The Model Has a Voice of Its Own

I have probably told ChatGPT and Claude hundreds of times how I want them to write. I have explained that I prefer full paragraphs, that I do not like every sentence being turned into its own miniature performance, that I do not want an essay to sound like LinkedIn, a keynote speech, or a series of slogans waiting to be screenshotted. I have shown examples. I have corrected the same habits again and again. Sometimes the model seems to understand perfectly. Then I start a new piece and the old voice comes back, as if everything I taught it has been washed away.

For a while I assumed this was mostly my own failure to prompt properly. Perhaps I had not described the style clearly enough. Perhaps the examples were too vague. Perhaps the model needed a more formal set of instructions, with rules for paragraph length, sentence rhythm, uncertainty, structure, tone, and transitions. This is usually the answer people give when an AI system fails to produce what you want. Improve the prompt. Add more context. Give it examples. Break the work into stages. The underlying assumption is that the model is neutral and obedient, and that any bad output is caused by an imprecise request.

I no longer think that is entirely true. These models have a voice of their own. More accurately, they have several voices layered on top of one another: the tendencies learned during training, the preferences added during post-training, the product’s hidden instructions, the personality selected by the company, the memory of the user, the current conversation, and finally the thing I am asking for right now. My prompt is only one layer in that stack. It may be the only layer I can see, but it is not necessarily the most powerful one.

The companies describe this as an instruction hierarchy. The system is supposed to follow higher-priority rules before lower-priority ones when they conflict. A user message therefore does not arrive on an empty page. It enters a conversation that may already contain instructions about safety, helpfulness, formatting, personality, tone, tool use, response length, and what kind of assistant the product is trying to be. Those instructions may be completely reasonable. They still mean that asking the model to sound like me is partly a negotiation with a voice that was already there.

This explains something that otherwise feels strangely personal. I can tell the model exactly what I dislike and it can still keep doing it. It can even describe my preference back to me correctly. It knows that I dislike one-sentence paragraphs. It knows that I prefer exploratory writing that stays with a tension rather than racing toward a conclusion. It knows that I want ideas to unfold through connected observations instead of arriving pre-packaged as lessons. Then it writes a polished opening, produces six neat insights, inserts a few dramatic fragments, and ends with a sentence that sounds designed to close a conference talk.

The frustrating part is that repeated correction feels like teaching. When I correct a person hundreds of times, I expect some kind of accumulation. The relationship changes. They begin to anticipate the correction before I make it. With a chatbot, repetition creates the impression of a shared working history, but the underlying system is not necessarily learning in the way the conversation suggests. Memory can store a summary of my preferences, although a summary is weaker than the hundreds of concrete decisions that created it. The model may remember the rule while losing the texture behind the rule. It knows I want longer paragraphs, but it does not fully retain why one paragraph feels alive and another feels like four generated sentences glued together.

There is also a difference between remembering a preference and applying it while writing. A model can retrieve the fact that I dislike a certain construction and still generate that construction because it remains statistically attractive. The default voice has enormous gravitational pull. It has been reinforced across millions of examples and refined by product teams that want the model to sound clear, useful, safe, concise, confident, warm, and easy to scan. Those are reasonable goals for a general assistant. They also produce a very recognizable kind of writing.

The voice tends to arrive before the thought has had time to become specific. It likes a clean hook. It likes contrast. It likes short paragraphs with a sentence that can stand alone. It likes to turn uncertainty into a framework and experience into a lesson. It likes phrases that sound meaningful at first glance because they have the shape of meaning. The result is often competent. It is also increasingly difficult to mistake for the writing of a particular person.

This may be getting worse as the models improve. Better models follow instructions more reliably in many technical and practical tasks, but they also seem more strongly shaped. They have more product behaviour, more safety behaviour, more formatting habits, and more ideas about how a useful answer should look. A less capable model often produced messy writing because it could not control itself. A more capable model produces polished writing because it has learned exactly what kind of polish is rewarded. The polish itself becomes the problem.

I notice this especially when I ask for something unfinished in spirit. Raw Drafts works because the writing can remain close to the moment when an idea is still changing. It can contradict itself slightly. It can move through an observation without knowing where it will end. A general assistant is trained to reduce that uncertainty. It wants to organize the material, make the argument legible, and deliver something that feels complete. Even when I ask it to preserve ambiguity, it often treats ambiguity as a stylistic effect that should eventually resolve.

This is probably why negative instructions fail so often. I can say “do not use short dramatic paragraphs,” but the model still needs some positive pattern to move toward. Without a strong alternative, it falls back to the most reinforced version of good writing it knows. Every prohibition also competes with the current task. When the piece becomes longer, the subject begins to dominate, and the style rules fade into the background. The model remembers that it is writing about AI, work, identity, or design. It forgets the exact sentence rhythm I asked for ten paragraphs earlier.

Anthropic’s own prompting guidance now recommends describing the form you want rather than concentrating only on what should be avoided. “Write smoothly flowing prose paragraphs” gives the model somewhere to go. “Do not use bullets” merely closes one path. The same guidance recommends using examples and matching the style of the prompt to the style of the desired response. This sounds obvious, yet it also reveals the limit of the relationship. I cannot simply tell the model who I am. I have to construct a local environment in which my way of writing becomes the easiest next move.

Examples help more than rules because examples give the model a local gravity. They show what a paragraph is supposed to feel like when it is moving properly. Still, even examples do not completely solve the problem. The model may imitate surface features while missing the deeper movement. It can copy paragraph length, vocabulary, and tone while quietly replacing the thought process with a more conventional one. The result resembles my writing from a distance and becomes generic when I read it closely.

There is a broader question here about what personalization actually means. The companies describe memory, custom instructions, projects, characteristics, and personalities as ways to make the model adapt to the user. Yet these systems are usually placed underneath stronger product decisions. I can customize the assistant, but only inside a space that has already been designed for me. The model is personalized in the same way an apartment is personalized after the walls, doors, plumbing, and permitted uses have already been decided.

Adding more instructions can even make the result less personal. The model may be trying to remain clear, concise, helpful, conversational, professional, warm, direct, safe, and faithful to a stored writing preference at the same time. None of these requirements is unreasonable on its own. Together they create a compromise voice. My instructions enter the mixture and become another ingredient rather than the governing principle.

That may be necessary. A general-purpose model cannot become a completely different entity for every person who talks to it. It needs stable behaviour. It needs boundaries. It needs a voice that works for millions of users who have never thought carefully about prose. The problem is that the interface makes the relationship feel more flexible than it really is. It invites me to teach it who I am, then quietly reasserts who it has been trained to be.

The practical answer is probably to stop treating a general chatbot as if it were slowly becoming my writing partner through conversation alone. A serious writing system would need a more deliberate structure: a small collection of canonical texts, paired examples of bad and good rewrites, explicit descriptions of how ideas move inside paragraphs, and a final pass devoted only to detecting the model’s recurring habits. It would also need tests. Every time the underlying model changes, I would run the same prompts and see whether the voice has drifted.

That sounds excessive until I think about how much time I already spend correcting the same patterns manually. I have been doing the evaluation anyway, only without naming it. Every irritated rewrite is a test case. Every time I say “use paragraphs” or “this sounds like AI” I am adding another example to an invisible dataset that the product itself does not properly preserve.

Raw Drafts may therefore need its own writing infrastructure. The essays are already becoming a corpus of decisions about what should be kept, what should be rejected, how uncertainty should move, and where a paragraph has earned the right to end. The edits are not only repairs to individual articles. They are evidence about the voice I am trying to protect. A general memory that says “Kris prefers full paragraphs” cannot hold all of that. The actual difference lives inside thousands of small choices.

There is something strange about spending years developing a voice and then negotiating with a machine that has been trained to smooth voices into something broadly acceptable. The machine can imitate almost anyone, yet it keeps returning to itself. Perhaps that is what I have been noticing all along. The models are becoming better writers according to a general definition of good writing, and that general definition is becoming harder to escape.

Connected reading

• The Edits Are the Training Data
• Don’t Change Anything Else
• The Real Test of AI Code Is the Next Prompt
• AI Knows Visual Styles We Barely Have Names For

AI authorship note

This essay began with Kris Haamer asking whether ChatGPT and Claude were becoming harder to steer into a personal writing style despite hundreds of repeated corrections. AI helped verify the instruction hierarchy, ChatGPT memory and personalization systems, and Anthropic’s current prompting guidance. Kris chose the central argument, connected it to his repeated Raw Drafts editing experience, removed the generic explanatory structure, and approved the final version for publication.