School of Planning, Design and Construction National Charrette Institute

How can planners use AI to support their public engagement work? The ideas in this article come from a recent discussion of the Public Participation Community of Practice group hosted by:
Jennifer Hurley.

President & CEO, Hurley Franks & Associates

Visiting Assistant Professor, Bryn Mawr College

Institute for Social Innovation Fellow, Fielding Graduate University

John Stephens, Associate Professor of public administration and government at the University of North Carolina at Chapel Hill, seeded the discussion with examples and questions. Some ideas below describe ways that planners are already using AI, while others envision possible uses that may be just around the corner. Each idea includes a sample prompt or AI tool to use. Next month’s piece ends by discussing some cautions and concerns.

MEETING SUPPORT USES

Another use of AI that is already possible is various forms of meeting support. The tools are less well-developed than the text tools and may be harder to use, especially in a public setting, but they show interesting promise. Part 1 covered the functionalities of; converting handwriting to text, developing slides from text, and providing ideas for meeting activities or a draft agenda to meet meeting goals.

A meeting support use mentioned but not expounded upon in Part 1 was AI's capability to generate images from community input.

Generate images from community input. AI image generators have come a long way. Even when they create somewhat odd outcomes, the results can help stimulate discussion. Planners and architects are trained to read plans and imagine built space based on description, but many members of the public have trouble doing either. Generating images based on what public participants say they want can help refine their vision. The series of prompts and results below the tool examples demonstrates some of the possibilities, as well as some of the serious problems, with these tools:

TOOLS:

Chat GPT Image Generator - Works through Chat GPT prompt window, but requires Pro subscription.

Midjourney - Considered one of the better image generators, but works through a Discord server, so takes some setting up.

Craiyon - Free site, supported by ads.

CraiBing Image Generatoron - Also uses Dall-E. Does not require a subscription, but generates images much slower than Chat GPT.

Add a touch of Dutch to your street - Takes a street view image for a particular address and “reimagines” it to be more supportive for cycling.

EXAMPLES FROM DALL-E, BY OPEN AI (PART OF CHAT GPT)

PROMPT: Create 3 images that show options for a neighborhood center that has walkable local retail.

Note that this result has two images, not three.

PROMPT: Adjust these images to have more traditional architecture and lower density.

Note that this result has added weird text artefacts, including some that are inaccurate (3 instead of 2) or illegible.

PROMPT: That architecture looks southwestern. Make it look more like New England.

Now this result seems like a pretty good response to the prompt, except that it is a completely different image and not just an adjustment to the architecture style on the previous image.

PROMPT: That image is good. Make it look more like watercolor markers and less like a cartoon.

This result definitely looks more like a sketch and less like a cartoon than the previous image, but it’s a completely different image.

PROMPT: Make sure some of the people in the image are people of color.

Again, Chat GPT has changed the image entirely rather than altering the previous image. It has also added text artefacts. Since the people are so small in both this image and the previous one, it’s hard to discern racial features, but the one obvious difference in the people in the two images is that the people in the image that was directed to “include people of color” are dressed much more casually than the previous image. Hmm, maybe it’s random, but that seems like bias to me.

EXAMPLES FROM DUTCH CYCLING LIFESTYLE

INPUT PHOTO FROM A STREET ADDRESS:

RESULTS FROM DUTCH CYCLING LIFESTYLE:

NEAR FUTURE POSSIBILITIES

It’s also possible to imagine some uses of AI that may be possible in the near future, such as:

• Analyzing survey input
• Developing avatars to represent missing points of view

Analyzing survey input. The use of community surveys as part of public engagement has exploded with the ease and cost effectiveness of using online surveys. AI can create summaries of survey data. At the moment, these tools all require expensive subscriptions or are difficult to use, but here are a few to watch for user improvements:

TOOLS: Tableau | Qualtrics | Data Robot | Rapidminer

Developing avatars to represent different points of view. We know that the vast majority of public engagement efforts reach a particular slice of the population that does not represent the whole.

Counter-balancing the skew in participation takes time and resources than many planning processes do not have. AI companies are working to develop independent agents that can act on your behalf (say, to plan a trip itinerary and reserve and purchase flights and other reservations). At some point in the not-too-distant future it is likely that an AI avatar could represent the point of view for a particular demographic. To see a glimmer of this possibility, see how Claude responded to the following prompts.

You can try the prompts in other models, such as Google Gemini or Bing. This approach could be helpful for generating the criteria to consider in evaluating options. Even if a particular demographic is not represented at a public meeting, at least concerns they would be likely to have could be included in the evaluation matrix.

CAUTIONS AND CONCERNS

Although AI has intriguing possibilities for improving public engagement, there are several cautions we should consider, including:

• Hallucinations and errors
• Creating records subject to public information laws
• User agreement issues
• Reinforcing stereotypes
• Bots in the discussion
• Disconnect from decision making process

Hallucinations and errors. Current AI is based on large language models (LLMs) that create very sophisticated text prediction, but it does not actually understand what it is saying, so it is subject to making things up, often in hilarious but sometimes alarming ways. For example, AI chatbots gave wrong answers to questions about the 2024 election 27% of the time. Anything created by AI must be reviewed carefully by a real human who knows the subject area. Even transcripts created by voice recognition have many errors; public agencies must be careful about how they present these files as representations of the meeting content.

Creating records subject to public information laws. Putting City Council meetings on local cable access television seemed like a clear win for good governance – increasing transparency and reducing corruption. But it’s obvious that it also had unintended consequences: in some cases, Council members grandstand for an audience at home, limiting the ability of the meeting to engage in real, constructive debate and compromise. Using tools to create an auto-generated text transcription of an internal planning meeting sounds like a good idea for everyone’s recall of what happened, but in many places, it creates a file subject to public records laws. And knowing that everything they say may become public through a records request is likely to change the content of people’s speech, making it harder for people to be candid and negotiate difficult conflicts.

User agreement issues. Many AI services have language embedded in their user agreements that say they reserve the right to save and use uploaded material for training and/or marketing purposes. That may be fine for more generic tasks, but when we are working with data that is specific to a community and is not yet public, we need to avoid those services. Unfortunately, most services make their user agreements hard to find and understand.

Reinforcing stereotypes. It’s clearly documented that AI has racial and gender biases. Any content generated by AI may display those biases. Although the idea of avatars as a possible tool for bringing more representative diversity into public debate is intriguing, it will be limited by the biases within the models. And research shows that, “People may learn from and replicate the skewed perspective of an artificial intelligence algorithm, and they carry this bias beyond their interactions with the AI.”

Bots in the discussion. Online discussions of local political issues and elections are vulnerable to non-local outside agitators. As avatars become more sophisticated and cheaper to generate, it will be harder and harder to tell if people participating in written or virtual public engagement venues are real people or just bots. We may become overwhelmed by the sheer volume of bot spam in public input. Bots may be developed purposely to skew a public debate, likely to benefit a particular segment with more resources.

Disconnect from the decision making process. We should never forget that public engagement is only effective when it is connected to the decision making process. Planners wield influence with their analysis and recommendations, but most consequential planning-related decisions are made by elected or appointed officials. Perhaps AI avatars could help engagement processes consider a more representative set of opinions and perspectives, but if elected officials only care about the constituents who call their office, they may ignore the AI input.

WRAPPING UP

Our discussion group was just 10 people, so take the following with a grain of salt, since we clearly don’t represent all planners. In our group, everyone had played around with a few AI tools and uses them on occasion, but no one yet found them to be a core part of their workflow. In some ways, AI seems to be a solution in search of a problem. Many in our group also had ethical qualms about using AI because of the method of its creation: current models were trained on material scraped from the internet without content creators’ consent or compensation. On the other hand, most in the group had a general sense that this technology is growing rapidly, and planners need to figure out how to harness it in a positive way.