Civic AI / Research R&D / 2026

2030canvas

Product design · Interaction · Engineering

A canvas-first civic research tool where residents can reshape the assumptions behind a planning scenario, factor by factor.

What changes when residents can edit the assumptions behind a planning model, instead of only reacting to its conclusions?

A public planning conversation starts in everyday language, but the systems used to evaluate it are made of factors, relationships, and assumptions. Somewhere in that translation, residents can lose authorship of their own ideas. 2030canvas turns the representation itself into the conversation: a shared diagram people can move, question, annotate, and revise. The crucial design decision was to stop treating the diagram as an output screen and make it the main workspace.

An ongoing human-centered AI and civic-participation research effort at Florida State University's School of Information, situated in conversations about Leon County's long-range planning. The interface is a facilitated research instrument, not an official county service or a validated forecasting product.

01

Start with residents, not a template

Fifteen exploratory stakeholder interviews informed the research questions and early interface needs. The canvas now begins with no preset local scenario. A participant can describe a preferred future or build a representation manually, without first accepting the research team's framing.

02

Make assumptions manipulable

Factors are draggable objects with editable definitions, possible measures, and provenance. Participants can add or remove factors, connect them with directional positive or negative relationships, reverse an arrow, and explain why a connection might exist.

03

Make disagreement part of the artifact

The interface distinguishes suggestions from confirmed or revised elements. Contextual inspectors expose each claim for review; session notes and a revision history help an interviewer capture where a resident agrees, hesitates, or pushes back. Sessions can be exported for later analysis.

A model people cannot question is not a shared model.

The research began with 15 exploratory stakeholder interviews and early thematic analysis. The manual canvas is a design response to recurring concerns about who defines the problem, how information is understood, and whether technology makes public participation more meaningful. These themes inform hypotheses for upcoming work, not proven outcomes.

01

Participation versus token input

Residents may be invited to comment without being able to influence how the problem is framed. Editing the model is an attempt to make that framing negotiable.

02

Information asymmetry

Planning representations can hide definitions and dependencies behind technical language. Clickable factors and relationships make the assumptions available for inspection.

03

Trust in AI interpretations

AI may help translate a vision into structure, but it can also introduce assumptions. Provenance labels and explicit human revisions keep that translation open to challenge.

Research in. Revisions out.

→

01 / Listen

Scenario interviews surface residents' priorities, uncertainties, and language before the research team imposes a diagram.

→

02 / Represent

An interviewer and participant construct factors and directional relationships directly, optionally beginning with an AI proposal.

→

03 / Contest

Participants move, define, add, remove, accept, and revise elements while describing what the representation gets wrong.

→

04 / Learn

Researchers examine exported revisions and interview observations to refine the interface and formulate later modeling questions.

Next: run facilitated walkthroughs of the canvas-first version, examine how residents interpret factors and links, and compare where manual edits or AI proposals help or hinder discussion. Quantitative land-use simulation and validation remain separate research stages.

The canvas became the interface

The earlier staged, panel-heavy scaffold put instructions around the model. I moved the graph to a full-width reasoning surface, collapsed the vision prompt to a compact pill, and made inspectors appear only when a factor or relationship needs attention. The object of discussion stays visible while it is being changed.

Direct manipulation comes first

A participant should not need to persuade an AI to make a correction. The Factor and Connect controls, drag handles, editable properties, and reversible relationship directions work directly. AI can propose a starting interpretation when configured, but it is one possible entry point, not the owner of the model.

An elegant arrow is not evidence

The interface marks where a factor came from and whether an element is proposed, revised, or accepted. Positive and negative links are hypotheses for discussion, not calculated effects. That distinction prevents a fluent or polished diagram from masquerading as scientific certainty.

Leave room for an unanticipated future

No Lake Jackson or other example scenario is preloaded into a real interview. Starting empty reduces the risk that a convenient sample quietly defines which planning priorities matter, while still allowing a separate illustrative portfolio preview.

The result is a working, canvas-first interview instrument: a spacious causal diagram, direct factor and link editing, optional AI-assisted interpretation, contextual inspection, local session saving, notes, revision history, and JSON export. This redesign is a testable design decision, not a demonstrated improvement in civic participation. The next task is to see what residents actually understand, contest, and change when given control of the representation.

What this does not claim

The initial 15 interviews guided exploration; they were not a usability study of this redesigned interface. The pictured portfolio scenario is illustrative, not participant data. Proposed causal links are not measured impacts, no validated land-use projections are generated, and browser-local session storage is not an approved research database.