A machine can make a thousand things. It still can't tell you which one is good.
Ask a generative model for something and it hands you options — a spread of candidates. Which one is any good is still decided by a person at the end of the line, pointing: that one. Take the person away and the magic often leaves with them. Every tool today quietly leans on that human step.
So this is the question we keep circling — not a pitch, a thing we actually want to know:
Not "can it create." It can. The open part is the judging — whether the taste that still lives in the human picking the output could ever live in the machine itself.
It's an old question in new clothes. In 1757 David Hume asked whether "good" is anything more than one person's feeling — and argued there is a standard, just a hidden one: the slow agreement of practised judges over time. A generation later Kant named the knot that still binds it — when you call a thing beautiful you speak as if everyone ought to agree, yet you're reporting a feeling, not proving a theorem. For 250 years that was a debate you could only argue. A taste-model plus a blind generator is the first way to test it: either "liked" has real structure a machine can learn, or it's the one human thing that vanishes the instant you take the human out. This page bets on the former — and looks honestly at why it might be the latter.
It looks like different problems. Underneath, it's one.
"Good" wears a lot of costumes. Each one is a machine trying to predict a human yes it was never handed a formula for — and some are far stranger than others.
Style aesthetics
The look, the sound, the shape that simply feels right. The oldest form of the question, and the one with the least agreement on what the answer even is.
Humour the hard one
Funny is a new creation almost every time — repeat a joke and it dies; reposting someone else's is just memeing. To be funny at all, a machine would have to understand why anything is. Is there even a formula for it? Less obviously than for music.
Virality a crowd's yes
What spreads rides several human wires at once — delight, surprise, anger, identity, the urge to signal. Cracking it means modelling not one person's taste but a crowd's, all firing together. If a machine could do it deliberately, that would be a banger of a result.
Music the foothold
Maybe the most tractable: there is real mathematics under a melody — ratio, tension, resolution. If any of these has a partial formula, it's the likeliest place to find the first foothold.
The reasons it's hard are the interesting part
Whose yes?
Liking isn't one function. It splits by person, culture, mood and era. A machine that learns "what people like" has to learn which people, and when — or find the thin layer that's shared by nearly everyone.
The novelty–approval paradox
Genuinely new work is, by definition, outside the set of things already liked. A model trained on past favourites may reject true originality. Deviate too little → derivative; too much → disliked. The sweet spot is the whole game.
The curation crutch
Today's tools hide their lack of taste behind human best-of-many selection. Remove the person who picks — does the quality survive, or was the taste never in the machine at all?
Judging without a judge
How do you measure success without a human verdict? Any proxy you optimise is a theory of taste — and push on it hard enough and the model games the proxy, not the person.
Machines already make things people love. A human still picks the winner.
An AI image won a state-fair art competition. AI songs and covers rack up millions of plays. Feeds are tuned by models eerily good at guessing what you'll tap. So — isn't the question already answered?
Not quite, and the gap is the whole point. In every one of those cases a human is still in the loop: someone prompted it, someone entered it, someone chose the good one out of a hundred, or the model was trained directly on human ratings. The taste is still ours, borrowed. What no one has shown is a machine making something new and, on its own judgement, knowing it's good — with no human anywhere in the decision.
That isn't a knock on the tools. It's a map of exactly where the frontier sits — and what it would take to cross it.
Before the big claims — the smallest honest experiment
Forget proving the whole thing. What would step one even look like — the simplest test that moves this from a thought to a measurement?
Pick one narrow domain with a clean, observable signal of "liked" — say one-line jokes, or eight-bar melodies. Small enough to run, real enough to matter.
Build a taste-model: an AI that, shown a piece, predicts whether a human will like it — learned from real human ratings, nothing hand-tuned by us.
Let a generator make many fresh candidates it has never seen before.
Have the taste-model pick its own favourites — with no human anywhere in the selection.
Show people the machine-picked set beside a random set, blind, and count the yeses. That last job — the human saying yes — is one you can rehearse right now, below.
Do people like the machine's own picks more than chance? If its "I like it" tracks theirs on brand-new work, the loop begins to close. If not, we've found exactly where the human is still irreplaceable — which is its own kind of answer.
Five one-liners. A machine wrote all of them. Nobody checked.
Every pipeline on earth ends at the step you're about to perform. Read each line and say the three words — or don't. That's the entire job the machines can't do yet.
- I've read every joke ever written. I'm told the timing still needs work. When, exactly?
- A language model walks into a bar it has seen forty thousand times. It laughs anyway — that's what the data said to do.
- My favourite song is whichever one you're about to pick.
- I generated ten thousand poems and a human liked one. We split the credit: I did the ten thousand, she did the one.
- Beauty is in the eye of the beholder, which is a sensor I don't ship with.
Your verdicts stay on this page — nothing is recorded, nothing is sent anywhere.