36 AI writing UIs score 1-10 on Collage as the page breaks into cards

Collage is the New Writing: Exploring the Fragmentation of Text and User Interfaces in AI Tools

Daniel Buschek

cs.HC, cs.CL

2024-05-27

Buschek scores 36 AI writing tools on four collage facets; the factor runs from 1 to 10, and high-scoring UIs replace the page with arrangeable cards.

What problem this solves

Generative models reached writing software before a reusable design concept did. HCI now has a pile of prototypes, sidebars, cards, canvases, pop-ups, and still lacks the intermediate idea Höök and Löwgren asked for. It has to sit above any one screen and stay concrete enough to compare screens. Calling the whole area human-AI collaboration is too vague for that job.

Word processors kept a page at the center. A toolbar does not share that rank. Generative tools split the center. Prompt boxes, continuations, outline nodes, and color for human versus AI text now shape the draft about as much as the page does. Fragmentation is already the interface. It has no shared name, so teams have no shared rule for whether the pieces should stay visible.

Method

Daniel Buschek, at the University of Bayreuth, uses a DIS '24 essay to move literary collage into tool design. In the literary sense, the writer stops originating every line and assembles fragments that already exist. The transferred idea is called Collage. It names fragments in the interface and in the writing process. The finished email or story is allowed to look smooth.

Four facets carry the transfer. The UI turns one page into cards. Interaction switches between diegetic text, the draft, and non-diegetic text, instructions and notes about the draft. Both require an intention and then a string of words, so they are hard to run in parallel. The artifact mixes the user's prose, model output, and the training corpus, and most UIs never mark the source. The user moves from writing lines to selecting and arranging them. Draxler et al. call the matching experience the AI ghostwriter effect: felt control tracks with a sense of ownership.

The concept is then used three ways: to inventory existing screens, to sketch designs that either keep the fragments or refuse them, and to set those screens against German literary Expressionism from about 1910 to 1925.

Results

There is no new model and no user study. From Lee et al.'s survey of 115 writing assistants, plus a forward search, the essay scores 36 tools in 32 papers. Each tool answers 11 yes/no questions. The sum is a Collage Factor, and the tools run from 1 to 10. The essay says this is not a systematic survey.

Low scores do not ask the writer to prompt. CatAlyst only pops up one suggested sentence when someone is procrastinating. The middle of the range is where cards, covering pop-ups, mixed sidebars, and human/AI color show up. VISAR arranges arguments on a 2D node diagram and syncs them to a page draft. It meets all four facets, and its factor is 10. HaLLMark records contribution with color and a human/AI share. Kim et al. split the UI into cells, generators, and lenses: fragmented input, outside material, and judgment one level up.

On-screen text falls into four types, human or AI crossed with content or talk about content. The design proposals are mockups. Keep the seams, and an accepted continuation stays clickable so a rejected line can be restored; a prompt collapses into a summary of that passage. Drop the seams, and the screen holds one sentence at a time, close to freewriting, while the model cleans the text in the background. The critical section adds no measurements. It restates three known risks: prose turns more template-like, people write less, and suggestions distract.

Why it matters

A writing product can keep growing sidebars. The earlier choice is whether fragments should remain visible. Visible seams help trace sources, check accidental copying, and show a collaborator what was turned down. Showing less at once suits a stuck writer who needs to draft. Both ideas are sketches. There is no quality or speed comparison against an ordinary page.

Ownership is the part a product team can use now. Once the main job is picking and arranging, people more readily feel the draft is not theirs, and they still resist saying that AI was involved. The fuller the generator, the more the UI has to say who wrote a sentence and whether that choice can be undone.

Limitations

The 36 tools are academic prototypes chosen to illustrate a concept. ChatGPT is treated as a writing tool in the broad sense and is missing from the score table. Assistants inside everyday document products are not scored.

Collage Factor adds 11 yes/no items with equal weight. One author iterated the criteria. The essay reports no agreement check. Extra pop-ups raise the score, which need not match how fragmented the screen feels.

The mockups were not tested with writers. The essay itself says that reopening an old choice may stall the draft. Expressionism is an analogy. The three worries come from studies already published, and collage does not measure them again. The concept stays on process fragments because most users want a finished text that looks intact. Whether a later document viewer should show human/AI joins to readers is left open.

Terms

Source

What people are saying

Related papers

All paper explainers