# Why do AI drafts sound like AI, and what fixes it?

AI drafts sound like AI because chat models are tuned toward one polite, complete and slightly long register, and they know nothing about you, the reader or the channel, so they fill the gaps with safe defaults such as a greeting, a hedge and an offer of more help. What fixes it is showing the model your own writing and telling it where the message is going, either in the prompt, in saved instructions, or through a voice tool such as Ownhand that does it on every draft. After that, cut anything you would never type.

By Jordan Gibbs, founder of Ownhand at Thalient Labs. Updated 2 October 2026.

Readers rarely need a tool to spot an AI draft. They notice it does not sound like the person who sent it. Six things cause that. Each one has a fix you can apply today, and most of them are free.

## 1. Chat models are tuned toward one register

A chat model starts as a base model trained on a huge amount of text. Then it gets tuned to follow instructions and be helpful. That tuning gives it a house style.

A 2025 PNAS study by Reinhart and colleagues compared human text and model text on the same prompts. The instruction-tuned models wrote in a noun-heavy, dense style even when asked to match informal speech. The gap was wider for instruction-tuned models than for the base models they came from. It held in larger models too.

The model has one voice, and it's not yours. Telling it to be "casual" moves it a little. Showing it how you write moves it a lot.

## 2. Longer answers were rewarded

Chat models are tuned with feedback from human raters and from reward models trained on those ratings. Singhal and colleagues found that reward models are easily swayed by length, and that a reward based on length alone reproduced most of the gains of the full method in the settings they tested. A model trained that way learns that more words tend to score better.

In a message, it shows up as the reason stated twice, a thank-you on top of a thank-you, and a closing line that offers more help. The fix is a word cap taken from what you normally send. If your Slack messages run 15 words, say so.

## 3. Everyone's drafts look alike

The same tuning narrows what a model writes. Kirk and colleagues found that this kind of training significantly reduces output diversity compared with plain fine-tuning. Millions of people now get drafts from a handful of models, so readers have seen the same moves many times.

That repetition is measurable. Kobak and colleagues tracked words whose use jumped after chat models arrived, and estimated that at least 13.5% of 2024 biomedical abstracts were processed with a model. Words such as "delve" and "showcasing" became tells because so many texts used them at once. The phrases readers now spot are listed in [the phrases that make a message read as AI-written](https://ownhand.dev/answers/ai-writing-tells).

## 4. It does not know where the message is going

A Slack DM to a teammate and a first email to a stranger need different lengths, greetings and formatting. If you don't say which one it is, the model picks a safe middle. You get an email-shaped message with a greeting, a sign-off, and sometimes bold labels that won't render.

Anthropic's prompting guide also notes that the style of your prompt carries into the reply. A prompt full of Markdown tends to get Markdown back.

Name the channel and the reader in plain words. "A Slack DM to Sam, who I work with every day" gets a different draft than "an email".

## 5. It cannot see the thread

A reply has to fit the conversation it joins. If the thread is lowercase, short, and full of shorthand, a reply with full sentences and a greeting stands out. Most assistants never see the thread unless you paste it, so they write the reply like it's the first message.

Paste the last few messages, including your own. Then ask the model to match how you write there.

## 6. It never sees what you sent

You fix a draft, send the fixed version, and move on. The model never sees that version. The next draft repeats the same habits. Memory features in ChatGPT and Claude can save a preference you state out loud, but a habit you fix quietly every time is unlikely to reach them.

## There is no single human version

Removing the tells gets a draft to neutral. What makes it sound like you is how you write. The same status update, rewritten by two demo writers:

These examples are real output from the live API. Each AI draft was rewritten by a demo Hand: a fictional writer (an engineer, a founder, an ops lead, a designer or a consultant) built from a handful of samples.

**AI draft:** "I'm pleased to report that the migration is approximately 80% complete. We remain on track to finish by Friday, barring any unforeseen issues."

**Demo Hand, ops:** "Migration: 80% done. On track to finish Friday."

**AI draft:** "I'm pleased to report that the migration is approximately 80% complete. We remain on track to finish by Friday, barring any unforeseen issues."

**Demo Hand, founder:** "The migration is about 80% done. We're on track to finish by Friday, unless something comes up."

Both keep 80% and Friday. The ops lead writes a label and two fragments. The founder writes full sentences and keeps a softer version of the hedge. Each sounds like one person. They don't sound like each other.

Padding is the other half. This decline loses the stock framing and keeps the reason:

**AI draft:** "Thank you for thinking of me for this opportunity. After careful consideration, I've decided to respectfully decline, as my current commitments would not allow me to give it the attention it deserves."

**Demo Hand, engineer:** "Thanks for thinking of me. I'm going to pass. My current commitments don't leave room for it."

## What fixes each cause

| Cause | Free fix | What Ownhand does |
| --- | --- | --- |
| One house register | 3 to 5 of your messages in the prompt, a Project or custom instructions | Rewrites from your Hand, built from 5 to 10 things you wrote |
| Length | A word cap taken from your own messages | Each occasion has a cap, such as 60 words for `chat_dm` |
| Shared stock phrases | Read the first and last lines and cut | A checker flags them, and the model retries up to 2 times |
| Wrong channel | Name the channel and the reader | 12 occasions set greeting, sign-off and formatting |
| No thread | Paste the last few messages | Matches up to 30 messages, never stored |
| No feedback | Update your saved instructions by hand | Learns from the exact text you send |

Step-by-step settings for ChatGPT, Claude, Cursor and Copilot are in [how to stop AI drafts sounding like AI](https://ownhand.dev/answers/stop-ai-drafts-sounding-like-ai). If you write most messages yourself and only need the odd draft, those free fixes are enough.

## Fixes that do not work well

- Asking it to sound human. The model doesn't know you, so it guesses what casual sounds like and adds exclamation marks and filler.
- Describing your voice in adjectives. "Warm but direct" fits half the people you know. Samples have the detail adjectives leave out.
- A long list of banned words. The model swaps in a near synonym and keeps the sentence shape. A long list also crowds out the samples that would have helped.

## Where Ownhand fits

Ownhand handles the six causes in one call. Your agent drafts as usual, then sends the draft and the occasion to the write tool. The rewrite comes from your Hand. It fits the occasion, matches the thread on replies, and keeps every name, number, date and link. Your agent shows it to you. Nothing is sent without your OK. When you send, your agent reports the exact text, and the Hand learns from it. See [how Ownhand learns from what you send](https://ownhand.dev/answers/how-ownhand-learns-from-edits).

## Follow-up questions

### Why does ChatGPT open every email with a greeting and close with an offer to help?

It's tuned to be polite and complete for a reader it knows nothing about, so it adds the safe opener and the safe closer. Tell it the channel, paste a message you sent to the same person, and most of the extras go.

### Do newer models still sound like AI?

The PNAS study found the style gap held from smaller models to larger ones. It was wider for instruction-tuned models than for base models. A newer model writes cleaner prose and still defaults to its own register until you show it yours.

### How many writing samples does a model need?

Anthropic's prompting guide suggests three to five examples to steer tone and format. Ownhand asks for 5 to 10 things you wrote, from the places you write most, like Slack, email and pull requests.

### If I remove the AI words, will the draft sound like me?

It'll sound neutral. What makes a message yours is mostly length, how you open and close, and punctuation. Those come from your own samples.

## Sources

Checked on 2 October 2026. If something here is out of date, email [hi@thalientlabs.ai](mailto:hi@thalientlabs.ai) and we will fix it.

- [Reinhart et al., Do LLMs write like humans? (PNAS, 2025; arXiv version)](https://arxiv.org/abs/2410.16107): the style gap is larger for instruction-tuned models than base models, and persists in larger models
- [Singhal et al., A Long Way to Go: Investigating Length Correlations in RLHF (COLM 2024)](https://arxiv.org/abs/2310.03716): reward models are swayed by length; a length-only reward reproduces most RLHF gains in their settings
- [Kirk et al., Understanding the Effects of RLHF on LLM Generalisation and Diversity](https://arxiv.org/abs/2310.06452): RLHF significantly reduces output diversity compared with supervised fine-tuning
- [Kobak et al., excess vocabulary in LLM-assisted biomedical abstracts (Science Advances, 2025)](https://arxiv.org/abs/2406.07016): at least 13.5% of 2024 PubMed abstracts processed with LLMs, found through excess style words
- [Anthropic: prompting best practices](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices): examples steer tone and format; the style of a prompt carries into the output
- [Wikipedia: Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing): editors' field guide to the patterns in model text
- [Ownhand source code](https://github.com/jordan-gibbs/ownhand): occasion presets, thread matching, the checker and the revision rounds

## Install

In Claude Code it is one line. You sign in through your browser, so there is no API key to copy.

```
claude mcp add --transport http --scope user ownhand https://ownhand.dev/mcp
```

For Claude, ChatGPT, Cursor, VS Code, Codex, Gemini CLI, Windsurf, Zed, Goose and Cline, see [Connect](https://ownhand.dev/connect). New accounts get $1 of free credit. After that you pay as you go, and each rewrite costs about half a cent.

More answers: [all questions](https://ownhand.dev/answers). Full reference: [Docs](https://ownhand.dev/docs).

---

Page: https://ownhand.dev/answers/why-ai-drafts-sound-like-ai. Index for agents: https://ownhand.dev/llms.txt