At MIT, a Greek architect questions who decides what is beautiful in the age of AI

Research papers, it goes without saying, are not meant arsenic objects to walk through. But aft Alexandros Haridis spent 4 years writing them, alternatively than conscionable submit them to a diary and determination on, helium started wondering what they would look similar arsenic a room.

Haridis is an designer by training, a computational researcher by practice, and a module subordinate astatine Harvard University. That combination, plan thinking fused with a precise understanding of however systems work, shapes everything astir the accumulation present connected view astatine the MIT Department of Architecture’s Keller Gallery successful Cambridge, Massachusetts.

“What you’re trying to do,” helium said, “is take a probe insubstantial that’s very dense, sometimes with mathematical formulas, algorithmic ideas, and alternatively of describing it successful words, transform it into a story that includes interaction, carnal material, and integer visualization.”

The exhibition, “Beyond Data-Driven Aesthetics,” is organized astir 5 installations, each drawn from archival sources and world literature, and structured astir a question Haridis has spent years working on: not whether AI tin create, but whether it tin evaluate. Whether it tin look astatine two things and know, successful immoderate meaningful sense, which 1 is better.

That question has a specific root point. In 1956, astatine the Dartmouth Summer Research Project, a radical of scientists gathered to specify what artificial quality should pursue. They identified 7 dimensions of quality quality that aboriginal AI probe would request to address, 1 of which was creativity and originality. Some of the radical successful that country went connected to win Turing Awards. The tract they named has since made bonzer advancement connected astir of their list. Creativity and judgement person proven harder.

Visitors interacting with a carnal wall-mounted sculpture, interpreting the work Aesthetic Measure (1933) by George D. Birkhoff, MIT Keller Gallery. [Adrian Yu, Beyond Data-Driven Aesthetics (2026)]

“There wasn’t adjacent the conception of machine subject astatine the time,” Haridis noted. “But they understood they were working connected communal subjects.” Nearly 70 years later, the accumulation returns to the items that stay unresolved, not to solve them, but to amusement however overmuch thought has already gone into trying.

The systems connected view span the 20th period and into the 21st, and each approaches the aforesaid occupation from a antithetic angle: How bash you teach a instrumentality to person taste? Haridis has organized them into 5 categories, progressing chronologically from the earliest attempts to enactment aesthetic rules connected insubstantial to today’s algorithms.

Most systems today stock a communal assumption: that a system learns by consuming vast amounts of examples alternatively than pursuing explicit rules. Given capable images, capable text, and capable internet, it learns to admit patterns, including aesthetic ones. Haridis is not dismissing that approach. He is contextualizing it. “The accumulation looks astatine antithetic perspectives astir aesthetic value and instauration that are not needfully based connected precedence,” helium explained. “It tries to marque the repertoire richer, the things you tin look astatine beyond the data-driven.”

One system successful peculiar has stayed with him. It comes from the 1980s. The thought is disarmingly simple: Instead of analyzing the qualities of a finished work, you write the instructions, oregon the “program,” that generates it. To understand the work, you study the recipe, not the dish. “It transforms attraction from the entity to the programme that creates it,” Haridis said. There is something softly extremist successful that inversion: the work of creation arsenic output, the logic down it arsenic the existent subject – a determination that has since been some influential and debated within the field.

The accumulation besides features works by Lillian Schwartz, a pioneering American creator who, successful the 1970s, became 1 of the archetypal to use computers arsenic a originative tool. Her works, drawn from the postulation of The Henry Ford depository successful Michigan, inhabit a pivotal presumption successful the amusement successful that they sit astatine the bound betwixt the earlier systems that followed explicit, handwritten rules and the image-based computation that followed. Haridis treats that transition not arsenic a rupture but arsenic a done line, grounds that the questions person ever been the same, adjacent arsenic the tools changed.

Haridis’ ain enactment follows a akin arc. Architecture, helium argued, has ever been astir taking what a decorator knows intuitively and making it legible. Giving signifier to process. Computation does the same. “I proceed to beryllium an architect,” helium said, “because the way I think is usually spatial. I can’t debar that.” The two disciplines, successful his framing, are not successful tension. They are asking the aforesaid question from antithetic rooms.

Exhibition view, MIT Keller Gallery, Cambridge 2026. [Adrian Yu, Beyond Data-Driven Aesthetics (2026)]

As generative systems flood originative industries, producing images and designs astatine a standard that would person been unimaginable a decennary ago, the questions astatine the bosom of this accumulation consciousness recently urgent. Haridis is cautious not to framework the amusement arsenic either a warning oregon a celebration. The data-driven paradigm, helium suggested, is not the extremity of the story. It is 1 section among many, and the earlier ones connection alternate models for valuation worth reading.

“These systems,” helium said of the humanities work connected view, “can pass however you determination today’s data-driven systems to caller levels, so they person caller capabilities, so they are much meaningful for a decorator oregon an artist.”

For those who want to spell deeper, a elaborate catalog accompanies the show, and wall labels supply discourse astatine each installation. But Haridis has designed the acquisition to work without immoderate of that. A visitor tin determination done the 5 installations for the spatial and visual pleasance unsocial and permission satisfied.

The statement underneath is there for those who spell looking, 1 that connects 1956 to the contiguous and insists that judgement cannot beryllium taught simply by feeding a instrumentality capable examples. That taste is not a dataset.

And if a instrumentality ever proves otherwise, Haridis suggested, that would not beryllium the extremity of the conversation. It would beryllium the opening of a harder one.

“Beyond Data-Driven Aesthetics” is connected view done June 30 astatine MIT’s Keller Gallery, 77 Massachusetts Avenue, Cambridge, MA 02139.

Gallery hours: Monday done Saturday, 9 a.m. to 6 p.m.

This accumulation has been realized with the support of organization partners, collaborating organizations, student contributors, and outer fabrication and accumulation services. For much information, visit aestheticsbeyonddata.com.