Establishing Interfaces for Provisional Computing

Verdict

Yes—with a deliberately modest novelty claim. “Interfaces for provisional computing” can become a useful HCI concept, but not because it identifies an untouched class of mechanisms. Previews, undo, branches, alternatives, uncertainty displays, mixed initiative, and editable intermediate representations all have long histories. The contribution is to identify their shared object of design: computational states that are useful now but not yet final in meaning, validity, continuation, singularity, or consequence.

The term has the greatest chance of uptake as a strong concept: intermediate-level knowledge between a particular system and a predictive theory.1 It needs exemplars, counterexamples, a design space, patterns, and measures. Without those, it will become a loose synonym for “uncertain AI” and add terminological debt rather than knowledge.

1. Method and naming evidence

This review used the project-local literature-review and scholar-search skills. An OpenAlex pass screened 100 records returned for “provisional computing interfaces HCI”; 17 were shortlisted by title and abstract and all had full-text routes. Five targeted searches added 200 records around alternatives, uncertainty, mixed initiative, concept formation, and subjunctive interfaces. Exact-title searches repaired known retrieval gaps. The synthesis combines abstract-level screening of the shortlist with targeted full-text analysis of the papers supporting its central historical and empirical claims; it is integrative, not a systematic full-text review of all 17 papers. Search state, candidates, shortlist, reasons, fetch plans, and evidence levels are retained in the project.

The phrase evidence is telling. Searches did not reveal an established HCI field called “provisional computing.” Instead they found local uses: a provisional input collaboratively refined by a grounding interface,2 a provisional “design world” in reflective conversation, and probabilistic systems described as unstable or non-final. The closest mature named ancestor is subjunctive interfaces.3

This absence is an opportunity, but not proof of value. Names matter when they improve retrieval, comparison, and design reasoning. Chignell argued early that HCI terminology and taxonomy are infrastructure for cumulative scientific activity.4 The burden is therefore practical: does this name connect literature that currently fails to find itself, and does that connection produce better questions?

2. The phenomenon already exists

Reduced commitment

Cognitive Dimensions established provisionality as the ability to try things and defer commitment, countering systems that demand decisions before their consequences can be understood.5 This gives the proposed concept a legitimate root—but a dimension is not yet an interface research program.

Alternative computational worlds

Subjunctive interfaces turn alternatives into simultaneous processing realities. Their evidence is nuanced: initial users were more satisfied but not faster because scenario setup was difficult; after redesign and repeated use, users completed tasks 27% faster, and scenarios became reusable information holders.3 Parallel Paths adds a crucial refinement: alternatives need independent and shared operations, not merely snapshots.6

Preview before consequence

Persistent command previews, feedforward, staged edits, and operation-level undo separate inspecting an outcome from accepting it. Parallel prototyping further shows that keeping several options alive can improve diversity, expert-rated outcomes, click-through performance, and self-efficacy relative to serial iteration.7

Editable intermediate computation

Interactive machine learning already treats the human–machine process as co-adaptive rather than a one-way command.14 AI Chains makes model steps and intermediate results inspectable and editable; users calibrate expectations, compare downstream strategies, and “unit test” components.8 Spellburst branches and merges generated creative code.9 DirectGPT combines manipulable generated objects, scoped commands, and undo, making participants substantially faster while shortening prompts.10

Approximation before validation

Interactive Speculative Planning separates a fast approximator from a slower target agent and lets a person interrupt. Its UI must reschedule asynchronous outputs so users do not see invalid downstream work as settled.11 That status-and-lineage problem is not captured by “uncertainty visualization” alone.

3. A discriminating definition

A useful definition must identify both a state condition and an interaction capacity. A computation is provisional when at least one relevant aspect is non-final; an interface supports provisional computing when users can act meaningfully on that non-finality before commitment.

The interaction capacity can be summarized as mark, inspect, fork, compare, revise, validate, commit, invalidate, repair. Not every interface needs every operation. But merely labeling a result “draft” or showing a confidence score is insufficient if the user cannot do anything appropriate with that status.

4. The neighboring-concept map

The term earns space only by preserving distinctions. It should function as connective tissue, not a conqueror that renames every adjacent field.

5. What the term can add

A common comparison unit

Current papers evaluate a preview, an AI chain, a branch canvas, or a confidence display within their local traditions. Quickpose shows how versions can become material for reflection and goal formation rather than mere backup.15 A shared vocabulary would let researchers compare which forms of non-finality they expose, how status is represented, when users may intervene, and what constitutes commitment.

Commitment as a first-class HCI object

HCI has mature vocabularies for action, feedback, error, trust, and undo but less consolidated language for the transition from candidate to authoritative state. Agentic systems make this gap visible: a plan can be generated, partially executed, verified, approved, invalidated, or compensated. “Commit” is not just a button; it is a social and technical boundary.

A bridge across design and AI

Giaccardi and colleagues argue that data and algorithms destabilize the artifact: probabilistic and agentive systems participate in making rather than merely being stabilized by design.12 Provisional computing can connect that design-theoretical insight to concrete interface mechanisms for status, alternatives, validation, and effects.

A design target beyond “human in the loop”

Human-in-the-loop often specifies presence without specifying useful control. A provisional interface asks sharper questions: what exactly remains open, what can the person alter, which downstream states depend on it, and what event makes effects durable?

6. The research program the term must enable

RQ1 — Sources and representations

How should interfaces represent intentional, plural, epistemic, computational, representational, and consequential non-finality? Can users distinguish them and choose the right remedy?

RQ2 — Provisionality budgets

How many live alternatives, caveats, pending validations, or staged actions can users productively manage? When does openness become branch debt or decision paralysis?

RQ3 — Propagation and semantic merge

When a shared region changes, which alternatives inherit it? How can AI-generated branches be merged by intent or evidence rather than line-by-line text?

RQ4 — Commitment quality

Can interfaces help people commit the right state at the right time? Candidate measures include premature-commitment rate, unresolved-risk awareness, invalid downstream work, repair cost, and appropriate acceptance conditional on correctness.

RQ5 — Collaborative status

A state may be a draft to an author, a proposal to a reviewer, and production input to an automation. How should rights to mark, revise, validate, or commit be allocated?

RQ6 — Accessibility and scale

Parallel views and marginal status cues can be visually powerful but navigationally expensive. What nonvisual structures preserve comparison, lineage, and status? How should systems summarize hundreds of branches?

RQ7 — Politics of finality

Who benefits when a system declares a result final, and who bears the labor of keeping it open? “Provisional” can protect inquiry, but it can also deny recognition, defer accountability, or make workers repeatedly verify machine output.

7. How to establish the term

Terms are established through useful work, not branding. A credible path has five cumulative steps:

The first paper should avoid a manifesto tone. It should present a falsifiable proposition: systems with identifiable non-final states will differ systematically according to source, available operations, lineage, and commitment boundary; matching interface support to those properties will improve commitment quality at an acceptable management cost.

A minimum viable design space

Code each system on six axes: source of non-finality; granularity of the provisional object; plurality and branch topology; visibility of status and dependencies; agency to revise or validate; and commitment consequence/reversibility. If these axes do not help reviewers classify disputed examples or designers discover missing controls, the concept has not earned its keep.

8. Failure modes

Renaming the literature

The most serious risk is citation erasure: presenting decades of subjunctive interfaces, previews, versioning, and design exploration as if generative AI created the problem. Any concept paper must put Green, Petre, Lunzer, Hornbæk, Terry, Mynatt, and related lineages at its center.

Becoming synonymous with AI uncertainty

If every low-confidence model output is called provisional, the term contributes nothing. The decisive test is whether status is coupled to an action or withheld consequence.

Conflating technical and social finality

A model can mark output “final” while an institution treats it as advice; a draft can become de facto policy before formal approval. Interface state, database state, organizational authority, and public consequence may diverge.

Assuming more alternatives are better

Subjunctive interfaces and LLM design-space tools both reveal attention and setup costs. Beyond Code Generation found broader exploration but also difficulty keeping up with LLM-originated changes.13 Provisional interfaces need compression, curation, and closure.

Declaring a field too early

A workshop, website, or catchy definition does not create a field. The appropriate first ambition is a strong concept and shared design space. A field claim should wait for independent adoption, distinct methods, recurring venues, and results not readily produced by neighboring traditions.

9. A near-term agenda

The most valuable first empirical study is likely a 2 × 2 × 2 commitment experiment: provisional versus conventional interface; correct versus incorrect system candidate; reversible versus consequential task. Measure final correctness, time, premature commitment, unnecessary rejection, unresolved-risk recall, and repair. This directly tests whether the concept predicts more than a generic “give users control” heuristic.

Conclusion

“Interfaces for provisional computing” has a viable intellectual niche. The phrase names a recurring but fragmented condition: computation is increasingly available for use while its interpretation, plurality, validity, continuation, or effect remains open. Existing HCI gives us many parts of the answer, but distributes them across vocabularies that rarely compare their assumptions.

The path forward is not to claim an empty territory. It is to build a better map. If the term clarifies boundaries, predicts interface needs, reveals missing commitment controls, and supports comparative evidence across design tools, AI systems, and agents, it can become useful intermediate-level knowledge. If it merely rebrands previews and uncertainty, it should be abandoned. That contestability is not a weakness of the proposal; it is the standard the proposal should set for itself.