Why do AI models sound so different from each other?

Different models sound different because each one has a house style that was trained in on purpose. One leans chatty and explanatory. Another leans short and direct. What you hear when a draft feels off isn't AI-ness leaking through. it's a default voice, and it belongs to the model, not to you.

In April 2025, researchers at Carnegie Mellon University measured how strong that signature is. Their classifier read a piece of text and named which of five models wrote it with 97% accuracy, using nothing but word choice. Earlier attempts at telling just two models apart managed 60 to 70%.

So the fix isn't to sound less like AI. it's to stop sounding like that particular model.

Why the same prompt sounds different on every model

The five models in that study were ChatGPT, Claude, Gemini, Grok and DeepSeek. Each produced text carrying its own signature, and the signature held up under pressure. The researchers scrambled it, rephrased it, translated it and summarised it. The personality survived all four.

ChatGPT trended toward detailed, explanatory prose. Claude trended toward concise, straightforward answers. Those are opposite instincts, and you have felt both of them even if you never put a name to the difference.

What makes this more than a curiosity is what the researchers were not testing. They were not separating AI from human. They were separating AI from other AI, and the signal was massive. Your writing is competing with a 97% fingerprint, and so is your ear.

We could use characteristic word choices to determine which large language model generated a particular bit of text with 97% accuracy. Mingjie Sun, Carnegie Mellon University, April 2025.

Each model has a default voice, and it was designed on purpose

OpenAI publishes a Model Spec, and the May 2024 version spells out how the assistant behaves when nobody says otherwise. One documented setting is called interactive. When it's true, and it defaults to true, the assistant uses markdown formatting and a chatty style with clarifying questions.

The same document argues that default behaviour should stay stable over time, because unstable defaults make a tool harder to learn. Read that again. The sameness is a feature somebody chose, not a failure somebody missed.

Anthropic described the same idea from another angle in a September 2023 post about Claude's character. Character training, the tuning that happens after the main training run, was added from Claude 3 onward. The stated goal was to seed traits like curiosity, open-mindedness and thoughtfulness.

Two labs, two very different personalities, and both of them written down. When a draft comes back bland, you're hearing a product decision.

What the default voice actually sounds like

Readers rarely describe a model. They describe patterns. The cleanest catalog of those patterns is maintained by Wikipedia editors in the Signs of AI writing guideline, which the AI Cleanup project keeps current and which was still being revised in August 2026. It lists negative parallelism, rule-of-three overuse, stacked transition words, title case, and paragraphs of tidy, positive, verbose prose.

That guideline is careful about one thing, and the caution is the most useful line on the page. The patterns are potential signs of a problem, not the problem itself. Sanding off the signs while the thinking stays mechanical just makes the writing harder to spot.

Keep that distinction. If you spend an afternoon deleting every tell from a hollow draft, you end up with hollow prose that no detector can name. For the working version of that list, sorted by how often each pattern shows up in real drafts, start with our breakdown of common AI writing patterns.

Why asking for less formal makes things worse

Here's the loop most people run. A draft comes back stiff, so you ask for something less formal. It comes back casual, and now it reads like a blog post from nowhere. So you ask for a bit more polish. It stiffens again. Three rounds later you have a worse version of the first draft.

Nothing broke. You moved the model along one axis and it answered with a different default. Its casual voice and its formal voice are both its own. Less formal isn't a setting that reveals you. It's a setting that selects a different house style.

Register is how formal the writing is. Provenance is who the writing sounds like. Those are two dials, and the instruction everybody reaches for first only turns the register dial.

That's why the request keeps failing politely. How to sound less formal in AI writing takes the same problem from the other direction, and the useful part of it isn't the list of fixes. It's that formality was never the thing being detected.

The two axes people keep mixing up

Draw register left to right, formal on one end and casual on the other. Draw provenance bottom to top, the model's voice at the bottom and your voice at the top. Every draft you write with AI lands somewhere inside that square.

The trap is trying to fix a problem that lives on the vertical axis with a move along the horizontal one. You can slide along register all day and never leave the model's voice.

The four corners are easy to recognise once they have names:

• Formal and yours. A proposal written the way you write proposals. High register, high provenance. This is the target.
• Formal and the model's. The same proposal with every tell intact. It reads like a policy document, and nobody quite trusts it.
• Casual and yours. A message to a colleague. Short, specific, slightly rough around the edges. Low register, high provenance.
• Casual and the model's. The one that opens with a rhetorical question. Low register, low provenance. This is what asking for less formal buys you.

Two of those fail and they sound nothing alike. Both failures are somebody else's voice wearing a different shirt.

Does the model you write with matter?

For mechanics, yes. For voice, no, and a 2025 study in PLOS ONE shows why the distinction is easy to get backwards. A team led by Wataru Zaitsu compared 100 human-written public comments with 350 texts produced by seven language models, then ran a random forest classifier over three stylometric features. It separated AI from human text at 99.8% accuracy. That part stopped being surprising a while ago. The second half of the study is the interesting half.

In a second experiment, 403 people were shown texts and asked to judge AI or human, then rate their confidence in the answer. Human detection was limited and often no better than chance, and participants were most easily fooled by the newest model in the set.

So a machine can tell seven models apart almost perfectly while you can't reliably tell any of them from a person. That's why reading a draft and waiting to feel something doesn't work. Your intuition is tuned to a signal that mostly isn't there.

How to override a default with an anchor from your own writing

An anchor is a sample of your own writing that you paste into the conversation before you ask for anything. Around 150 to 250 words, in the register the new piece needs, taken from something you already sent or published. Not a description of your voice. The actual sentences.

It works because a default is a distribution rather than a fixed output. The model samples from everything its training made likely. An anchor narrows that distribution before the first word is generated, so the model picks among your habits instead of its own.

Three lines do the work. Everything else is decoration:

• The role. Say what this is and who reads it: an update for a client who already knows the project, or a first message to a stranger.
• The anchor. Paste the sample and label it. 'Here is how I wrote the last three of these.' The label matters, because it tells the model the text is a target rather than a topic.
• The constraint. One or two rules specific to this draft. 'Sentences under 20 words.' 'don't open with a question.'

Short text needs a shorter anchor. For an email, two or three of your own sent messages is plenty, and pasting a blog post will make the email read like a blog post. Our guide to fixing robotic email tone has the two line version for messages that have to go out in the next five minutes.

One failure mode is worth planning for. In long documents the style drifts back toward the default as context grows, and it usually drifts where you're least likely to look. Why ChatGPT stops sounding like you in long documents covers the drift signals and the re-anchoring step.

Once an anchor works, save it. A short reusable instruction set beats pasting a sample into every new session, and how to train AI to write in your voice is the longer argument for that.

A five minute test to hear your model's default

You can run this today with a draft you already have. No tools, no detector, no tab switching.

• Write one sentence about something you know well. A decision you made last week, or a problem you solved. Keep it plain.
• Ask the model to rewrite that sentence three ways. Same meaning, about the same length, no other instructions.
• Read all four out loud. Your original has one rhythm. The three rewrites share another one, and usually it's the same as each other's.
• That shared rhythm is the house style. Describe it in one line, because that line becomes your first constraint.

Now run it once more with a different model and compare the two sets of rewrites. You'll get two rhythms, and you'll be able to say which one sits further from yours. That comparison is the useful output, and no detector score will hand it to you.

What changes when a new model ships

Every release tunes the defaults again. Some drift toward longer answers, some toward more structure, some toward a flatter, friendlier tone, because those are the qualities the labs measure and optimise. Your anchor survives that. Your collection of prompt workarounds does not.

Detector scores move for the same reason. When the model underneath changes, the number you got last month stops meaning what you assumed it meant. How to interpret AI detector results has the current reading rules, including the two lines you should ignore entirely.

One group pays for every model change first. Stanford research published in 2023 tested seven widely used detectors on essays by non-native English writers. 61.22% of the TOEFL essays were classified as AI-generated. All seven detectors agreed on 18 of the 91 essays, and at least one flagged 89 of them. On essays by US-born eighth-graders, the same tools were close to perfect.

James Zou, the senior author, named the cause: the detectors score perplexity, which tracks how sophisticated the writing looks, and non-native writers naturally trail on those measures. The lesson isn't that detectors are useless. It's that a detector score measures distance from an average, not authorship.

Which brings it back to the only question worth asking about a draft. Not whether it sounds like AI, but whose voice it's. Answer that one and the rest gets much simpler.

Frequently asked questions