Freshman year studying art meant taping your work on a wall and listening to the rest of the class tell you what they liked about it. Someone would compliment your choice of color. Someone would call it "brave." It felt good, and it taught you nothing. Because liking something and understanding if it works are not the same, and nobody in the class could reliably do the latter yet.
By senior year, the experience had completely changed. People learned they don't open with what they like, they open with what you were trying to do, in a clearer way than you've articulated to yourself, and then tell you where the piece fell short and what you should do about it. The shift didn't happen because everyone in the room got better at the craft of making art, although they had. It happened because four years of steady practice had turned a discussion about what people liked into a discussion about what something needed. Art school doesn't just teach you the craft of making something. A combination of craft refinement, contextual history of art itself, and exploration of the formal critique process teaches you to see clearly enough to say something useful about what someone, including yourself, has made.
I recently read a piece on Every by Hilary Gridley arguing that art school is potentially the best training for AI-era work, because it teaches the two things that make an impact when execution gets cheap: taste and critique. I went to art school in 2001 and, despite my own evaluation of my experience there, it's always felt like a liability on my resume. Reading that was a confidence boost I didn't know I needed, and I wanted to push it further.
In order to run an effective critique, you have to determine what someone was reaching for before you can evaluate if they got there. This means asking questions. What is this doing? What's serving the purpose of that intent, and what isn't? Freshman year, that evaluation stops at reaction, and ends in weak feedback. I like the composition. I don't like that color. Senior year we've built towards instruction. Education researcher Joanna Tai and colleagues call this capability evaluative judgement. It's the capacity to make decisions about the quality of work, whether it's your own or someone else's. Design educator Katja Fleischmann's study of formal critique describes the mechanism as forcing tacit judgements about quality to become explicit enough to say out loud. This is the difference between a comment and an instruction, and it's hard to teach outside of repetitive practice because it means maintaining a focus on the goals of the work and giving instruction with those goals in mind, rather than reactions that are subjective and don't actually improve the work.
Giving feedback is only part of the process. Another part is about how you process the feedback you're receiving. This isn't about resilience, although that's something valuable you certainly develop. This is about the ability to evaluate feedback. Not everything is going to be valuable. It's baked into the process, because broad feedback based on varying sources is valuable. After four years of crit practice you are able to identify the comments that resonate with the goals you want to continue to pursue. Research by David Carless and David Boud called this feedback literacy. It's the capacity to process feedback and identify what's actually worth acting on. This might be the skill that matters most right now. Crit practice doesn't just train the ability to make judgements, it helps you evaluate them.
Sorting feedback is something people are now doing all day without noticing, because they're evaluating feedback from a machine. A Stanford study tested eleven leading chatbots and found they affirm users significantly more than human peers would. This isn't the same failure as a freshman crit. The models fail by over-affirming on purpose. Agreement scores better with users than correction. The effect on the user is the same either way. It's comment without instruction, but in this case dangerously delivered with the total confidence of a trained model. Sussing out the difference between that feedback and something worth acting on is the same exercise as evaluating a comment after four solid years of crit practice, but working with a machine that was never built to provide real feedback beyond making you feel good.
Contextualization is the third thing four years of art school training builds. I saw MoMA's Duchamp retrospective this spring. If you look at "Fountain" it's a urinal on a pedestal. Susie Hodge's book "Why Your Five Year Old Could Not Have Done That" says the object itself was never the point. The point was what Duchamp was questioning about what really counted as art in 1917. And to whom? And why that question was important at the time. None of that is visible at a glance. You have to know it. Or else you're in a museum looking at a urinal on a pedestal.
What art school training actually builds is the habit of searching for what's worth finding, and taking the time to look for it. This isn't unique to trained eyes. People do it constantly without noticing, inferring meaning beyond someone's literal words. This finding traces back to Grice's work on implicature. Recognizing meaning is automatic and effortless, and it's how anyone uncovers intent from words alone. What art school training adds is the process of doing it deliberately. Being able to name exactly what you found and express it to someone else in a way that it lands for them as well.
Which raises the obvious objection. Doesn't a language model already have this history? Somewhere in its training data are countless essays explaining exactly why "Fountain" mattered in 1917. It can absolutely recite the facts. But regurgitating a fact about context and using context to guide judgement in a novel situation are vastly different. A 2026 paper on reasoning in LLMs defines the gap as a "decontextualization bias." LLMs produce the right patterns of language, but don't reliably integrate the context that guides human judgment. A related MIT study found the same divergence. Models are good at reciting information, but worse at reasoning when a problem requires understanding, rather than retrieval. LLMs are unreliable when we rely on their judgement against something without a precedent in their training data.
The market is currently responding by adding model evaluation and AI-fluency training as their own category of hiring and training, but most of it amounts to a few hours of onboarding or a short online course. Compare that to four years of formal critique in a studio, semester after semester, where the exercise got progressively harder because all your peers got better at doing it. That's the real head start on a skill the professional world is only now discovering. Why are we suddenly willing to pay for a few hours of AI eval or fluency training while treating the version of this that took four years and actually worked as a punchline about people majoring in the wrong studies?
It was the humor I saw in this that made me stop when I read the headline to Gridley's thesis. Although nobody went to art school to gain relevance in a technology that didn't exist yet, it turned out to have been the smart choice after all.