Beyond Data-Driven Aesthetics frames AI aesthetic debates as reruns
MIT's Keller Gallery is hosting 'Beyond Data-Driven Aesthetics,' an exhibition by MIT Architecture alumnus Alexandros Haridis that, according to the researcher, reframes current debates about AI-generated creativity as continuations of older work. The exhibition, on view through June 30, organizes five case studies spanning Birkhoff's 1930s attempts to quantify aesthetic value mathematically, shape grammars as a rule-based alternative to data-driven generation, and AICAN's cognitive-aesthetics model for judging generated images. Haridis frames current discussions around ChatGPT, Stable Diffusion, and AI's role in creative production as repeats of questions identified in the 1956 Dartmouth Summer Research Project as one of seven dimensions of human intelligence. The practical tension the exhibition exposes is whether the field is reinventing evaluation frameworks rather than building on prior work in design computation and aesthetic theory.
The exhibition's five thematic areas, Aesthetic Measure, Aesthetic Guidelines, Algorithmic Aesthetics, Aesthetic Appropriation, and Aesthetic Novelty, function as selective 'windows' into distinct computational approaches to aesthetic judgment, each drawn from a specific publication. The source describes 'measure' as rooted in Birkhoff's 1930s mathematics, while 'novelty' examines AICAN, the machine-learning system that judges generated images using a cognitive-aesthetics theory balancing familiarity and deviation from known artistic styles. By placing a contemporary system like AICAN next to mid-century formalism, the exhibition argues that evaluation logic for machine output was theorized long before generative models became a public concern. The practical question the exhibition raises is whether AI teams running image-generation systems today are replicating these earlier frameworks implicitly, without naming them or accounting for their known failure modes.
The methodological framing matters as much as the historical content. According to the source, the exhibition treats design itself as a method of interpretative translation: software reconstruction, physical fabrication, and data visualization convert dense algorithmic and mathematical texts into spatial, interactive experiences. Haridis cites parallel practices across computer science, including neural network visualization and software reconstruction, as evidence that reconstruction techniques are increasingly used to make opaque systems more tangible. The implication is that interpretability is treated as a curatorial and material problem, not only a technical one. That framing has practical implications for teams working on explainability, but the source does not claim any specific method for translating model behavior into built form; the exhibition is positioned as a research platform rather than a validated methodology.
A separate tension sits in the comparison the source draws between rule-based methods and data-driven learning. Shape grammars, the rule-based tradition Haridis names, encode design generation through explicit syntactic rules rather than learned distributions, and the source presents them as an alternative lineage to current generative AI. The source also draws on figures including Samuel Taylor Coleridge, Oscar Wilde, and John von Neumann to argue that aesthetic value and comparison have been theorized in philosophical and literary traditions that long predate computation. This is not a claim that data-driven methods are wrong; it is a claim that data-driven evaluation of aesthetic output is not the first attempt at the problem, and that practitioners working on AI evaluation may be ignoring a body of theoretical work that already addresses questions like novelty, familiarity, and value.
The source positions the exhibition as both a research artifact and an ongoing platform. Haridis states he is interested in moving these ideas into broader applications related to the built environment and in continuing to develop the methodological role of design as an interpretative device. The closing implication is that the exhibition is meant to function as a prompt for further cross-disciplinary work, not as a settled answer. The relevant question is whether the historical continuity the exhibition identifies is descriptive, meaning these patterns really do repeat, or prescriptive, meaning the field should actively build on this prior work. The source does not adjudicate that question, and the exhibition's value is in making it visible rather than resolving it.