Language models are good at writing. Arguably too good. When you ask them to draft an email, an article or a reply, they produce fluid, well-structured, grammatically flawless text. The problem is not the technical quality of the output — it is that the output tends towards an average style: competent, but anonymous, recognisable within a few sentences to anyone who reads widely.
This is not a design flaw. It is the statistical nature of how these systems work: they learn patterns from millions of texts and generate what is most probable, most conventional, most expected. The result is useful. It is rarely distinctive.
The risk of sounding like everyone else
There is a quiet homogenisation happening in professional writing. Corporate emails, LinkedIn posts, explainer articles — increasingly, they share the same cadence, the same opening formulas, the same rhythm of medium-length sentences followed by bullet points. It is not that they are poorly written. It is that they are interchangeable.
When you use a language model without calibrating it, the resulting text aligns with that average. The AI does not know how you write. It does not know your pauses, your digressions, the way you build an argument step by step, the words you use most often or the ones you avoid. It writes well, but it does not write like you.
For someone who writes regularly, that loss of voice is not a cosmetic detail. Your voice is recognition, trust, distinction. It is what makes someone read what you write instead of another text on the same topic. And it is precisely what language models, by their nature, tend to dilute.
The phenomenon already has a name in some editorial circles: generic AI prose. Text that nobody would sign their name to, even though it is technically error-free. Correct but hollow.
Where it helps and where it gets in the way
The solution is not to abandon the tool. It is to understand with precision which functions add value and which subtract it.
Where AI genuinely helps:
Breaking through the initial block. Asking it to generate a rough structure or a first sentence removes the friction of starting, which is often the biggest barrier. The resulting draft may be mediocre; its function is to break the silence of the blank page.
Reviewing clarity from outside. Questions such as “what part of this paragraph is ambiguous?” or “is there anything I’m assuming without explaining?” work well because they ask for an external reading, not a rewrite.
Detecting problems in long texts: argument inconsistencies, logic gaps, grammatical errors that self-review misses.
Generating variations when you cannot find the right form of an idea. “Give me three different ways to express this” is a useful instruction because it keeps the decision in your hands.
Where AI gets in the way:
When it drafts the full body of a text from scratch without your prior input. The result is often technically correct and tonally foreign at the same time.
When you use its opening or closing sentences directly. Those sentences tend towards recognisable formulas that mark the text as generated.
When you ask it to “improve” a text without specifying what you want improved. It tends to soften the edge, neutralise the tone, round off the corners — the opposite of what good writing does.
When it replaces your process of thinking-through-writing. Writing is not just transcribing already-formed ideas: it is the process through which many ideas take shape. Fully delegating that process means losing it.
Strategies for preserving your own style
The key lies in where AI sits in your workflow, not in whether you use it at all.
Write first, revise with AI second. The initial draft should be yours, even if it is rough. AI can help in revision, but starting from your own draft preserves your perspective and thinking structure. If you reverse the order, the AI text becomes the starting point and you become its editor.
Define your voice explicitly. If you use a persistent context or system prompt, include examples of your writing, the patterns you want to preserve and those you want to avoid. Something as direct as “short sentences, no unnecessary bullet points, never start with ‘In today’s world’” can make a notable difference to output consistency.
Use AI for alternatives, not substitutions. Instead of “rewrite this”, try “give me three different ways to express this idea so I can choose or combine them.” It keeps creative decisions in your hands and the final text remains a synthesis that is genuinely yours.
Iterate from your text, not theirs. If a paragraph from the AI contains a good idea, extract it as a concept and rewrite it yourself from scratch in your own words. Direct copy-paste is usually the entry point for generic voice in a text that started as yours.
Work section by section, not article by article. Asking for help with a specific section, with your text as context, produces more integrable results than asking for the full piece at once.
The criterion that matters most
There is a simple test that works better than any list of rules: would someone who knows you recognise that you wrote it?
This is not a rhetorical question. If the answer is “probably not”, the text has lost something that goes beyond grammatical correctness. If the answer is “yes, even though AI helped me structure or revise it”, you have used the tool well.
This distinction is not nostalgic. It is not about defending artisanal writing as a matter of principle. It is about understanding that your own voice is functional: it generates trust in those who read you, differentiates your work from others, and makes your ideas land with more force than they would coming from a neutral, anonymous source.
AI changes the writing process in the same way word processing changed it decades ago: it makes some things easier, removes specific frictions, but cannot replace the judgement or perspective of the person writing. That perspective, well applied, remains scarce. And what is scarce, by definition, has value.
Using it well does not mean using it less. It means using it in the right place.
