Texterial: A Text-as-Material Interaction Paradigm for LLM-Mediated Writing, a paper by MIT Media Lab and Microsoft Research researcher Jocelyn Shen with Nicolai Marquardt, Hugo Romat, Ken Hinckley, Nathalie Riche, and Fanny Chevalier, received an Honorable Mention (top 5%) at CHI 2026 in Barcelona. Its core argument: today’s LLM writing interfaces are rigid and linear — a text box, a prompt, a returned block of text — and that structure actively hides the rich, non-linear operations a language model is actually capable of performing on a piece of writing.
The pitch: text as something you touch
Instead of treating a document as a static string edited through typing and prompting, Texterial reframes text as a material — something with the malleability of clay, the layering of paint, or the growth pattern of a plant, and maps large-language-model text operations onto expressive, touch-based interactions with that material. The idea is that physical-material metaphors are ones people already have strong, intuitive mental models for, and borrowing those models makes an LLM’s editing, restructuring, and generation capabilities legible in a way a chat interface simply doesn’t.
Two working probes, two different metaphors
The paper presents two concrete technical probes built to test the idea. Text as Clay lets a user refine text through gestural sculpting — kneading, pulling, or reshaping a passage the way you’d work a physical material, with the LLM interpreting those gestures as edit operations. Text as Plants takes the opposite temporal approach: ideas grow serendipitously over time rather than being sculpted in one deliberate session, treating a piece of writing more like something cultivated than something authored in a single sitting. Together, the two probes stake out a design space between immediate, hands-on control and slower, emergent development — letting the same underlying LLM capabilities surface through very different interaction rhythms.
Why this matters beyond one CHI paper
The broader stakes here reach past writing tools specifically. As generative AI systems accumulate more and more latent capability, the interface layer between a person and that capability increasingly determines how much of it actually gets used — a powerful model hidden behind a narrow chat box leaves most of its capability undiscovered. Texterial is a concrete example of a research direction gaining traction across CHI and SIGGRAPH this cycle: designing interaction metaphors specifically to reveal AI capability, rather than defaulting to conversation as the only interface. For anyone building creative tools on top of LLMs — not just writing tools — it’s a useful case study in what that kind of interface design can look like in practice.