In my recent article on how artificial intelligence will shape the future of learning and work, one conclusion raised in many comments was the importance of good taste.
For example, AI agents can now write large amounts of code, but human experts still seem better at deciding what code should be written. AI tools are good at finding solutions, but are less useful at figuring out the best problems to solve in the first place.
In other words, today’s AI may be skilled, but it lacks taste.
but what exactly yes taste? And if taste matters (and may be the remaining bastion of human creativity), how do you actually acquire it?
What exactly is taste?
Good taste is hard to define, but easy to understand. Taste is the ability to discern good ideas from bad ones, promising opportunities from dead ends, elegant solutions from purely practical ones.
In cognitive science, there are two competing views on the nature of taste.1
One school of thought holds that taste is just a subset of general expertise. This is the view of Nobel Prize winner Herbert Simon. He believes that humans find solutions to good problems in the same way they learn to play chess or solve logic puzzles.
Another school of thought holds that there is a difference between finding a good problem and solving it. Mihaly Csikszentmihalyi is a proponent of this view. In a book he co-authored with Jacob Getzels on art students, they found that taste (as I call it) is related to later artistic success.
I suspect both sides are correct. Taste seems to be acquired through practice, observation, and feedback like regular proficiency, but the two are distinct. It is possible to have good taste without being skilled (as a discerning critic) or to be skilled but not have taste (as a skilled journeyman dependent on the eye of a master).
The psychological mechanism of discrimination
I agree with Simon, taste is a kind of expertise. This means it relies on intuition and understanding.
Intuition sounds mysterious, but it’s mostly just memory in disguise. When we have an intuition that something is good, it’s because we pattern-matched it to examples of good things we’ve seen in the past.
This may sound overly simplistic, but compelling work in many fields shows that experts build vast libraries of perceptual “chunks,” rivaling the number of words you know in your native language, and use them to quickly form intuitive impressions about what to do in a given situation.
Given that intuition is primarily memory, this part of expertise requires extensive exposure to acquire. The vast knowledge of patterns required for expert performance is the primary motivation for the 10,000-hour rule, or the ten years it usually takes even the greatest creators to produce their first masterpiece.
Understanding is largely a process of building mental models. These models are mental simulations you generate of a situation so you can figure out the appropriate action. Unlike intuition, understanding is a process that requires attention, but it can provide more reliable results (at least when you don’t need to consider many different possibilities at the same time).
Savoring seems to involve both intuition (quick judgments about the quality of a work or the potential in a given direction) and understanding (working hard at simulation to anticipate possible design problems and opportunities).
How do you learn taste?
While some of the same psychological mechanisms may underlie taste and proficiency, there are differences.
For example, taste is largely the default. The rules that guide it cannot be easily written down. These rules, which can be expressed clearly, usually only apply in a limited number of situations and have many exceptions, which are of no help to the unintuitive person.
Skill proficiency certainly also depends on tacit knowledge, but to a smaller extent. Syntax, algorithms, and “best practices” are all easy to express, but what makes code elegant or a codebase maintainable is much harder to put into words.
This flavor of default may help explain why AI models get into trouble. LLM pre-training is essentially nothing more than book learning, absorbing all the texts in the world but ignoring any expert judgment that cannot be expressed in words.
It is sometimes claimed that the only way to acquire tacit knowledge is through direct experience. After all, if someone can’t tell you what to do, you need to figure it out yourself.
This undoubtedly plays an important role in how experts learn to judge quality. Practice with feedback will definitely fine-tune our sensitivity. (And, interestingly, this seems to be the next frontier in artificial intelligence research, with labs creating countless customized environments for agents to learn from practice and feedback, rather than simply typing in more text.)
However, I suspect that practice and feedback may only play an ancillary role in how one actually acquires taste.
Many people show taste before doing a lot of work. I’ve met a variety of writers who had less than a year of experience but quickly started gaining attention and audiences with their writing. While their proficiency improves with practice, it is clearly guided by pre-existing taste.
In other words, the reason why their practice is efficient is because their inner sense of quality already has a strong feedback signal. They don’t need audience feedback about good or bad to tell them much. (Incidentally, this is such a noisy signal that it’s unlikely to play an important role in people’s acquisition of creative skills.)
Furthermore, despite the tacit nature of taste, there appears to be a lot of direct transfer of taste between tutors and students. Artistic movements, scientific dynasties, and centers of innovation all seem to suggest that much of our taste is acquired through guided observation rather than direct experience. I learned what was good mostly by seeing good and bad examples. If we have to generate examples ourselves, few will generate anything of sufficient quality to guide the learning process.
The role of taste transmission shows that even if taste rules cannot be written down, they are communicated in some way.
My guess is that the reason taste is so difficult to teach through books is that taste is conveyed largely through the emotional judgments people make. Students in elite research labs witness their mentors’ responses to a variety of experiments, ideas, paradigms, and questions. They begin to exhibit the same emotional reactions, even if the basis for those reactions is not spelled out.
This makes the development of taste more like a process of acculturation than a training process. It’s something you gain through close contact with other tasteful people, not just through personal observation or direct experience.
Perhaps one possible effect of the rising value of taste over proficiency is the emphasis placed on the transmission of knowledge from person to person over mere book learning. It would be an ironic outcome if the democratization of knowledge on the Internet ultimately made this knowledge a commodity, with professional value derived from esoteric knowledge that can only be communicated face-to-face.
footnote
- The term “taste” that I use here is not used in the literature, but terms such as “problem finding” versus “problem solving” are used to refer to parallel concepts.
