A coach I work with read my line-by-line edits on his AI-written copy and told me this: "Even if I agreed with 100% of your feedback, I don't think you could edit the copy in the same way I would."
I've been building AI systems for coaches and consultants for over a year. 50+ skills. A rule file for nearly everything. And I sat with that sentence for a while, because he's right, and because I don't have a rule file for the thing he's describing.
Everyone talks about voice. Almost nobody talks about taste. They are not the same thing, and the gap between them is where the last 10% of the work lives.
TLDR
- Voice is what a system can learn from your material. Taste is the choice you'd make between 2 things that both sound like you.
- 2 drafts of the same document, his and mine, were both defensible. Neither was finished.
- I left 17 comments on 1 AI-written document. 14 were things no lint rule could have caught.
- A system that runs without you still needs a place for taste to enter. That place is smaller than you think, but it isn't 0.
- The best description I've managed is a dog and a robot, and I'll explain.
The Sentence a Coach Said That I Can't Argue With
Context, because the sentence only lands with it.
He'd been playing with the model himself and sent me a brand document it produced, saying he thought this one was much better than what we'd been shipping. I sent back a version of the same restaurant that I'd built through my system. Same restaurant. 2 drafts.
His argument for his: it was more emotionally evocative, with nuance from deep research that mine didn't have. Objectively better writing, in his words. His read on mine: less verbose, more structured, hits the important notes in a formulaic way.
My argument for mine: it may not be as evocative, but it's the core spine. It pulls on the threads that make the place one of a kind. The feel-led language is the easy part to add afterwards.
Both true. And then he said the sentence. Even with 100% agreement on the feedback, I couldn't edit it the way he would. What he's still solving for, in his words, is taste.
2 Drafts, Same Restaurant, and What Each One Had
Here's what I think was actually going on, and it took me a week to see it.
His draft had voice. Evocative, specific, researched. It sounded like something a person who loved the place would write. Mine had structure. Every section did its job, the differentiators were named, nothing was table stakes. If you ran both through a checklist, mine would score higher. If you read both out loud, his would sound better.
Neither was finished, because the finished version needs both, and the decision about how much of each is not a rule. It's a call. It's the call he makes differently from me, and the call I'd make differently from a third writer, and none of us would be wrong.
That's taste. I've written about how to get a system to your voice, and I stand by all of it. Voice gets you to a draft you recognize. Taste is what you do to the draft after you recognize it.
17 Comments on 1 Document and the Rule I Couldn't Write
Here's the version of this that scares me a little.
I left 17 comments on 1 AI-written brand document. 3 of the 17 were me writing "I like this." 14 were not. 2 were me asking what a sentence meant. One of those sentences: "that you can feel good about eating and that makes you feel good after." My comment was, roughly, is this gibberish? 2 more were on section hooks that read as table stakes. Overused. Not unique.
The one I keep pointing at was in a different document from the same night. The section meant to prove the restaurant is unforgettable had been filled with a list of that restaurant's failures. An order handed over wrong. A counter that made people feel like a transaction. The sentences were fine. The section was doing the exact opposite of its own job.
I have no lint rule that catches a section doing the opposite of its job. I've thought about it and I don't know how I'd write one. The document read well. It passed every check I own. And it was off in 14 places that only a person reading it as a person could see. That's the honest limit of the skill files I build, and I'd rather say it than pretend the checklist is the whole answer.
Where Taste Sits in a System That Runs Without You
So does this mean the system doesn't work? No. It means the system has to be designed with a small, specific place for taste to enter, rather than pretending it isn't needed.
What I proposed to that coach is a step that runs before the model writes anything. It returns 3 to 4 short phrases for the intent and the approach of each section, drawn from a growing list of table stakes to avoid, and he picks from those before the draft exists. His edits then only ever touch the language, never the spine.
That moves his taste to the front of the process, where it costs 2 minutes, instead of the end, where it costs a rewrite. It doesn't eliminate it. I don't think anything eliminates it. One study I keep seeing quoted scores AI content at about 68% of the human baseline for emotional resonance, and I don't know how much to trust that number, but it matches what I see: the last stretch is human, and the job is making it a short stretch.
The system holds the standard when the day doesn't. Taste is still yours. Both of those are true at the same time.
The Dog and the Robot
My actual words to him, when I was trying to describe the hybrid we were both circling: a cute dog that is also as efficient as a robot. Not a pug. Not a future robo-dog from a lab either.
His draft was the dog. Mine was the robot. The finished thing is neither, and the person who decides how much dog and how much robot is the person whose taste it is. I can build the robot to a very high standard. I can teach it a lot of dog. I cannot decide for him where the line goes, and the day I claim I can is the day you should stop trusting the rest of what I say.
If you've read this far and you're still hoping a system will have taste so you don't need any, I can't help you.
Just kidding. That's most of us. What I can do is get you a draft that sounds like you, so your taste only has to touch the last 10%. Send me 1 piece of your work. You get a finished asset back in your voice within 48 hours, and you'll see exactly where the dog and the robot meet.
FAQs about taste and AI content
What is the difference between voice and taste in AI content?
Voice is the set of patterns a system can learn from your material: rhythm, vocabulary, structure, banned words. Taste is the choice you make between 2 options that both sound like you. A well-built system reproduces voice reliably. Taste stays with the person, and a good system is designed to make that person's input small and early rather than large and late.
Can AI learn taste from enough examples?
Partly. Approved and rejected examples move a system much closer than rules do, and a table-stakes list catches the generic move before it is written. But a section can pass every check and still do the opposite of its job, and no rule I have found catches that. Examples narrow the gap. They have not closed it.
How do I add my taste to an AI system without editing every draft?
Move the choice to the front. Have the system return 3 to 4 short intent phrases per section before it drafts, pick one, and let it write from your pick. Your edits then touch only language, never structure. That turns a rewrite at the end into a 2-minute decision at the start.
Why did an AI draft pass every check and still need 14 edits?
Because the checks test properties, and taste tests fit. The document had clean sentences, correct sections, and no banned words. It also had a hook that was table stakes and a section proving the wrong thing. Those are judgments about whether the writing does its job, which is what a human reader is still for.