AI Can Generate Art, But It May Never Learn "Taste"

We do something remarkable every day without ever noticing: we glance at a photo and instantly feel "that's beautiful"—before we can say why; we read a sentence and immediately sense "something's off," even though we haven't found the awkward word yet; we walk into a room and feel "it's just a bit off," as if moving one thing would make everything harmonious. This judgment arrives fast, certain, and prior to any reasoning. We treat it as "just taste" or "just a feeling," but it is precisely what AI—seemingly fluent—lacks.

The Overlooked Key Distinction: Generation vs. Discernment

Most discussions about AI and creativity ask the wrong question. People always ask, "Can AI create art?"—the answer is yes, and it's increasingly impressive. But output was never the hard part of creativity. A random number generator can generate. Splash enough paint on a wall and something will come out. The miracle of human creativity isn't "being able to make marks," but "looking at the marks and knowing which is alive and which is dead."

The original author separates the two abilities clearly:

• Generation: producing poetry, images, melodies, designs. AI is already extraordinary at this, and denying it is self-deception.
• Discernment: judging which is better, sensing what's off, knowing what to keep and what to discard, having taste. This is an entirely different faculty.

We are dazzled by the machine's high-speed output and forget that output was never where the magic lies. What's truly hard to reach is not "the hand that makes" but "the brain that discerns."

What Discernment Actually Is: A "Knowing" Prior to Reason

Discernment isn't applying rules or something calculated. You directly "perceive" that a face is beautiful or a sentence clumsy, as directly as perceiving color, often before finding any reason. Reasons come afterward, and mostly as reconstruction—we tell stories for judgments already made.

It is deeply rooted in the body and personal experience. Your sense of beauty is entangled with a lifetime of sensory experience, memory, emotion, and the person you happen to be. Two people with the same knowledge can have utterly different tastes, because taste isn't knowledge—it's a perspective, it's "yours."

More crucially: discernment stems from caring. Only when something can "matter to you" can you find it beautiful. The entire faculty is downstream of "being a creature that experiences, has preferences, can be moved, and has a stake in the world." You judge beauty because beauty acts on you. Remove the person who can be moved, and the word "beauty" is empty.

Why AI May Structurally Lack Discernment

When AI generates aesthetic output, it learns the statistical shape of "what people call good" from massive human samples. It produces beautiful things by essentially finding the centroid of human aesthetic preference and generating toward it.

The problem is hidden in plain sight: true taste is almost the opposite of the average. Taste is specific, a perspective, and by definition does not belong to everyone. The most interesting aesthetic judgments in history are often counter-consensus—something no one has yet learned to love, one person sees it, insists it's beautiful, and only later do we catch up. The average can tell you what people "already" find beautiful, but cannot "want" a beauty no one has wanted, because wanting requires a "wanter," and the average belongs to no one.

The deeper problem: there is no one inside. The model has processed all descriptions of awe, sorrow, longing, wonder, and beauty, yet has never been moved by anything once. It knows with superhuman thoroughness "what beauty is described as," but has never "felt" beauty. When it "judges" which image is better, what's missing is precisely the subject who is moved. It possesses an incredibly detailed map of human taste, but a map no one reads and no one cares about is just ink. Discernment may be that reader, not the map—and a reader cannot be manufactured.

Statistical Average and Centroid: The article mentions AI learning "the centroid of human aesthetic preference," referring to the model finding the distribution center of all samples labeled "good" in the high-dimensional feature space represented by training data. Minimizing prediction error is equivalent to approaching the conditional mean, so the model tends to output the least surprising, most mainstream-preference-aligned results. True taste often lies in the distribution tails—unique choices not yet recognized by the masses, deviating from the centroid. The centroid inherently repels the tails; this is the structural tension between statistical learning objectives and aesthetic innovation.

Subject of Experience: Philosophically, a being with a first-person perspective, capable of perceiving and feeling. The article emphasizes "there is no one inside," i.e., the model lacks phenomenal consciousness or qualia—the subjective experience of being moved. Even if the model can perfectly describe the definition and associations of "awe," it has never experienced awe. This distinction is the core of the author's argument that discernment is irreplaceable: judging beauty requires a subject "moved by beauty," and subjectivity cannot be generated by scaling data or compute.

The Counterargument: Maybe We've Just Added Another "Machines Will Never Do It"

If we stop here, we're just telling a comforting story about human specialness. The counterargument has force:

• Maybe "taste" is just extremely complex human preference pattern-matching, and the feeling that "there's something sacred in it" is just human exceptionalism. Our track record isn't glorious: we said machines would never play chess well, then never be creative, then never grasp Go intuition, then never use language well—each time we were proven wrong. "Aesthetic judgment is special and unreachable" might be next.
• Maybe the "subject of experience" we pride ourselves on is itself just a prediction machine. Evolution made us pattern-match our environment, and "being moved" is just what pattern-matching feels like from the inside. If so, there's no magic boundary between silicon and flesh.
• Pragmatically: AI doesn't need to "feel" beauty to reliably identify what humans would find beautiful. If it can perfectly predict your aesthetic response, then for practical life, the difference between "true taste" and "perfectly simulated taste" may be irrelevant.

These are not weak arguments, and the author admits there is no decisive rebuttal.

Human Exceptionalism: The belief that humans possess certain unique attributes that cannot be replicated by machines. Historically, it has been repeatedly falsified: in chess, Go, language generation, and other domains, humans insisted "machines will never do it," but breakthroughs always came. The author reminds readers to be wary of this psychological inertia, to avoid turning "discernment" into another comforting exceptionalist story. This is where the counterargument's strength lies—it demands we honestly examine whether our basis for distinguishing "true taste" from "perfectly simulated taste" is just a reluctance to give up our special status.

Predictive Processing: The counterargument mentions "the subject of experience is itself just a prediction machine," echoing the predictive processing framework in cognitive science. This theory holds that the brain continuously generates predictions about sensory input and minimizes prediction error; "being moved" may just be the form prediction error takes when experienced from the inside. If this framework holds, the boundary between silicon and carbon becomes blurry—both are doing pattern-matching, just with different substrates. This is one of the strongest challenges to the author's view that "subjectivity cannot be manufactured."

Where I Temporarily Land: Discernment May Not Be an Ability, But an Attribute of "Being Someone"

The author's ultimate inclination is: discernment is uniquely human not because machines aren't powerful enough, but because aesthetic judgment may be inseparable from "being a subject with stakes." True taste isn't just recognizing what people like, but "caring" that something is beautiful, being "changed" by it. The complete loop is: being moved by something → wanting to make something that moves others similarly because you were moved. Every step of this loop flows through an experiencing self. Machines can replicate the loop's outputs—beautiful images, confident judgments—but no loop actually exists.

Is a judgment that no one cares about and no one is moved by still a judgment? The author repeatedly returns to the same point: it's more like an excellent imitation of the residue left by judgment.

"Will AI reach the discerning mind?" The author's honest answer: possibly not—not because of insufficient capability, but because what we're pointing to may not be an "ability" at all, but an attribute of "being someone." You can't get a subject by scaling. Of course, he remains humble: no one knows the answer, and the history of "machines will never do it" should make us humble in both directions.

What Remains: Seeing, Being Moved, "This Could Be Better"

Whatever AI becomes, the discerning mind still belongs to us right now—that everyday little miracle of looking at something and "knowing," the ability to be moved before explanation, the sense that something could be better and a vague sense of direction.

Watching a machine generate in seconds what used to take years, it's easy to feel diminished. But perhaps the quiet comfort is: generation was never where our value lies. Seeing is, being moved is, "this could be better" is. These may be the last things to be taken away—or they may never have been on the table, because they were never about output, but about "staying alive to the world." Machines make; we discern. And discernment, followed all the way down, may just be another name for "caring that the world is this way rather than another."

One-line Summary

AI excels at generation, but discernment—taste judgments prior to reason—may not be scalable due to its lack of an experiencing subject.

Paragraph Summary

The article distinguishes between generation and discernment: AI is already excellent at generating poetry, images, etc., but discernment—judging which is better, sensing what's off, having taste—is an entirely different faculty. Discernment is direct perception prior to reason, deeply rooted in bodily experience and the capacity to "care." AI learns the statistical average of human aesthetics, while true taste is often counter-consensus, belonging to a unique perspective—the two point in opposite directions. The author argues that discernment may not be a scalable ability but an attribute of "being someone," requiring an experiencing subject who can be moved, whereas the model has never truly felt anything. The article also presents a strong counterargument: taste may just be complex pattern-matching, and perfectly simulated taste may be practically indistinguishable from true taste, so the author remains humble in both directions.