Arizona Water ChatBot

Three things to know in three seconds.

Everyone talks about water. Few know who to ask.

A statewide survey run through the Arizona Water Innovation Initiative (n = 128) showed what residents actually discuss, and how patchy their sources are. People pieced together answers from water bills, local news, neighbors, and Instagram Reels:

"Water quality. A lot of my friends scared me away from drinking the tap water." survey respondent, AWII Arizona Water Survey, 2023

That's the gap the chatbot targets: turn anxiety and hearsay into sourced, plain-language answers about drought, conservation, and the Colorado River.

Personas built from real Arizonans, not vibes.

From the survey and our own interviews we built personas of the residents the bot had to serve. Alex is the one that drove the most design decisions, a young, tech-savvy homeowner drowning in hard-water problems and scattered information:

Before designing the bot, we evaluated the brains.

We ran a comparative heuristic evaluation of the two available LLM engines (ChatGPT and Bard) through a water-questions lens. The differences shaped the conversation design more than any wireframe did:

• Editable answers beat re-prompting. Bard's built-in answer modification saved users a whole prompt cycle, so edit and regenerate controls became a core requirement, not a nice-to-have.
• Accessibility is a differentiator. Speech-to-text input widened who could actually use a public-service tool.
• The blank prompt box is a usability cliff. Without a guide, users need practice to phrase effective prompts, which argued for guided starter questions instead of an empty text field.

Moderated think-alouds, with an adversarial streak.

Each of the four of us recruited at least two screened Arizona residents (residency, age band, self-rated water knowledge, tech comfort) for moderated 30 to 45 minute think-aloud sessions, consent-formed, recorded, and structured around goals in the five Es: efficient, effective, engaging, error-tolerant, easy to learn. Remote results were collected through Maze.

• Scenario 1, first impressions: open exploration after hearing the bot exists.
• Scenario 2, the homeowner deep-dive: soft-water research with escalating tasks, including deliberately challenging the bot ("I've heard soft water corrodes pipes, is that true?") to probe trust and error recovery.
• Scenario 3, quality & regulation: ask, edit the question, regenerate, rate the answer, follow up.

Analysis was deliberately small-n honest: counts instead of percentages ("4 of 6 participants" says more than "67%"), every outlier reviewed individually, and findings triangulated across the survey, sessions, and post-test satisfaction ratings before anything earned a recommendation.

Every persona expectation became an interface decision.

The tested design, an Arizona Water Chatbot research preview in ASU's Global Futures Laboratory environment, answered the research directly:

• "I don't know what to ask"→Guided starter topics: Diving into Drought · Mastering Conservation · the Colorado River's role
• "Can I trust this?"→Curated source links (ADWR, AMWUA) under answers + a plain-spoken accuracy disclaimer
• "The answer isn't quite right"→Edit the question, regenerate, and rate answers inline
• "I'll need this again later"→Named conversation history (Water policy · Lake Mead & Lake Powell · Conservation practices)

Public-service AI is a trust problem first.

The chatbot's hardest problems were never technical. They were about whether a resident would believe, and act on, what an AI said about the water coming out of their tap. Sources, honest disclaimers, and user control (edit, regenerate, rate) did more for trust than any answer-quality tweak. That thread, how people calibrate trust in automated systems, runs straight through the rest of my research.