Eric Rodenbeck, a lecturer in architecture at the Harvard Graduate School of Design (GSD) and founder of Stamen Design, is making the case that AI is fundamentally mischaracterized by the design industry—and that mischaracterization is costing us something valuable. In an essay published on Harvard Design Magazine, Rodenbeck argues that treating AI as a productivity tool focused on efficiency and automation misses what makes it genuinely interesting: its potential as a medium for inquiry, critique, and new forms of knowledge creation.

The Archive as Testbed

For the past two years, Rodenbeck has been running a GSD course called "Re-imagining the Archive" where students work with collections from MoMA, Harvard Art Museums, the Institute of Black Imagination, and the American Museum of Natural History. The premise is deliberately uncomfortable: these archives are supposed to stabilize knowledge, but Rodenbeck's students treat them as something you can "work on, work through, and sometimes work against." The goal isn't cleaner visualization or faster access—it's seeing what happens when you stay with material long enough that its seams start showing.

What Happens When You Ask the Same Thing Twenty Times

One of the most striking student projects came from Roy Zhang, who graduated from the Master of Design Studies program in 2025. Zhang took a single archival image from Harvard's Houghton Library and asked ChatGPT to generate twenty keywords describing it—then did it again, and again, twenty times with identical prompts. What he found reveals something fundamental about how these systems behave under repetition: early responses are wide, exploratory, slightly strange. Over successive iterations, they collapse toward narrower, more conventional descriptions. Variability decreases. Language standardizes. The model becomes more confident—and less interesting. This is treated as a feature if you're optimizing for efficiency. But Rodenbeck sees it differently: "If you are using it as a design medium, that is a problem—and an opening." Zhang's project transforms prompting from something you perform into something you can analyze. He's not chasing better outputs; he's mapping trajectories—the shape, drift, and bias toward the familiar that these systems produce over time.

Re-Indexing by Idea Instead of Episode

Kevin Tang and Yuanqing Xie, also MDes graduates in 2025, built a project for the Institute of Black Imagination that takes this interrogation further. Working with Dario Calmese's podcast archive, they used AI to break conversations into individual sentences, analyze semantic content, then rebuild the entire collection as an alternate audio player—browsable by idea rather than episode. Themes surface across conversations. A sentence spoken in one context finds resonance in another. The result feels less like a player and more like a landscape you navigate by association. At one point they had AI generating interstitial sentences between podcast segments, which "sounded like someone trying way too hard to sound cool by mediating in between." They caught it, iterated, pushed further. The important move isn't the use of AI—it's treating its output as something to work on rather than accept.

Misuse as Method

Rodenbeck's third principle might be the most counterintuitive: "The most interesting projects almost always slightly misuse the technology." MDes students Cindy Emefa Coffee and Ieva Lygnugaryte took archival photographs from MoMA where both authors and subjects are unknown—images already unstable in their attribution—and used AI to animate the figures exiting the frame entirely. It's a simple gesture, but it lands hard. Archives depend on containment; objects get cataloged, stabilized, held in place. By animating subjects walking out of the picture, these images become even less fixed. The frame fails to hold them. What you're left with isn't a better description of the image—it's a different relationship to it, a sense that the archive cannot fully account for what it holds. Technically straightforward, conceptually difficult: using AI not to fill gaps but to widen them in controlled ways.

Prompts as Sketches

Rodenbeck draws an analogy from GSD lecturer Edward Eigen who compared LLMs to the Talmud—not because they're sacred, but because they're dense, generative, and endlessly interpretable. You don't read them once; you return, argue, annotate, build traditions around them. That framing cuts against most AI discourse in design circles, which still sounds like management consulting: efficiency, acceleration, staff reduction, faster pipelines. The distinction matters operationally. If you treat AI as a tool, you ask "how do I get the right answer?" If you treat it as a medium, you ask "what happens if I push this?" A prompt becomes provisional—a sketch you can revise, distort, overwork—rather than an instruction returning correct or incorrect output. The point is not to nail it on the first try; you're trying to see what these models do under pressure.

Key Takeaways

  • AI "settles" under repeated prompting, collapsing toward conventional language—useful as a feature, but also a design problem worth interrogating
  • Treating outputs as something to work through rather than accept opens new forms of spatial and visual knowledge
  • Showing your prompts isn't a footnote; it's part of the work, revealing the dialogue, back-and-forth, and moments where things clicked or didn't
  • Misuse can be method: pushing systems against their intended use reveals properties that efficiency-focused approaches miss entirely

The Bottom Line

The design industry's AI rush toward "faster pipelines" is missing what makes this stuff actually interesting. Rodenbeck's framing—that these systems are closer to ink than automation—isn't just philosophically elegant; it's practically useful. When Google Maps launched in 2005, he thought map design was finished. Instead it sparked a golden age of web mapping as the technology became democratized and designers pushed into new territory. The same could happen here—but only if we approach AI with more ambition than "better throughput." Treat it like ink. See what bleeds.