Longitudinal research is rare in interaction design, because it is expensive and slow and because the field’s publication cycle rewards neither. A four-wave study that has kept hold of the same practitioners for four years is therefore worth reading on method alone.
From Prefabricated Content to Situated Generation: A Four-Wave Longitudinal Study of AI-Mediated Museum Surfaces and Spaces (arXiv:2610.24076) interviewed 23 Chinese science museum practitioners annually from 2023 to 2026, retaining 19 of them across the full span, and analysed the transcripts with reflexive thematic analysis. It was posted to arXiv on 16 October 2026 and its comments field identifies it as an accepted ACM ISS 2026 full paper for PACMHCI.
Authors: Yibo Meng, Xiujun Chen, Bingyi Liu and Xiaolan Ding. The arXiv page does not list affiliations.
The shift the title names
The paper’s central claim is a change in where content comes from. The prefabricated model is the one every museum professional knows: text panels, interactive kiosks, projection loops and audio guides, all authored, approved and installed before a single visitor arrives. The paper proposes that practice is moving toward situated generation, which it characterises with the line that “visitor-facing content can be produced during interaction rather than fully authored before it.”
That is a bigger institutional change than it sounds. A museum’s authority rests substantially on the fact that what is on the wall was checked. Content generated at the moment of display has not been checked by anyone, which is why the paper’s design propositions read less like a feature list than a set of brakes.
Four propositions, mostly about restraint
The paper offers four design propositions:
Provenance at the point of display. If the sentence a visitor is reading was generated thirty seconds ago, the interface has to say where it came from. This is the hardest of the four to do well — provenance notices either get ignored or get in the way — but it is the one that keeps the institution’s claim to accuracy intact.
Spatially appropriate fallback. Generation fails: the model is slow, the network is down, the request is out of scope. In a gallery, failure is not a spinner. A surface in a physical room needs a degraded state that still belongs in that room, which means the fallback is an exhibition design problem, not an error-handling problem.
Multi-intent negotiation. A museum surface is approached by a school group, a specialist, a bored teenager and a parent with a pushchair, often at once, and often literally at the same screen. A system that resolves to one intent per interaction is wrong about who is standing in front of it.
Curatorial control before generation. Not review after the fact — constraint beforehand. Curators set the bounds inside which generation may happen, which is the only version of editorial control that survives content made at display time.
Where it stops short
The paper studies practitioners, not visitors. It tells you what the people building these exhibits believe is happening and what they are worried about; it does not measure whether visitors learn more, stay longer, trust the labels, or notice the difference. That is a reasonable scope for a four-wave interview study, but it means the propositions are grounded in professional judgement rather than in visitor outcomes.
The sample is also specific: Chinese science museums. Science centres have a different relationship to authorship than art museums do — interactive exhibits are native to them, and the didactic voice is already an institutional product rather than an artist’s. Whether the same trajectory shows up in a contemporary art institution, where the wall text is often the artist’s or the curator’s own writing, is an open question the paper does not claim to answer.
Why it lands now
There is a version of this research that would have been purely speculative in 2023, when the first wave was collected. Four years later the propositions describe decisions institutions are actually making, and the paper’s value is that it watched the belief change rather than sampling it once. Provenance at the point of display, in particular, is the design problem the next few years of gallery interactives will be judged on.
Sources
- From Prefabricated Content to Situated Generation: A Four-Wave Longitudinal Study of AI-Mediated Museum Surfaces and Spaces — arXiv:2610.24076
- arXiv listing for cs.HC
- ACM ISS 2026