A decade ago the World Bank conducted a study. It found that nearly a third of their reports had never been downloaded, not even once. Another 40% had been downloaded fewer than 100 times. Similar studies are floating around about the paltry lack of readers for the 1.8-2 million academic journal articles published each year. While the numbers vary, it’s clear that too much knowledge goes unused. We are living in a world of information overload. But artificial intelligence is helping us overcome that challenge so we can get smarter faster about how to solve the world’s hardest problems. We are testing this proposition by building the Reboot Democracy chatbot. Try it for yourself at the lower right of the screen!

In a 2014 article about the World Bank’s findings, the Washington Post asked the provocative question: “What if someone had already figured out the answers to the world's most pressing policy problems, but those solutions were buried deep in a PDF, somewhere nobody will ever read them?”

With generative AI, we can ingest large amounts of information, including neglected PDFs and then have the AI summarize and extract knowledge. For example, it is estimated that Claude, Anthropic’s generative AI tool, can process the equivalent of about 75,000 words in a minute, or the length of an average novel, meaning that generative AI can make short shrift of all those studies, PDFs, reports, policies and articles by “reading” the information. We can even ask questions about the information. That kind of conversational interaction with data is called a chatbot.

A chatbot gives you the opportunity to effectively search through such a big data set, and also to make sense of the data in a way that's accessible to regular people. 
Róbert Bjarnason, Citizen Foundation

As Róbert Bjarnason at Citizens Foundation explained in a recent interview with me about the work the GovLab is doing with Citizens Foundation to create a Reboot Democracy chatbot based on our research, including posts from this blog: “A chatbot gives you the opportunity to effectively search through such a big data set, and also to make sense of the data in a way that's accessible to regular people.” Stephan Schmidt at the GovLab, who is working with Robert, says, “the added value of a chatbot is that it works with your raw universe of data but combines it with the power of ChatGPT to offer something much more specific for a given field. It's kind of like a search engine with the added value of AI technologies.”

Robert and Stephan are building a toolkit designed specifically for making sense of large quantities of documents. The problem with simply searching for information on ChatGPT is articles on a topic like governance or democracy are not in the training data. Also, you often want to be able to query specific documents, such as the policies of a single government agency to help educate and onboard new hires. That’s one reason to build a custom chatbot as my students are doing for the Massachusetts Department of Transportation.

To be sure, anyone (with a Gmail address) can upload a set of documents to a free tool like Google’s NotebookLM or (for a fee) do the same with OpenAI’s GPT Builder, which will then answer questions based on your documents.

But using these one-size-fits-all tools comes with its own problems. “ChatGPT,” says Stephan, tries to catch the whole world. But the data that funnels into it is opaque to us. By contrast, the data in the chatbot is much more transparent. Also, ChatGPT takes a longer time to contemplate and to reply.”

Hence Robert and Stephan have built what they call “an ingestion pipeline,” comprising 30 different instructions to ChatGPT (known as prompts) that are helping with the organization of information to ensure that retrieval is both efficient and relevant. Their approach involves compressing information to speed up searching while giving the AI directions to help ensure that output is accurate. The tools they are building split documents into small chunks so that instead of simply retrieving a whole PDF, the chatbot returns the right paragraph with a citation and source.

If you click on the icon at the lower right you can ask a question: How will AI impact democracy? What AI tools can I use to engage the public? How can AI help governments work better?

While we are using this chatbot to enable anyone to ask questions about the impact of AI on democracy and on governance, Citizens Foundation and TheGovLab are trialing this approach to create a reusable toolkit for any context that involves synthesizing large quantities of information.

Every branch of government at every level possesses tremendous knowledge about the economy, society, science, the environment, public health, and the populace. There are public documents—from reports and regulations to strategic plans and studies—then there is a treasure trove of internal documents, including memos, letters, and more. 

Imagine an incoming politician who wants to know all the actions the last administration took on a particular issue. Or an eighth-year incumbent who wants to know the position she expressed back in her first year of office. Imagine a public servant who wants to search all of his agency’s policy documents to get an answer to a question. Imagine a member of the public who wants to teach themselves everything about a policy area. The possibilities are endless and extraordinary. Such a chatbot can assist policymakers, the public, and researchers in accessing a wide array of existing research, legislative and regulatory documents, case studies, and reports, enabling us to craft policies that are informed by a comprehensive understanding of historical context and current best practices.

So play around with the Reboot chatbot. Tell us how it works, what we can improve, content we should add, and what you think!