A Cheat Sheet for AI in Government

Governments across the country are thinking about ways to use artificial intelligence (AI). Some state agencies are using chatbots to simplify the customer experience when people accessing a service need direction to basic information. Others are using it to sort applications so that caseworkers can identify who can easily be approved for a benefit and whose cases need more personal attention. Use case examples are proliferating, and just like any new tool in government, there’s a learning curve. 

In order to responsibly decide when and where to adopt AI, governments should be familiar with the terminology around the technology—so here’s our cheat sheet with key terms you should know.

So first, what is artificial intelligence? Put simply, AI is a computing system designed to perform tasks that might normally take human intelligence—things like speech recognition, language translation, or scenario planning. AI systems can vary in their complexity, and examples are all around us. It’s how services like Gmail recognize patterns in your writing and suggest the next words to type, and how Spotify learns the kinds of music you listen to and creates personalized playlists.

Expert systems, one of the earliest forms of AI, are built around a predetermined set of rules.  Governments often use these systems, usually known as eligibility rules engines, to automate the determination of whether or not someone is eligible for a benefits program. Modernizing these systems with more advanced machine learning technology could help reduce error rates and streamline the benefits delivery process.

Machine learning is an application of AI in which machines “learn” without being specifically programmed to do so. In an example metaphor, traditional programming is compared to baking, where a recipe gives ingredients and specific instructions for the baker to follow; similarly, programmers give computers detailed instructions to follow. With machine learning, computers decide what to do next based on what they’ve learned after being trained with a data set. Engineers build a learning model, input the data, and let the model train itself by drawing patterns or making predictions based on the data.

Machine learning is used in a lot of different contexts. One way that governments could use it in the future is to help automate renewals for programs where someone’s eligibility is unlikely to change. For example, Medicaid clients usually need to renew their benefits each year—but for clients who are disabled or elderly, their eligibility is unlikely to change. Machine learning algorithms could analyze applicants’ past data, determine whether they still meet eligibility requirements, and surface that information to caseworkers for them to make an assessment. The system could then automate chunks of the renewal process, notifying recipients of what they need to submit and when, and pre-filling some fields—reducing the administrative burden for clients and caseworkers alike.

Another type of machine learning we work with regularly is called entity resolution—a fancy phrase for deduplicating records in a system. Government often works with duplicative and messy legacy data, where multiple records for the same person exist across multiple systems. Entity resolution technology helps identify and link those records. You may have also heard this referred to as master data management.

AI IN ACTION

Our criminal justice team used entity resolution to help Utah implement its Clean Slate law. The technology solved a key challenge: how to match duplicate defendants in the court’s case management system to provide a more complete view of a person’s criminal history. We integrated this machine learning technology into an eligibility algorithm that resulted in the clearance of over 500,000 criminal records.

One subset of machine learning is deep learning, which uses large neural networks, a computer’s way of emulating the structure of the human brain, to analyze large datasets and learn patterns. Some example use cases of deep learning include speech recognition, language translation, sentiment analysis, and document summarization software.

Perhaps the best-known type of deep learning is generative AI: a form of the technology that can respond to prompts and create new content such as text or images. One form of generative AI is a large language model (LLM). LLMs are programs that are fed massive amounts of language data so that they can produce coherent language output via prediction. This is good for tasks like answering a question—though they may generate inaccurate or biased language. When you ask a program like ChatGPT to summarize a piece of legislation at a fifth-grade reading level, what it produces is an example of generative AI. But a note of caution here: though these programs sound confident, they can make errors, and their interpretations and logical leaps might not be ones that a human would make. 

While this technology is evolving to use more reliable data sources, human-reinforced learning, and fact-checking mechanisms, we currently advise government against using chatbots for direct public-facing interactions. A more responsible approach would be to pilot the use of this technology internally as a form of augmented-intelligence to support civil servants in their daily workflow. 

AI IN ACTION

We sometimes use chatbots—but only in specific circumstances, and with a great deal of caution. Our GetCalFresh chatbot responds to simple questions people have about their applications to California’s SNAP program, but the responses people get are templates written by members of our staff—not information generated by AI. Instead, an AI model uses classification to identify messages that qualify for an automated response.

Governments might be more interested in other forms of automation that solve for monotonous tasks that would otherwise take people a long time. While not technically under the artificial intelligence umbrella, Robotic process automation (RPA) is a type of technology that can do repetitive tasks, such as data entry, filing in forms, and moving files. Some states use RPA to help process benefits applications—when clients upload their verification documents, an RPA tool can integrate the uploader with the state’s document management system and add these documents to their case file. 

Another useful implementation of AI for improving service delivery is optical character recognition (OCR) and handwritten text recognition (HTR), which refers to the process of converting an image or PDF into text. When handwritten forms are scanned into data management systems, OCR/HTR can effectively reduce the time it takes to process the information on them. Our criminal justice team has worked with governments across the country to implement OCR to determine if criminal records are eligible for expungement after a state passes a Clean Slate law. 

Though the world of AI may be filled with new terminology, it doesn’t have to be intimidating. The keys to experimenting with this technology are starting small, mitigating risk, and deploying it with a human-centered approach. When we follow those, we can find ways to leverage AI to make government work better for everyone.