Using AI during a participatory process, part 2
/Using AI during a participatory process, part 2
Transcribe speech
Example platforms with this feature: Fora, Loomio, CartoDEBAT, Cortico, The People Say, Dembrane ECHO. Plus mainstream platforms like YouTube and Zoom, for some languages.
Generative AI isn't limited to working with text. It's also pretty adept at working with audio and video. AI is now able to capture speech as text (and even convert it back to speech again). The more that participation platforms allow people to speak as they naturally would, as if to another human, the more accessible participatory processes can become.
Literacy levels vary widely, for many underlying reasons, and typing out text paragraphs on a screen just isn't a very accessible activity for many people. AI is already being used to capture and transcribe people's spoken voices onto digital participation platforms. By lowering barriers to participation in this way, AI could help broaden participation beyond those comfortable using tech. It can also free up organizers to more meaningfully engage in the program. For example, in Germany’s Siegen-Wittgenstein district, planners piloted DeliberAIde during their participatory social-planning conferences. The AI helped streamline note-taking, improved documentation, and allowed facilitators to focus fully on engaging with participants in dialogue.
Once the speech is transcribed into text, digital platforms can do many things with it. An obvious application is search. But storing interview recordings as text also allows participation platforms to include that participant input in topic clustering, translation, auto-tagging, alerts, and recommendation engines, among other features. Platforms like The People Say can also provide deep links to specific timestamps in a video or audio interview, for example, allowing others to jump right to the relevant part of an hour-long discussion.
Another example worth calling out here is Cortico. The team has developed an entire user interface for transcribing speech and surfacing patterns in the recordings, while retaining the specific audio segments to use later. The project demonstrates that the power of the human voice doesn't need to be lost while collecting large amounts of interview data.
Let people review and approve transcripts of their speech
Example platforms with this feature: CartoDEBAT, Converlens, deliberAIde
Although rapidly improving, speech transcription is still not yet 100% accurate. Certain languages, accents, and specialized terminology can produce errors in the transcripts. For this reason, it's important that platforms offer participants the opportunity to review and if necessary, correct the transcripts of their speech transcripts.
Public meeting transcription
Example platforms with this feature: OpenCouncil, citymeetings.nyc, Block Party
An exciting use case for AI speech transcription is public meetings. For years, civic innovators have wanted to enable alerts for meetings. Imagine getting a proactive notification when your child's school is being discussed in a community meeting. Rather than attend meetings for hours each week, you can engage when something relevant to you has been discussed.
A variety of civic platforms are developing AI speech transcription tools to make public meetings more accessible in this way. They're navigating the need to train the AI on special, locally-relevant vocabulary and acronyms, and exploring ways to engage residents in the outputs of the meetings. deliberAIde, for example, is using AI to detect and delete Personal Identifiable Information (PII) in transcripts.
Scanning handwriting and other analog documents
Example platforms with this feature: Go Vocal
Optical Character Recognition (OCR) is an important technology for integrating and digitizing offline participation channels like handwritten input, or paper print-outs. As AI improves, OCR is also getting better, with fewer transcription errors. Like with speech transcription, it can still struggle with niche language or symbols.
AI chat
Example platforms with this feature: Polco, Konveio, Your Priorities
ChatGPT popularized the general human-to-AI chat interface for AI interactions. Between the familiarity of the interface and the ease of replicating it technically with APIs, many products adopted a similar chat approach when adding AI. This includes some digital participation platforms, which have already integrated chat modes in some of their products.
An example might be taking a long government document, like a law, and adding a chat interface where people can "chat with" the document in natural language terms. This allows people to ask for just the answer to the question they care about, for example, rather than read 50 pages of dense legalese. On participation platforms, this interface can be extended to "chatting with" the results of a participatory process, so they can make sense of what took place and how the group came to the final decision.
Tech companies are already exploring other interfaces for interacting with LLMs, so we'll see more ways of interacting with LLMs in the future.
A downside can be that a chat interface expects people to come up with questions to ask the LLM or dataset, for example, rather than proactively guiding them. Which leads us to…
Assist participants through the process with bots
Example platforms with this feature: deliberAIde, Polco, Empurrando Juntas, All Our Ideas, Konveio, Your Priorities, Loomio
Another form of chatbot is explicitly framed as a user-support agent. Due to the cost of employing humans to provide customer or constituent services, this has been an active area of software development for many years. The chatbot automates or semi-automates helping users get answers to their questions, sometimes routing the request to a human support agent if the chatbot fails to handle the initial query.
Digital participation platform developers have also customized AI support bots to assist people through a participatory process. The user support bots are provided with contextual background on the process, and tasked with supporting people as they navigate it in order to help them participate. They usually appear in the composition box of a participation platform, where participants are tasked with contributing ideas or drafting a proposal. For example, All Our Ideas generates brief, helpful Pros and Cons to explain the options participants are voting on.
Some bots are designed to accompany people throughout the entire participation process. The bots can proactively encourage and work together with people to brainstorm ideas, expand on their draft proposals, and reflect on what they want to say to the group. The result can be that users' proposals contain more supporting material, enhancing its legibility to other participants.
A risk here is that different participants might have varying experiences with the same support bot, leading to unequal treatment. For example, a bot might provide two different people a different set of pros and cons about the same proposal. And due to the (valid) competing priority of keeping these conversations private, the process hosts might never know how users' experiences with a support bot might differ.
There is also a risk that the user will submit whatever the bot suggests, since it's significantly easier than composing an original proposal, distorting the human inputs into the participatory process. AI is already shaping our language, and people are short on time. There's increasing potential that participants might submit something that the AI contributed without sufficient review (see our discussion of 'AI slop'). As one People Powered member put it, "If you shape people's input with AI, you miss the point of engaging people."
Allow participants to generate text
Example platforms with this feature: Assembl, All Our Ideas, IdeaScale
One of the key ways user-support bots help people participate is by helping them improve their participatory proposals with AI draft assistance. This can mean expanding upon their initial thoughts, clarifying initial ideas, and generally helping people make their contributions easier for others to understand.
All of the existing concerns about text generation apply here. So far, the user is always the one who decides to submit their communications. So even if AI helped out, the participant hypothetically signs off on it by submitting. But it's not hard to see how an AI model's inputs into a participant's proposal could also shape it.