is no vendor statement saying the models were trained to use em dashes, or trained to stop. What OpenAI does document, in the GPT-5.1 announcement of November 12 2025, is that "the updated GPT-5.1 models are also better at adhering to custom instructions, giving you even more precise control over tone and behavior," and that changes made in personalization settings now take effect across all chats immediately rather than only in new ones.

That is the real story. The punctuation was never the problem. The problem was that telling the model to stop did not reliably work, and instruction adherence is what changed.

The tell is density and uniformity, not the character. A writer who uses three em dashes in a 2,000-word piece is using a punctuation mark. A draft with one in almost every paragraph, always performing the same interruption, is running a template. Count them before you accuse anyone, including yourself.

How do I stop AI writing from sounding like AI?

Give the model a specification instead of an adjective. Anthropic's guidance is explicit about the direction: tell the model what to do rather than what not to do. Its example is exactly on point. Instead of "Do not use markdown in your response," try "Your response should be composed of smoothly flowing prose paragraphs."

Three layers do almost all the work. Start with a style contract, add a voice sample, and finish with a stripping pass.

1. The style contract

Paste this once at the top of a thread, or save it as a reusable system prompt.

STYLE CONTRACT: applies to everything you write in this thread.

VOICE
Write as a working practitioner explaining something to a peer.
First person where it is natural. Concrete nouns. Real examples.

RHYTHM
Vary sentence length deliberately. At least one sentence under
eight words in every paragraph. Never three sentences of the same
length in a row.

STRUCTURE
Prose paragraphs by default. Use a list only when the items are
genuinely discrete. No bullet shorter than one complete sentence.
No headers on anything under four paragraphs.

OPENINGS
Start with the answer. Do not restate my question. Do not open
with praise or with "Here is" / "Based on".

ENDINGS
Stop when the answer is finished. No summary paragraph. No offer
of further help.

BANNED
"Great question", "It's not just X, it's Y", "dive into",
"when it comes to", "it's worth noting", "I hope this helps",
"genuinely", "honestly", "straightforward".

PUNCTUATION
At most one em dash per 500 words. Prefer a comma, a colon, or a
full stop.

CLAIMS
One claim per sentence. If you are unsure of a fact, flag it in a
short parenthetical rather than hedging the whole sentence.

Two of those banned words come straight from Anthropic. The published Claude Opus 5 system prompt, dated July 24 2026, says: "Claude avoids saying 'genuinely', 'honestly', or 'straightforward'. Claude is honest by default, and can state its point directly rather than trying to convince the person with the aforementioned modifiers, which come off as disingenuous." When the lab that trained the model bans a word in its own instructions, that word is a tell.

2. The voice sample

This is the single highest-leverage block on the page, and almost nobody uses it. Adjectives get resolved against the model's average. A sample gets resolved against you.

Below is a sample of my own writing. Study three things: the
rhythm, the vocabulary range, and where I choose to stop sentences.

<voice_sample>
[paste 300–500 words you actually wrote: an email, a Slack
message, an old post. Unedited. Typos are fine and useful.]
</voice_sample>

Before writing anything, list the three most distinctive habits you
noticed in that sample, in one line each.

Then write: [TASK].

Match the voice. Do not match the topic, the format, or the length
of the sample.

Anthropic's prompting guidance calls examples "one of the most reliable ways to steer Claude's output format, tone, and structure" and recommends three to five, wrapped in <example> tags so the model can tell instructions from content. The same principle drives few-shot prompting generally, which is covered properly in few-shot vs zero-shot prompting.

3. The stripping pass

Run this on a finished draft, whether a model wrote it or you did.

You are a line editor. Do not rewrite for content, argument, or
length. Remove machine tells only.

PASS 1: Delete opening flattery, any preamble that restates the
brief, the closing summary paragraph, and the closing offer of
further help.

PASS 2: Every bullet shorter than one full sentence gets folded
back into a sentence.

PASS 3: Every three-item list of adjectives becomes two items or
four.

PASS 4: Cut every hedge that does not change the meaning:
"it's worth noting", "generally speaking", "in many cases",
"that said".

PASS 5: Count em dashes. If there is more than one per 500 words,
convert the surplus to commas, colons, or full stops.

PASS 6: Flag any sentence making two claims at once.

Output the edited text. Then list what you changed, by pass,
in no more than one line each.

4. Make it permanent in ChatGPT

A contract you have to paste is a contract you will stop pasting. ChatGPT's Custom Instructions live under Settings → Personalization on web and desktop, or Customize ChatGPT on iOS and Android. Per OpenAI's help centre, Free and Go accounts can save up to 1,500 characters; Plus, Pro, Enterprise, Business and Education accounts can save up to 5,000, raised from 1,500 on July 15 2026. Since November 12 2025, changes apply immediately to existing conversations rather than only to new ones.

[Custom Instructions — "How would you like ChatGPT to respond?"]

Answer first, then explain. No preamble, no restating my question.
No opening compliments.

Prose paragraphs by default. Lists only for genuinely discrete
items, and never a bullet shorter than a full sentence. No headers
under four paragraphs.

Vary sentence length. At most one em dash per 500 words.

Do not end with a summary or an offer of further help. Stop when
the answer stops.

Never say "genuinely", "honestly", "straightforward", "dive into",
"it's worth noting", or "great question".

When I share something I made, give me the strongest objection
before anything you liked.

OpenAI also ships tone presets. The GPT-5.1 announcement lists Default, Professional, Friendly, Candid, Quirky and Efficient, with Cynical still available; the Nerdy preset was retired on March 17 2026. Efficient and Candid are the two worth trying if the default reads as too eager.

Where should the fix actually live?

The instruction is not the hard part. Remembering to apply it on every task, in every tool, six weeks from now is the hard part.

This is the part of the problem Prompt Architects exists for, so treat what follows as interested testimony. Our own usage data from 2,170 customers, analysed 15 July 2026, says the average customer uses 1.16 of our 7 features, and that adoption falls off a cliff after the first one: 69.7% use the enhancer, 23.8% get as far as saving anything to a library. The people who never save the contract are the people who go back to pasting adjectives.

Save your style contract as a template with the task as a variable, and it stops being something you have to remember. The mechanics are the same whether you use our library, a notes app, or a text expander. Just pick one. Building a brand-voice context covers the version that persists across every chat rather than sitting in a single prompt.

What does not work

Worth stating plainly, because these consume a lot of people's time.

"Write like a human." An adjective, not a specification. The model resolves it against its average of human writing, which is where the tells came from in the first place.

Turning the temperature down. It changes how adventurous token selection is, not the structural habits. On Claude you cannot even try any more: Anthropic's Messages API reference marks temperature, top_p and top_k as deprecated, and states that models released after Claude Opus 4.6 reject non-default values with a 400 error.

Thesaurus swaps. Replacing "utilise" with "use" is good editing and does nothing about symmetry, hedging or the shape of the paragraph.

Detector-driven rewriting. See the classifier numbers above. You are optimising against a tool whose own author withdrew it.

Asking the model whether its output sounds like AI. It will agree with you, which is the sycophancy problem answering a question about itself.

The thing that does work is unglamorous: decide what your writing sounds like, write that down as rules a machine can check, and keep the rules somewhere you will actually reuse them. Everything else is a workaround for not having done that. If you want a worked example of the rules-first approach applied to a specific role, persona prompting is the closest neighbour to this piece, and the free prompt enhancer shows what the same idea looks like as a one-click operation.