Why your AI can't fix its own writing

When you ask ChatGPT or Claude to "make this sound more human", it rewrites the draft with the same habits that made it sound like AI in the first place. Post-training (the human-feedback stage that turns a raw model into a helpful assistant) narrows what a model writes toward one typical, safe style. Researchers call this mode collapse. The model can swap words, but its sense of good writing stays the same, so the voice does too.

Why does all AI writing sound the same?

All AI writing starts to sound the same because the models have been tuned toward the same target: answers most raters would mark as good. Here's the simplified version of what happens.

The base model, trained on large swaths of internet text, is capable of writing in thousands of different voices. It's seen a lot of writing: forum rants, legal documents, love letters and personal essays.

Then, during post-training, a person (or another model trained on people's judgments) is shown two model outputs for a single prompt and asked which is better. This is done many, many times, to train the model to give more outputs like the one chosen.

The problem is, in a lot of cases the better output is going to be clear, polite, balanced and complete. That's a sensible target for a help desk answer, but the side effect is that the model's range shrinks. All the odd, specific, slightly messy choices that make writing sound like one person are squeezed out, because no single odd choice is what most raters think is best.

What is mode collapse?

Mode collapse is the name for what happens when an AI model is capable of a wide range of outputs, but after post-training, a narrow range of very similar outputs becomes overwhelmingly likely. This most likely output is known as the "mode."

Mode collapse has been measured. A 2023 paper by Robert Kirk and colleagues compared models that had been fine-tuned using RLHF (reinforcement learning from human feedback, a common form of post-training) with models that had only undergone supervised fine-tuning (SFT). The authors found that "RLHF significantly reduces output diversity compared to SFT across a variety of measures."

What determines which output is the most typical? A 2025 paper by Jiayi Zhang and colleagues, including Christopher Manning, identifies the root cause as the preference data used during post-training. The authors found "typicality bias in preference data, whereby annotators systematically favor familiar text." In other words, when people rate model outputs, they tend to prefer text that feels familiar, so the model gets rewarded for writing the most familiar possible version of everything.

However, the model still has access to a wider range of outputs. Using a prompting technique the authors call verbalized sampling, where the model is asked to list several possible responses along with their probabilities, they increased diversity in creative writing by 1.6 to 2.1 times compared to prompting the model directly. This can be useful for brainstorming, but it doesn't help if you want the model to rewrite a piece of text in a natural human voice, which is a different problem.

Can ChatGPT humanize its own text?

Not reliably. One of the issues is that if you prompt a collapsed model to make its draft sound more human, you're going to get something based on how the collapsed model defines what human sounds like, and that definition was shaped by the same training. So you'll get things like more contractions and maybe some slightly more casual language, but it'll still have the same even rhythm, the tidy structure and its favorite words.

You can see it in the words models add. In our data, words like "honestly" and "genuinely" showed up several times more often when AI rewrote human writing than in the human originals. They showed up when AI wrote from scratch too: "honestly" 8.3 times as often as in the human documents, and "genuinely" 4.2 times. Those are the words a model reaches for when it wants to sound candid.

What happens when AI polishes human writing?

Even when you give a model text a person wrote and instruct it to polish it, the model will tend to rewrite the text in its own style. To verify this, we took 7,554 passages from four well-known blogs, all published before ChatGPT came out in November 2022, and "polished" each of them with 2026 models. Here is how the original text differs from the rewritten text.

AI rewrites have more uniform sentence lengths. If we measure the coefficient of variation of sentence length (standard deviation divided by the mean), we get 0.65 for the original human-written text. In comparison, we get 0.59 for the OpenAI rewrites and 0.56 for the Claude rewrites. The lower it is, the more uniform the sentence lengths: the steady beat readers notice even when they can't name it.

The bloggers were more likely to start sentences with "And", "But" and "So". In their original text, 12.0% of the sentences start with one of those words. In comparison, 9.7% of the sentences in the OpenAI rewrites start with them, and only 5.0% in the Claude rewrites. That was partly a blogger habit: across the wider range of human writing in our from-scratch study, it was 3.8%, against 3.4% for the AI drafts.

The AI models use their own favored words even when rewriting someone else's text. For example, the word "genuinely" is 12.5 times more common in the rewrites than in the original text. Other examples: "plenty of" (7.2 times more common), "simply" (4.7 times more common) and "for instance" (almost never used in the original, but about 0.25 per 1,000 words in the rewrites). See more words that AI tends to overuse.

Note that we didn't tell the models to write in an AI-like way; we just told them to polish the text. And apparently they interpret polishing as pushing the text toward the mode.

What happens when AI writes from scratch?

The same pull toward one style shows up here too, and most of it is formatting. In a second study, four 2026 models (Claude Sonnet 5.5, GPT-5.4 mini, Gemini 3.8 Flash, GLM-5.3) were prompted to produce 1,177 drafts based on provided notes for real human-written documents, asked to write in the same kind of writing, tone, and length as the original document. Without being prompted to do so, the models added formatting, writing on average 10.1 bolded phrases, 4.2 header elements, and 8.9 bullet point lines per 1,000 words (yes, they even bold things in emails). The human documents, for comparison, averaged 2.4 bullet point lines.

In addition, the models used about six times as many em dash characters as the people: 4.31 per 1,000 words, compared to 0.72 in the human documents. (People use dashes in other ways too, and when those are included the AI models still use 1.6 times more dashes.) The models also wrote sentences of more similar length than humans did; the coefficient of variation in sentence length was 0.47 for the AI models and 0.50 for the humans. GPT-5.4 mini was the most homogeneous at 0.43. The difference is smaller than in the rewriting study, though the two studies are not directly comparable because they involve different texts and use different metrics.

Why don't prompts like "write like a human" fix it?

Prompts help with the surface but not the shape. You can tell the model not to use em dashes, give it a list of words not to use, or even tell it to write shorter sentences, and most current models will respond well to that. That's worth doing, and for a short note that will only be read by your team, it may be enough. We've given some of these prompts in our guide to making ChatGPT sound more human.

But the issue is that those rules only remove the tells you know about. For everything else, you're still leaving the model to use its default cadence, structure and phrasing. And if you prompt out its most typical phrasing, it's just going to go for its second most typical phrasing, which is only one word away from the default.

What actually works?

There are two ways to do this, and both put a different judge on the final pass.

A human editor. Have a person rewrite it: someone who reads the draft out loud, cuts the safe filler, varies the rhythm and adds the specific detail only they know. Yeah, it's a pain, but if you need to send something that carries your name and really matters, like a launch post or a hard email to a client, that's still your best option.

A different model for the rewrite step. When the AI rewrites its own output, it's grading its own homework, so the fix is a second model with a different sense of what good writing looks like. The approach Zavus takes is a separate model, trained on how humans write, that rewrites the finished draft your AI agent hands it. The agent still writes the draft and makes sure the facts are right, then passes it to Zavus, which only changes how it sounds.

AI draft

Hi Sarah, honestly, I just wanted to reach out and share a quick update on the project. We've made genuinely great progress this week, and the new dashboard is roughly 80% complete. There are plenty of small details left to polish, but the core features are working smoothly. For instance, the export tool now handles large files without any issues. I truly appreciate your patience as we finish things up, and I'm confident we'll have everything ready for review by Friday. Let me know if you have any questions!

After Zavus

Hi Sarah, Just wanted to give you a quick update on the project. We've made some great progress this week, and the new dashboard is about 80% done. There's a bunch of little details to take care of, but the big features are all working. We fixed the export tool so it can handle large files now. Thanks for being patient while we work on this. We should have something for you to review by Friday.

Note: you should always review rewrites, as details can be changed regardless of who or what does the rewriting. Read it once for meaning before you send it.

How do I tell if a rewrite still sounds like AI?

Read your writing out loud, and check if your sentence length varies. If every sentence is about the same length, chances are it still sounds like AI. Also check for word tells ("genuinely", "plenty of", "for instance", "simply", "honestly"), and whether all your paragraphs end with a neat one-line summary. Our list of signs of AI writing covers the rest, and for an in-depth guide, check out our pillar guide on making AI writing sound human.

If you're writing a lot of content with ChatGPT, Claude or Cursor, you can use Zavus as a separate rewrite step inside the agent you already use. Try it free for 5 days and compare it with your current prompt on your own drafts. If you're a founder writing your own updates and posts, you can get started with Zavus for founders.

Frequently asked questions

Can I fix mode collapse with a better prompt?

Partly. Prompts can remove the tells you know about and ask for more varied sentences, but the model still falls back on its default style for every choice you didn't specify. A different model or a human editor changes the default itself.

Is mode collapse the same as model collapse?

No. Mode collapse is a narrowing of one model's outputs, usually caused by post-training. Model collapse is a separate idea: models getting worse when they're trained on too much AI-generated text.

Does every AI model have the same tells?

No. They share the even rhythm, but their word habits differ. When rewriting, Claude Opus 5.5 used almost no em dashes, while GLM-5.3 and gpt-oss-120b used about four times as many as the bloggers they were rewriting. Writing from scratch, Claude Sonnet 5.5 used 0.09 per 1,000 words and GLM-5.3 used 11.13.

Will the next model fix this?

Some tells fade as models change. "Delve" has almost vanished from 2026 rewrites and drafts. Others take their place, because the training pressure toward typical text is still there.

Sources

1. Understanding the Effects of RLHF on LLM Generalisation and Diversity, Kirk et al., arXiv, October 2023 (revised February 2024)
2. Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity, Zhang et al., arXiv, October 2025 (revised July 2026)
3. AI writing tells in 2026: data from 7,554 rewrites and 1,177 drafts, Zavus research, November 2026