Stop Outsourcing Your Taste to a Language Model

Ira Glass has a famous bit about the gap. When you start out, your taste is good but your work isn't, and the distance between the two is what almost makes you quit. The only way to close it is volume: you make a lot of work, and slowly your ability catches up to the taste that was always ahead of it.

Notice what that story assumes. It assumes the taste was yours to begin with, ahead of your skill, pulling it forward. That gap is the engine of getting better. So here's the question nobody's asking loudly enough: what happens to a generation of writers whose taste was never theirs in the first place, because they let a language model supply it by default? Your AI writing taste, the judgment you'd normally be sharpening on every draft, is the asset most quietly at risk, and almost nobody is defending it.

Taste is the writer's actual core asset

Strip away everything else and a writer is a taste machine. Not a sentence-producing machine, a taste machine. Sentences are cheap now; the model makes them by the thousand. What's not cheap is knowing which sentence is good, and why, and what to cut, and when the thing is working.

That's taste: the trained capacity to tell good from bad in your own domain, with reasons. It's what lets you reject the obvious metaphor for the strange one that's better. It's why two writers given the identical premise produce wildly different books. The premise is shared. The taste is not, and the taste is the whole value.

Everyone agrees with this in the abstract. The problem is that taste is invisible while it's being outsourced, because outsourcing it feels like efficiency.

Models regress to the mean, by design

A language model is, at its core, an averaging machine. It's trained to predict the most probable next token across an ocean of text, which means its native output gravitates toward the center: the most expected phrasing, the most common structure, the metaphor everyone reaches for. This isn't a flaw to be patched. It's the math. The model is a portrait of the middle of the distribution.

The middle of the distribution is, definitionally, where taste does not live. Taste is deviation from the average that happens to be better. The off-kilter image, the unexpected cut, the rhythm that breaks the pattern on purpose. All the things that make writing worth reading are departures from the mean, and the mean is precisely what the model hands you.

So when you accept the model's default, you're not getting a taste. You're getting the absence of one, dressed up as competence. It reads fine. It's smooth. It's also the exact center of everything that's already been written, which is another way of saying forgettable.

The development loop you're skipping

Here's the mechanism that should actually worry you, because it's not about any single draft. It's about the loop.

Taste develops one way: you make a choice, you live with the result, you notice whether it worked, and you adjust next time. Choice, consequence, correction. Thousands of times. That loop is how a beginner's taste, good in the abstract, becomes a working writer's taste, good in the specifics. It runs on the friction of you deciding.

Accept the model's defaults and the loop never spins. You didn't make the choice, so there's no consequence that's yours to learn from, so there's no correction. You skip the rep. Do it for a year and your taste isn't a year more developed. It's a year more dependent. You've trained yourself to recognize the model's center of gravity as "good," because it's all you've been exposed to, which is the worst possible thing for taste, since taste is supposed to be the thing that rejects the average.

This is the same disease we've described as writing with AI making people worse at thinking: the moment the default stands in for your judgment, the judgment stops growing. Taste is just judgment pointed at quality.

How to keep building yours

The fix is not to swear off the model. It's to refuse to let it make the taste calls, while using it relentlessly to sharpen the ones you make. A few concrete habits:

Make the choice before you ask. Decide what you think is the better line, the better cut, the better image, before the model weighs in. Commit to it. Then bring the model in to argue with you. If you ask first and choose second, you've already anchored on its average and your taste sat the round out.

Reject defaults on principle, then justify the exception. Treat the first output as the thing to beat, never the thing to keep. When the model gives you the expected metaphor, your job is to find the unexpected one that's better. If you end up keeping a default, fine, but only after you could say why it beat your alternative. That "why" is taste being trained.

Use the model to interrogate, not to decide. "Why might this line be working?" "What's the cliché version of this scene, so I can avoid it?" "What am I overexplaining?" These point the tool at your judgment and make you defend it. That's the coaching posture, and it's exactly why an AI writing coach is built to ask questions rather than supply answers. The questions build taste. The answers replace it.

Read more than you generate. Taste is downstream of exposure to good work, not model output. The writers with the strongest taste are almost always the most voracious readers, because they've absorbed thousands of examples of deviation done well. No amount of prompting substitutes for that intake.

Run the blind test on your own preferences. Once in a while, ask yourself whether you actually prefer the model's version or just defaulted to it because it was there and smooth. If you can't articulate why yours is better, that's a sign your taste went quiet, not that the model won.

This connects directly to protecting your writing voice, because voice is what taste produces over time: the accumulated residue of all the idiosyncratic calls you made and the model wouldn't have. Surrender the calls and there's no residue, just the average wearing your name. If you want a setup that's built around keeping the calls yours, what Polyz does is organized around that principle, putting the model in the critic's seat instead of the author's.

Your taste is the one thing worth protecting

Sentences are a commodity now. Taste isn't, and it's not going to be, because it can't be averaged into existence. It has to be built, one stubborn choice at a time, by a person willing to deviate from the center and find out whether they were right.

The model will always offer you the middle. Politely decline. Keep making the calls yourself, keep getting them slightly wrong, keep correcting. That's not the slow way. It's the only way taste was ever made.

---

Polyz uses AI to challenge your choices instead of making them, so your taste keeps developing. Start a free trial.