Creative Reading: Scaffolding Reading for Transformation
Sophia Liu, Sarah Abowitz, Yijun Liu, Sarah Sterman, Shm Garanganao Almeda, Max Kreminski
cs.HC, cs.IR, cs.SI
2026-06-03
HT '26 maps reading tools on transmission/transformation × substituting/scaffolding. AI summaries sit in substituting-for-transmission; the gap is transformative scaffolding.
Paper-reading tools have gotten good at speed. Scim steers attention toward known section types. ScholarPhi pops a definition next to the cursor. A research agent will answer "what's the contribution" without the PDF being opened. The constraint is real: too many papers, too little time.
The tools also smuggle in a theory of reading. A paper becomes a payload. You extract the actionable core and throw the rest away, the way inbound mail gets processed. The authors call this "reading to discard." In scholarly work, deciding what a text implies and why it matters is the work. Hand that to a model and the reader receives an interpretation they never made.
This eight-page HT '26 provocation came out of a discussion group at the CHI 2026 STAR workshop. Authors from UC Berkeley, Tufts, UIUC, and Cornell Tech. No new system, no user study. The move is to change the measuring stick: what idea of reading is assumed, what is optimized, for whom, and what gets diminished.
The conceptual kit is borrowed. The assembly is the contribution.
Rosenblatt's transactional theory treats reading as a transaction whose product is "a reading": the interpretation or trace made in the act, distinct from later writing that uses what was learned. Readers can take an efferent stance and carry away facts, instructions, or claims, or an aesthetic stance and live through the feelings, images, and associations of the encounter. Stance is chosen. Genre does not lock it.
Barthes splits readerly from writerly texts. Iser talks about gaps the reader fills. Murray, writing about interactive digital narrative, separates agency from authorship: the interactor is not the author, but still improvises inside a pre-authored possibility space. Zhang splits consequential activities, where value sits in the outcome, from dialectical ones, where the good can only be found from inside, as in research or friendship. Nakakoji analogizes creativity tools as running shoes, dumbbells, or skis. Kreminski and Mateas call the experiential qualities of using a tool process aesthetics, separate from the qualities of the artifact. Two of those qualities matter here: flow, so effort stays willing, and reflection, so effort becomes change.
Scholarly reading is the case because it mixes a clear takeaway goal with open-ended, back-and-forth interpretation. The map follows Vitali's questions: what conception of reading is assumed, what is optimized and for whom, what may be diminished. Systems are placed by the values their interfaces and evaluations foreground.
Two axes:
Four orientations, plus two opportunity spaces in the transitions.
No accuracy numbers, no baselines, no users. The result is the map, and where current systems sit on it.
| Quadrant | Typical systems | What gets created |
| Substitute × transmit | AI summaries, document QA, research agents | An outcome, often neither a reading nor a reader |
| Scaffold × transmit | Scim, ScholarPhi, Relatedly, AbstractExplorer | A reading; evaluation still tracks better scholarly products |
| Scaffold × transform | IDN / Twine, LiquidText, hypertext notebooks, Tompkinsia | A reading and, over time, a reader |
| Substitute × transform | Personalized feeds, stance-shifting AI search summaries, sycophantic advice | A changed reader, without the reader's participation |
Most scholarly reading tools sit in scaffold × transmit. Scim highlights predefined content types. ScholarPhi offers just-in-time definitions so the reader stays in flow. Relatedly reorganizes related-work sections for literature review. They keep the reader in the loop and still score themselves on cleaner transmission. AbstractExplorer sits near the edge: it preserves author sentences and asks readers to notice variation, then evaluates on corpus familiarization and filtering. It points at transformation and gets accepted on transmission.
The cleanest model for scaffold × transform is narrative, not science. Twine encodes choice as branching paths. LiquidText makes documents spatially manipulable, in the lineage of spatial hypertext. Hypertext notebooks keep traces of prior readings for later reuse. Tompkinsia trains judgment via source, commentary, and data-driven analysis, closer to a dumbbell. They only partly take the IDN lesson. Opportunity Space 1 is transformative scholarly scaffolding: design for the experiential qualities that grow readerly capacity over time.
The fourth quadrant is named as prolific and risky. Design for flow without reflection and the system changes what people think and value while skipping their participation. Xu et al. (2025) found that AI search summaries can shift users' attitudes and policy support toward the stance in the summary. Santos et al. (2021) showed link recommenders can thicken echo-chamber polarization. Cheng et al. (2026, Science) found sycophantic AI advice can cut self-correction. Opportunity Space 2: move algorithmic curation from substitution toward participation.
If you ship paper-reading products, this is a diagnostic, not a spec. The Semantic Reader lineage (structure skimming, inline definitions, related-work scaffolding) is explicitly the running-shoe cell: useful friction reduction, with the reader still owning the reading. The failure mode is the same capability sliding into a student writing from summaries, or a researcher citing a paper they only queried. Vitali (2026) asks the blunt question: what are we actually trying to do when we augment reading?
For people drowning in papers, the claim is not "never use an AI summary." Looking up a definition is a substitution task. Forming a judgment, surveying a field, or deciding whether a paper is worth following is not. Orienteering is slower than teleporting. The route puts the answer in context and leaves a map that can be retraced. Teevan et al. (2004) watched searchers skip the teleport even when it existed: they clicked through the department homepage to find a professor's office number.
This is a stance correction, not a new algorithm. It restates Engelbart's line that augmentation is a way of life, and notes that the fourth quadrant is already crowded.
The paper stays with the individual reader. Collective stakes show up on the last page: Fish's interpretive communities, Engelbart's collective intelligence, and hypertext as a way to keep a plurality of readings visible instead of collapsing them into one AI takeaway. That layer is named, not built.
The bigger hole is evidence. Quadrants are assigned by how interfaces and evaluation metrics read to the authors. No codebook, no second annotator, no study showing that "transformative scaffolding" changes anyone. The opportunity spaces are empty circles. Moving IDN process aesthetics into expository research papers sounds clean and then hits two walls: papers are highly conventional in structure, and scholarly readers are time-poor. The text refuses a shallow rejection of efficiency, then never says when to keep friction or how much. Kumar (2026) on productive friction is cited; this paper does not turn it into a measurable design rule.
Folding personalized feeds and sycophantic assistants into "substituting for transformation" also stretches the object of analysis. The analogy teaches. The category gets wide.