Research Radar: The Six-Word Problem: Will Voice Improve How We Research the Impacts of AI?

The Limits of Survey Research

When we ask public professionals an open-ended question at the end of an InnovateUS course, post-course feedback averages six words. A typical answer: "Mostly for everyday tasks at work."

That is the six-word problem. And it is the reason we built the Public Voice tool.

AI use is expanding across governments. Agencies are combining platform access, responsible use policies, and training to ensure public professionals use these tools for the public benefit. 

But is this three-pillar strategy working?

Is learning changing how public servants do their jobs? What is the impact of that behavior change on the public?

Measuring whether learning translates into impact is the most important question we set out to answer when we established our research initiative: The Observatory of Public Sector AI at InnovateUS. 

We were stuck in that tension, getting six-word answers that don't tell us enough about what public servants were actually doing differently.

The Observatory runs pre- and post-course surveys and workshop surveys on public-sector AI use. But surveys have a basic tension: if they're too long, no one finishes them. If they're too short, you don't learn enough to act on. Survey response rates are notoriously low to begin with, so we were already not hearing from most participants. We were stuck in that tension, getting six-word answers that don't tell us enough about what public servants were actually doing differently.

Survey Tools: What's Out There

We first looked at what already existed: general survey tools, voice survey platforms, AI-moderated interview tools, enterprise feedback systems, and story-collection tools. Each category solved part of the problem. 

General survey tools could collect open-ended responses, but they often produced short, generic answers. Voice survey tools demonstrated the value of speaking rather than typing, but they were not designed around InnovateUS's course improvement workflow.

AI-moderated interview tools could generate richer qualitative data, but they were built for longer research interviews, not a quick post-course reflection. Enterprise feedback platforms had strong reporting infrastructure, but they were not designed around a lightweight, anonymous learner experience.

The tool can check whether an answer is sufficiently specific and ask a targeted follow-up only when needed.

What InnovateUS needed was an evaluation tool designed to turn a brief learner reflection into more robust evidence of how training may show up in practice.

The Design Challenge

The design challenge was tuning the judgment layer. AI made it possible to add that judgment without requiring a human interviewer: the tool can check whether an answer is sufficiently specific and ask a targeted follow-up only when needed. 

The goal of Public Voice is to help InnovateUS move from short reactions to concrete evidence of use. It becomes a way to understand whether AI training is helping public servants do their work more effectively.

We focused on specificity rather than length. A long answer can still be vague, and a short answer can be useful if it names a real task or workflow. "I will use it for writing" does not tell a program team much. "I will use it to draft plain-language public notices for road closures," tells us the task, the workflow, and the kind of public-facing work the training may affect. 

That is the difference between a reaction to a course and evidence that a learner can apply the training in practice.

Follow-ups are limited and targeted — asked only when the first answer is too general to act on, and designed to invite a real example. We capped the exchange at two follow-ups because the experience still has to feel like a check-in, not an interview.

The experience itself is straightforward. You finish the course. A link takes you to a short feedback experience. After a landing page and a consent screen, you move through a few standard survey questions. When you reach the open-ended questions, you can answer by speaking or typing.

One of those questions is direct: As a result of this course, what is one specific task or workflow you now feel better prepared to do?

If your answer is too brief, Public Voice asks for a short follow-up for a real example. If your first answer already includes a concrete example, no follow-up appears. At the end, you see a short summary of what you shared, in your own words, organized around the topics you covered. If something is wrong, you can edit it. If it looks right, you submit. The whole experience takes under two minutes.

Privacy was designed into the experience from the beginning. Public Voice does not ask for personal information, does not store audio, and is designed to detect and remove identifying details before responses are saved. The InnovateUS team reviews responses at the cohort level, not on an individual learner profile basis. That design choice matters because public professionals may be describing real work, internal processes, or barriers they faced. Honest reflection depends on trust.

Why Voice

Voice often changes the kind of answer people give. For Public Voice, that difference matters because we are trying to understand how learners may apply what they learned.

A 2025 comparative analysis by Glaut, a company that builds AI-moderated interview tools, tested the same open-ended questions with 252 participants across voice, text, and hybrid groups. Voice answers were longer and surfaced more themes than typed answers. Because the analysis comes from a vendor, we treat it as directional rather than definitive.

 Academic research points in a similar direction: a 2024 IZA discussion paper by Galasso, Nannicini, and Nozza found that oral responses to open-ended survey questions were longer and contained more personal experience than written responses. Peer-reviewed survey research by Höhne, Gavras, and Claassen also found that voice responses to open-ended questions in a smartphone survey contained more words and topics than typed responses.

When people are invited to speak, they often have more room to explain context, sequence, and experience.

That evidence fits what qualitative researchers have long known: when people are invited to speak, they often have more room to explain context, sequence, and experience. Public Voice uses that insight carefully, as a complement to interviews and traditional surveys, testing whether a short spoken reflection can help learners share more useful evidence of how training shows up in their work.

What We Expect to Learn

In May 2026, we launched an initial two-week pilot of Public Voice (to be repeated in June) with learners taking the InnovateUS' Responsible AI for Public Professionals course, comparing voice-based feedback with the existing typed survey. We are looking for the difference between a response that summarizes a feeling and a response that describes a behavior.

A typed answer of "I use it for drafting" gives a program team little to act on. We anticipated that a voice answer to the same question, with a follow-up if needed, might come back like this:

"I used it to draft a public notice for a road closure last week. I gave it the policy language, asked for a plain-English version, and edited it before publishing. It saved me about an hour. I have done it three more times since the course."

That answer contains a specific task, a specific workflow, an estimate of time saved, and evidence of behavior change that lasted past the day the course ended. A program team can take that answer and revise the curriculum. A funder can read it and see what their investment produced. A state partner can take it back to their staff as a real example of what the training led to.

If the pilot confirms existing research, Public Voice will become a standard part of the Observatory's approach to measuring the impact of every InnovateUS course.

Skill training is not peripheral to how the government performs. It is central to creating institutions that are more effective, responsive, and deserving of public trust. 

The goal of Public Voice is to help InnovateUS move from short reactions to concrete evidence of use. If learners can quickly explain what they tried, what changed, and where they still need support, then course feedback becomes more than satisfaction data. It becomes a way to understand whether AI training is helping public servants do their work more effectively.

That is worth more than six words.

Read more about the Observatory: https://rebootdemocracy.ai/blog/launching-public-sector-ai-observatory