Taste Is a Hypothesis:
Why Aesthetic Judgment Is Falsifiable Too

The Claim Hiding Inside "I Like This"

Almost no real aesthetic judgment is atomic. "This draft doesn't work" is a bundle wearing one sentence as a coat, the same way "this plan won't work" is. Unbundled, it turns into several separate, checkable claims, and the bundle is only true if at least one of them actually holds.

A note that says "this doesn't work" gives the maker nothing to act on, because there is nothing to check. A note that says "the stakes aren't clear until paragraph four" gives them a specific, testable claim, one they can go verify for themselves by handing the piece to someone cold and watching where attention drops. The difference between useless taste and useful taste is almost always just this: whether the judgment has been unbundled into something with a truth condition.

What Would Prove It Wrong

Once a claim is isolated, the next question is the one that makes it a real hypothesis rather than a restated opinion: what would show this is false? Three tests do most of the work, across writing, design, and anything else built to be experienced by someone else.

The Outside Reaction

Your own read of your own work is the least reliable data you have, because you already know what you meant. An outside reader, viewer, or user doesn't have that advantage, which is exactly why their confusion is informative. If you predicted "the button will get missed" and three people out of five miss it, the prediction held. If nobody misses it, it didn't, whatever your instinct said walking in.

The Time Test

Does the piece read the same on the tenth pass as it did on the first, or on day thirty as it did on day one? A judgment that only holds while the work is fresh, and the maker's excitement is still doing some of the reading for them, is a weaker hypothesis than one that survives distance. Time is a slow, honest test that most people skip because waiting is uncomfortable.

The Subtraction Test

Remove the part you are proudest of and see if the whole still stands. If it doesn't, that part might be load-bearing, evidence the earlier judgment was right. If it does, and stands just fine, that's evidence the attachment was personal rather than structural, which is its own useful, if less flattering, result.

Trained Pattern-Matching, Not Magic

None of this makes taste less real or less valuable; it explains where it actually comes from. A good editor's instinct that an opening won't land is not a mystical sense. It is a model, trained on thousands of past openings and what happened to each one, compressed into something that fires fast enough to feel like intuition. That is not so different from what a well-trained AI model does with its own training data: pattern-matching from prior cases to a prediction about a new one. The instinct is real, and it is also, structurally, a claim about the world that happens to run faster than conscious reasoning.

Said out loud like that, the "instinct" is obviously a prediction, and obviously checkable: is P₁ actually true, and does this draft actually match the pattern, or does it do something different enough that the old cases don't transfer? A model that was trained mostly on thrillers may misfire on a draft that's deliberately, structurally slow. That is not the instinct being wrong so much as the instinct being applied out of its distribution, which is exactly the kind of failure worth naming and checking rather than either blindly trusting or dismissing outright.

Where This Actually Helps

The payoff is the same one formal decomposition gives a stuck argument: it moves the disagreement off the unarguable conclusion and onto the checkable premise. "I don't like it" invites only agreement or a shouting match. "I predict readers will lose the thread here, and here's the pattern I'm basing that on" invites a test both people can actually run. Sometimes the test proves the note right, and the maker has something specific to fix instead of a vague cloud of dissatisfaction to absorb. Sometimes the test proves it wrong, and the model that produced the instinct just got a genuinely useful update. Either way, the conversation ends somewhere more honest than where it started.