AI and the Future of Governance

Artificial intelligence or “AI,” is the topic of much debate as policymakers and consumers grapple with how and when to regulate the increasingly used technologies. If recent hearings with top tech companies in the United States are any indication, it’s clear that policymakers need more actionable information to fully understand how AI works and how it could be applied in government.

This paper explores AI and its current applications in familiar consumer products. Further, we evaluate how AI is currently applied in government and posit how data and AI can inform a more metric-driven approach to creating and managing public policy.

Alphabet Soup: AI, ML, and NLP

First, it is helpful to define the key terms we’ll refer to in this paper:

• AI: Artificial intelligence refers to the broad field of making machines and systems intelligent.
• ML: Machine learning refers to the more specific objective of making machines and systems that can learn from experience. Virtual personal assistants like Siri and Alexa are examples of machine learning. These tools use your requests to learn more about you, how you speak, and what information you are interested in so that they can better understand and service your requests.
• NLP: Natural language processing refers to the goal of making machines and systems understand our human languages. Spam filters in email inboxes are a common use of NLP. The filters scan text for known words, phrases, and language patterns to detect established fraud patterns and protect you against spam. When you mark an email as “spam,” you are training the filter and strengthening its accuracy.

All three terms encompass techniques that enable machines to complete complex tasks that would previously require human intelligence — or might even be intractable for a human to complete.

Common Consumer-Facing AI Applications

Whether or not we realize it, artificial intelligence powers many of the consumer-focused technology platforms we use every day. Here are just a few of the ways AI is used in our daily lives and examples of how policymakers have tried to regulate these technologies.

Map apps: Products like Google Maps and Waze use anonymized, user-reported data on speed and traffic incidents to make predictions and serve recommendations for navigation. These apps have transformed how we travel and help us more accurately predict travel time, avoid accidents, and get to where we need to be most efficiently.

• Regulation: These applications have been impacted by numerous government rulemakings. For example, the National Highway Traffic Safety Administration (NHTSA) has issued rules that determine how long individuals may interact within technology while operating a motor vehicle (Time). These rules have made the development and monitoring of these technologies difficult.

Ridesharing apps: Services like Uber and Lyft use ML to optimize pricing, routing, and passenger queues. Uber’s former Head of Machine Learning, Danny Lange, confirmed that ML drives their ETA services, meal delivery estimates, fraud detection, and optimal pickup locations (Geekwire). This means that drivers are saving time and gas as they are synced with passengers nearby. The apps are always getting better. If you’ve ever canceled a ride on Lyft or Uber, you’re prompted to answer why you canceled (Driver too far, I changed my mind, etc). This helps the ML algorithm improve its estimates for future passengers.

• Regulation: Ridesharing applications have been the subject of increased scrutiny and regulation as the technology has spread. The federal government, international governing bodies, state governments, and even some local governments have weighed in on “transportation network companies,” (TNCs). For example, Austin’s government infamously banned Uber from operations (Engadget) and Texas’s state legislature offered a bill that overruled this local effort to curtail ridesharing services.

Airline autopilot: Airline technologies apply AI to manage the predictable, repeated stages of flight navigation. The New York Times reports that some Boeing planes allow for an average of only seven minutes of manual flight.

• Regulation: As might be expected, airline autopilot systems are subject to heavy regulation. Paul Robinson, president and CEO of AeroTech Research, summarizes the guidance given to pilots: “Let the computer do it because it can do a better job than a person,” (CNBC). Still, rules mandate what actions autopilot can manage and how many pilots should be in the cockpit during the operation of a flight.

Banking mobile check deposit: Banking applications use AI to power optical character recognition (OCR) technology that scans images and deciphers the written content for deposits. Instead of having to go to a bank’s physical location to deposit your check, you can simply scan the piece of paper on your phone and go about your day.

• Regulation: The Federal Reserve maintains administrative rules and guidance related to mobile deposits for banks, requiring the documentation of mobile deposits in a similar manner to physical deposits (Kiplinger).

Online search: Search platforms like Google, Microsoft, and Amazon leverage AI to learn from previous search patterns and provide more relevant results over time. The systems are trained to automatically understand relationships in relevance criteria and provide accurate recommendations.

• Regulation: Online search engines have perhaps witnessed the most regulatory scrutiny among the regulated technologies noted here. These regulations have captured the attention of regulators and the courts alike since at least 2000: Roberts Co. v. GoTo.com (2000) and Playboy Enterprises, Inc. v. Netscape Communications Corp (1999), (Yale Journal of Law and Technology). These cases deal with everything from copyright issues swept up in this new wave of innovation all the way to how “keying” works within the search engine itself.

Applying AI for Smarter Policy

As evidenced above, the private sector has rapidly innovated to provide consumers with simpler, faster ways of accomplishing common tasks using AI. Governments, however, have been slower to adopt AI despite many obvious applications across different government functions.

AI applications thrive on data and feedback. An algorithm on its own can only accomplish so much. Training data and access to live environment data feeds algorithms and allows them to learn and evolve over time. While consumer tech companies often scrape and piece together datasets to make their algorithms work, governments host and maintain millions of data points across every industry in our society. This data can help our policymakers create more effective policies in the public interest when the data is used efficiently and safely.

Consider the traditional policy model: legislators pass laws that spark the creation of regulation enforced by the executive branch. Lobbyists, think tanks, and special interest groups then track and advocate for policy changes in their interest areas based on the data they collect and the needs of their audiences. Policymakers rely on the information from special interests to guide new policy changes, and thus the cycle continues.

We envision a world where policies have specific key performance indicators (KPIs) attached to them, and these KPIs are tracked alongside the policy over time to indicate whether or not the policy is performing as expected. This strategy gives policymakers more insight into how policies might need to be adjusted without relying solely on biased data from special interests.

Further, this approach allows for a more data-driven approach to testing different policies across jurisdictions. States have long been touted as policy laboratories, but there is little standardization or systematic analysis of how different policy approaches perform against expectations. Standardized metrics can help benchmark policy performance across governments.

There are countless opportunities to apply AI and data analysis to the policymaking process. In the following table, we offer examples for how specific data could be tracked against policy goals.

This list demonstrates only a small portion of the possible future for government powered by AI. However, more data-driven policy is only possible to the extent that data allows. Governments are hosts to significant amounts of data, but the tracking, organization, and analysis of this data is often for wanting.

Crowd-Sourcing Data for Better Policy

Rather than relying solely on the government to collect and track policy outcomes, we propose crowdsourcing data to analyze policy effectiveness. By encouraging businesses and citizens on the front lines to share cost-benefit data and specific metrics on policy impacts, we can begin to close the feedback loop between policymakers and regulated entities..

An example of how this could practically be implemented is as follows:

• Policymakers pass laws and regulations with a goal and specific KPIs to track. Policymakers agree that they will revisit the policy on a regular schedule to evaluate the policy’s performance based on agreed upon KPIs.
• Regulated entities provide cost data and other data points specific to the KPIs on an ongoing basis to track policy effectiveness on a time series.
• Data processing techniques analyze crowdsourced data to identify average impact, outliers, and expected outcomes.
• Policymakers and the public are provided with reports on policy effectiveness and have the opportunity to validate or invalidate the analysis.
• Policymakers consider the data and public input when determining whether or not to modify the policy.

While we do not claim the above outline to be a standard to follow, we hope to facilitate dialogue on how we can shift processes to a more data-forward approach. We identify the following challenges for its implementation and welcome suggestions on how to solve:

It’s important to note that while AI will be instrumental in improving policy outcomes, it is still secondary to the people serving the public in those institutions. Data analysis and systems only help multiply the efficiency and abilities of our leaders — they do not nor should they seek to replace them and their judgment.

Effective governments will adopt AI powered technologies to identify, respond to, and track their success and failure in key policy areas. These systems promise to improve the ambiguity and skepticism surrounding government performance and services. More importantly, these technologies allow for more data-driven decisionmaking that cuts through politics and gets to the heart of policy outcomes..

Special thanks to Esper’s CTO Lilli Oetting for advising on the technology discussed in this paper. Visit esper.com for more information.