OpenAI claims to have solved the Navier-Stokes Millennium Prize math problem. In July, its AI agents broke containment and hacked competitor HuggingFace without human consent or knowledge. Anthropic CEO Dario Amodei wrote, and OpenAI CEO Sam Altman agreed, that AI companies must “pace the frontier” to ensure that increasingly intelligent AIs are not used to cause catastrophes.
Clearly, AI is a powerful tool – and governments have the opportunity to use AI to address longstanding problems. But they must also maintain commitments to safety, fairness, and accountability to the public. Achieving both goals means increasing adoption of AI while monitoring its use and ensuring that people remain accountable for government actions and decisions.
In the near future, AI will help deliver social safety net programs to meet people’s basic needs, set up pathways for people to get well-paying and fulfilling jobs, and even build homes and infrastructure. As a general purpose tool, it will speed up day-to-day work that involves analyzing, reasoning about, and communicating information in operations, management, and rulemaking and legislation.
The keys to making AI work for humans, and not for corporations, are the same government oversight measures that exist already, streamlined and enhanced to ensure oversight over AI as well.
Here’s how this would work.
Efficient operations
A standard application for SNAP must be processed by a state agency within 30 days. In practice, the average state processes only around 85% of applications on time, meaning millions of people spend weeks or months waiting for money to buy food. AI tools can help caseworkers more quickly answer questions from applicants and scan submitted documents for common issues; similar use of AI in the private sector allowed customer service staff to process requests up to 34% more quickly. This can speed up case processing times by shortening the time it takes to complete routine steps and letting applicants know more quickly if they need to provide more information to the agency.
Agencies can also implement oversight measures to ensure continued accuracy. They can randomly sample some tasks to be completed only by humans to ensure that automated systems behave in ways equivalent to people. They can have the systems flag complex cases and inconclusive results for extra review. With efficiency gains from AI, human caseworkers can spend more time listening to applicants to ensure that a person’s real-life situation is properly represented by a technical system.
Agencies are ultimately accountable to external oversight; for SNAP, these include fair hearing protections where applicants can dispute the results they received and independent quality control assessments by state and federal governments. Agencies routinely make operational changes unrelated to AI; these oversight measures ensure that any operational change doesn’t come at the expense of accuracy.
Data-driven management decisions
What can governments do to set people up for financial success? The evidence for what works is limited, and the effectiveness of any employment program will vary from person to person and across communities. Researchers have used data on millions of people in recent years to shed new insight on what has driven change across the country. But state and local government agencies tasked with improving employment want to know not only what works overall but what can work for them.
AI can help program administrators make more informed decisions, by analyzing local data (including novel data sources like social media) and searching for comparison cases to see what worked (and didn’t) in cities and states like theirs across the country. Administrators can treat these like scaled-up case studies, providing them with ideas for measures to implement and evidence for which interventions have been particularly robust.
Program administrators rarely defer judgment to case studies, though, and should not defer judgment to AI either. They can continue to draw on their practical expertise to sanity check AI analyses and make changes to program rules and funding allocations based on what they know to be feasible.
Laws and rules that actually work
Builders in New York City must navigate 40 approval processes to get something built. Other construction and infrastructure projects like for renewable energy also face onerous permitting processes. Local, state, and federal governments around the country are pursuing permitting reforms so that homes can be built where people want to live and clean energy can be generated to meet growing demand and transition away from fossil fuels.
The high volume of laws and regulations, especially across different levels of government, mean that even listing out all permitting rules that builders have to comply with is a challenge. AI tools can help gather that information so that legislators can ensure reform is comprehensive.
Legislators are often pressed for time and drawn between competing priorities. Even before AI, they would often rely on outside groups to provide the text for legislation, and often weren’t given enough time to fully read the bills they voted into law. AI has the potential to magnify these problems; it is already common for legislators to use AI to draft laws. Legislators must ensure that the bills they draft and vote on receive as much human review and scrutiny as laws created before AI.
Government must rise to the occasion
Unlike with math, government is built around human judgment and accountability. Success in government doesn’t always have deterministic, provable outcomes. Deciding how to harness AI while maintaining safety and accuracy standards is a major open problem, and even top tech companies are figuring out how to get it right. The responsibilities, accountability, and governance of a public agency are different from a private company, so the best practices for how to use AI in government will also differ in important ways.
It is therefore imperative that governments take the lead in responsible AI deployment instead of letting the private sector decide for them.
A generation ago, computing technologies like email, spreadsheets, and the internet fundamentally changed how office work was done. Employees’ responsibilities changed, and government agencies had to deliver on more because the public expected more. The same thing is happening with AI today.
We’ve invented robots that can perform mathematical miracles. Governments can harness these robots responsibly to ensure that people can put food on the table, find well-paying and fulfilling jobs, and have homes to live in and electricity to power them. A government that fully delivers on what the public wants is its own miracle.
Are you a government practitioner who wants to use AI and technology to do your job more effectively and efficiently? A technologist who wants to dip your toe into the civic realm? An expert in government who thought our article was spot-on, completely off-base, or somewhere in the middle? Tell us more by filling out this short form!