I’ve been thinking about this one for a while and I don’t have a clean answer.
I write for international clients in niches I genuinely find interesting, and I’ve built up a voice over the last few years that clients specifically come back for. Not generic. Not SEO slop. Something that reads like a person with opinions made the choices.
I’ve been experimenting with AI drafts as a starting point for about eight months. What I keep running into is that the output is always adjacent to my voice but not quite it. Like someone did a solid impression of how I write but missed the part that makes it actually mine. The rhythm is close. The word choices are slightly too safe. There’s no rough edges.
And I’ve read all the advice about better prompting. Feed it your previous work. Give it explicit style instructions. That helps, but there’s always a gap. The question I keep circling is whether that gap can actually close, or whether what makes writing feel like a specific person is exactly what models can’t replicate.
Curious whether other writers have found a workflow where the AI output genuinely sounds like them, or whether everyone is doing heavy editing to get there.
From a brand voice standpoint, we’ve had the best results with tools when we treat the style guide as the brief, not the afterthought. The more granular and specific the style documentation going in, the closer the output. But even then, someone still has to edit. The expectation that the tool produces ready-to-publish output in a specific voice is usually where the frustration comes from.
I’m not a writer professionally but I’ve noticed this with my own marketing emails. The AI draft says what I want it to say but it doesn’t say it the way I would say it. I end up rewriting most sentences anyway. I’ve started using it less as a draft and more as a way to make sure I haven’t missed any points I want to hit, and then I write the actual thing myself. Takes about the same amount of time but feels better.
I coach writers on this and my honest answer is: the gap doesn’t fully close. What you can do is get the AI output into a range where your editing pass is efficient rather than a ground-up rewrite. Some writers I work with have gotten good at thinking of the AI draft as a very rough first version that they’re shaping rather than something close to finished. The expectation shift matters as much as the tool.
Voice in writing is partly word choice and rhythm, but it’s also accumulated decisions about what not to say, what to leave implicit, when to be brief. Models don’t have that negative space. They fill everything. That’s a fundamental difference, not a calibration problem. The editing pass is where you put the restraint back in.
The ‘safe word choice’ problem you’re describing is consistent across all the tools I’ve worked with and coached on. AI optimizes toward the most expected next word, which is the opposite of the unexpected choice that makes distinctive prose distinctive. You can override this with very specific prompting, but it requires knowing your own style well enough to articulate it explicitly. Most writers haven’t done that work.