AI Has No Taste Because It Innately Seeks the Average
Deep Thoughts on AI and Aspirations —— ByteDance Deep Thinking Circle
Open any AI-generated content platform and you’ll get a familiar feeling: the output is growing, but it increasingly looks like it all came from the same person. The color schemes are those few sets, the copy has that tone, the design drafts look like they rolled off the same assembly line. This phenomenon has a name: AI slop.
Most people attribute this to “models can only imitate, they have no creativity.” This explanation sounds reasonable but actually says nothing—it blames the model without explaining why the model behaves this way.
I recently heard an explanation that dissects the underlying logic very cleanly. It came from Thais, founder of Taste Labs, a company specifically researching “why AI fails in subjective domains.” Her conclusion can be condensed into one sentence: AI has no taste, and the root cause lies in training signals—they consistently reward the average.
Models Only Improve Where Things Can Be Measured
Start with a background observation. We habitually say “models are great at writing code,” as if this were an inherent capability. Thais sees it exactly the opposite: this is actually a property of code itself.
Code can be executed and verified. Write it wrong, run it once and you get an error; write it right, tests pass. The feedback signal is clean, clear, and directly usable for training. Design can’t do this. Ask someone to look at a page and they’ll likely just say “something feels off,” unable to articulate what; ask another person and the judgment might be completely different. Aesthetics also change—styles fashionable five years ago might be dated today. But 2 plus 2 equals 4 won’t become 5 just because the era changed.
She summarizes this difference in one sentence: capability follows measurability. Where things can be measured, models improve; where they can’t be measured, models stand still.
This framework is far more useful than debates about “whether AI has creativity.” It transforms the question from mysticism into structure: subjective domains are hard because their “good” has no objective answer, but rather a judgment dependent on people and context that drifts over time. Discussing both types of problems on the same capability map was never going to work.
The Model’s Nature Is to Seek the Average
Next is her sharpest concept, and the underlying mechanism explaining AI slop.
The model’s training method essentially predicts “the most likely next content to appear.” In math and code, the most likely answer happens to also be the best answer: 2 plus 2 equals 4 is both the most frequent answer and the only correct one—the two coincide. In design and writing, there’s a wide chasm between the most common answer and the best answer.
Truly memorable creativity almost always happens at the edges of the distribution: the ad that deliberately breaks convention, the poster that deviates from norms, the color palette others wouldn’t dare use. Meanwhile, the model’s optimization objective naturally pulls output toward the center, toward “most common.” The result is output increasingly resembling the mean: not bad, but never stunning—like something you’ve seen a hundred times before.
And this problem won’t automatically disappear with larger parameters. Bigger models just perform more precise mean estimation on larger datasets; they won’t learn on their own “when to break the pattern.” When you complain that AI-generated content has no soul, what you really should ask is: have the training signals ever rewarded “non-average”?
Breaking Taste Into Verifiable Questions
Thais offered several actionable paths, with decomposition at the core.
“Is this design good?” can’t be answered, but “does this page align with brand tone?” is far more concrete. Break down a brand: what colors are used, what fonts, how is spacing defined, are animations fast or slow, is there texture? Each individual item has right and wrong answers. An AI-generated page can be scored item by item: are the colors correct, is the font from brand guidelines, does the spacing follow the design system? The value of decomposition is that it transforms a vague “good” into a string of verifiable sub-questions, and these sub-questions themselves can serve as ground truth for training.
There’s another more hidden trap in the data. The common approach is to have a bunch of people score things and average the scores as training signals. But different people’s aesthetics differ to begin with—these disagreements are part of the real world. After averaging, you don’t get a better answer, just a mixture that represents no one’s preferences.
Her approach is to create a preference vector for each user type and bind it to the training data. The model learns not “the unified standard of good design,” but rather “for people with this preference, what counts as good.” She also handles cases where experts clash: when two experts evaluate the same design oppositely, if it’s a disagreement on style preference, the data has genuinely captured preference diversity and should be preserved; only when they clash on objective questions like alignment does it indicate the data itself has problems.
One-sentence summary: in subjective domains, data quality matters far more than data quantity. Feed in the mean, you get the mean out; feed in expert data that preserves real diversity, and you might get taste out.
What This Means for Product People
Combining these mechanisms, several judgments can be directly applied.
First, don’t count on “waiting for bigger models” to solve taste problems. Think clearly about which parts of your product have objective right and wrong, and which are purely subjective. For parts with right and wrong, confidently hand them to model iteration; for purely subjective parts, either get in there yourself and break “good” into rules, or honestly accumulate high-quality preference data. Hand vague problems to vague signals and the result will inevitably be mediocre.
Second, personalization needs to go deeper. Most products doing “personalization” now tag at the behavior level: you clicked this, I recommend that. Aesthetic-level personalization is something else—understanding a person’s judgment criteria, what they like in what context. Build this layer and products can shift from “recommending what you might like” to “helping you generate and filter things that match your taste.”
Third, the average itself is a business. Safe design that pleases most people can make money by “never making mistakes.” So the real question to answer is: do you want most people to think it’s okay, or a small group to find it stunning? Choose the latter and you need to start accumulating differentiated taste data now—this path is expensive and slow with no shortcuts, because it directly determines the ceiling of model output.
Key points: The root of AI slop is training signals rewarding the mean (most common ≠ best); code trains well because it’s verifiable, capability follows measurability; the solution is decomposing into verifiable sub-questions and using preference vectors to preserve disagreements; the mean is also a business—first decide whether you want the safety of mass appeal or the brilliance of niche appeal.