Artificial intelligence is reshaping how governments operate, how workers do their jobs, and how citizens interact with public institutions. But while the technology is advancing quickly, the frameworks for using it effectively—and responsibly—are still taking shape. In this collection of perspectives, Harvard Kennedy School experts on policy, economics, education, and technology outline how governments can think about AI: as a tool for improving services, a source of new risks, and a force that will require careful oversight, new skills, and ongoing adaptation.
• Mark Fagan: On how governments should think about AI
• Jason Furman: On principles for AI regulation
• Videos from Teddy Svoronos, Mark Fagan, and Bruce Schneier
• The Project on Workforce: On tracking AI adoption
• Bruce Schneier: On AI and elections
• Teddy Svoronos: On teaching generative AI
Mark Fagan: On how governments should think about AI
People should be thinking about AI as another tool in the toolkit for delivering constituent services with quality, efficiency, and fairness. AI is a very powerful tool. It’s an evolving tool. But it is only a tool.
Every governmental organization—and for that matter, nonprofit—is faced with the current dilemma of doing more with less. Budgets have been reduced, and yet the demands of constituents or stakeholders are increasing, which puts a lot more pressure on those in government to be able to meet those expectations. AI can provide us with a more efficient way of delivering a fair number of constituent services.
Today, if you have questions about enrolling your child into school, you need to get on the phone with someone in the school system. Typically, the window for doing that is the business day. There may also be some language restrictions. An AI chatbot might be available 24/7. It might be able to operate in multiple languages and therefore provide much more reliable and efficient information. That’s a benefit that we can get from AI.
That’s the good side.
There are some concerns. Number one is data privacy. Governments have a lot of information, and one of the obligations of the government is to protect that information. When it is put into an AI algorithm, we need to make sure that, at the individual level, you can’t be identified. Protecting that information through anonymization, being very careful about what you use it for, being parsimonious about it—those are ways to address data privacy. A second real risk is hallucinations. The way in which a large language model works is it simply tries to predict the next word based on a wealth of information. You can get foolish information. The solution there is to verify, verify, and verify.
There are two other issues to worry about. One is bias. All models are trained on data. If that data is not representative of the underlying population, you’re likely to get biased information. This has been a problem that’s surfaced around hiring people. The last big risk is disinformation. Today, it is very easy to use AI to modify information, images, and videos and to share information that’s inaccurate.
Here, the role of the government is to monitor and to build trust with the community. You need AI expertise: technical expertise, people who know about machine learning, data science, and cybersecurity. You also need expertise in terms of AI ethics and regulations. What are the laws in the regulatory environments? What constitutes ethical decision making? Another type of expertise that’s important is operating under uncertainty.
The development of AI is happening real time. Being agile, learning, and adapting is really important.
People in your organization need to know what AI is all about. You can also leverage the AI community. There are a lot of people, communities, governmental organizations, and nonprofits who are working in this space. Tap their expertise.
Let’s talk about concrete guardrails. These consist of policies and ethical considerations. The European Union passed the EU AI Act. It’s a risk-based pyramid. It starts at the top with things it does not allow AI to do: cloning human beings, doing social tracking, and social scoring. Then there’s a set of things you can do, but you have to be transparent. And then, at the bottom, anything that doesn’t create risk, you’re free to do.
Ethical considerations tie into what data to use. That chatbot I mentioned in discussing school enrollment? Well, that’s fantastic if someone has access to technology to be able to access the chatbot. Many people don’t have that. Moreover, it isn’t enough to have the access. You need to have the literacy to be able to know how to interact with it. Guardrails and ethical considerations ensure quality, efficiency, and fairness.
Mark Fagan is a lecturer in public policy. His research focuses on regulation, and he is the co-author of a new book, “Governing With AI: How the Public Sector Can Use Artificial Intelligence to Improve Performance.” He also recently co-authored a risk assessment framework for AI.
—
Jason Furman: On principles for AI regulation
As policymakers consider regulating AI, they should keep five principles in mind:
- Consider both the benefits and the risks. It is tempting to apply the precautionary principle, requiring AI to prove that it is safe and eliminate all risks before it can move forward. But doing so would risk missing out on the enormous upsides AI has to offer in areas like scientific research, education, and the labor market more broadly. Delaying any of those benefits would be costly too. So, the right approach is the old-fashioned one of balancing the potential costs and risks of AI against the costs and risks of delaying its deployment.
- Compare AI with humans not the Almighty. AI is biased. AI gets into car accidents. AI can be overconfident and make stuff up. But guess what? Humans do all that too—and in many cases they do all of that even more and worse. Plus, AIs can be better at learning from their mistakes. So don’t ask for perfection, just ask for an improvement on us.
- Domain-specific regulation not an AI super regulator. AI will be used in everything from cars to medical devices to stock trading. We already have regulation for cars, medical devices, and stock trading. We shouldn’t have a super regulator for all uses of AI but instead beef up the AI capabilities of our existing regulators so they can judge not the technological input but the output—Is the car safe? Is the medical device effective?—and ask the questions they already ask today.
- Don’t let regulation be a way to protect incumbents. Sometimes the big AI companies welcome regulation for well-intentioned reasons. Other times I worry they think they can comply with the rules and smaller competitors cannot. We have benefited from vibrant competition in the AI space, we should not inhibit it—and help entrench monopoly power—with unnecessary regulation.
- Not every problem has a solution. We need to work hard on making sure AI does not help people make sexually abusive imagery or bioweapons. Those problems urgently need solutions. But I don’t see how AI regulation can ensure that the technology does not increase inequality or make jobs more meaningful. Instead, we will need to find the solutions to problems like that outside the technology, for example with more progressive taxes and transfers.
“AI is biased. AI gets into car accidents. AI can be overconfident and make stuff up. But guess what? Humans do all that too—and in many cases they do all of that even more and worse.”
These principles do not tell you the exact answer to any question, but they do help you think about how to answer them—and the answers themselves will be constantly evolving as the technology and our understanding of it evolve together as well.
Jason Furman is the Aetna Professor of the Practice of Economic Policy and was a top economic advisor in the Obama administration.
—
The Project on Workforce: On tracking AI adoption
Generative AI is now used by 55.9% of working-age adults in the U.S., according to the Project on Workforce’s latest nationally representative survey data from November 2025, available at the GenAI Adoption Tracker. Over 40% of workers now use generative AI at work.
To put this in perspective, we can compare the pace of generative AI adoption to that of other general-purpose technologies, relative to the release of the first mass-market product for each technology. November 2025 is roughly three years after the release of ChatGPT, the first mass-market generative AI product. The current generative AI adoption rate of 55.9% exceeds the internet’s 30.1% adoption rate in 1998, three years after the internet was opened to commercial traffic.
Where AI adoption is having the most meaningful workplace impact
In industries such as private management (71.5%), information (70%), finance and insurance (63%), and professional, scientific, and technical services (62%), worker adoption rates are remarkably high. But workers in traditionally blue-collar industries are also adopting it at significant levels: 42.9% in construction, 39.6% in manufacturing, 21.1% in transportation and warehousing.
What policymakers and employers should be paying attention to now
Workers report an increasing share of work hours saved due to generative AI use: 2% of total worked hours in November 2025, up from 1.7% in August 2025. When time savings estimates are fed into a standard aggregate production model, they suggest that generative AI may have increased labor productivity since the introduction of ChatGPT at the end of 2022.
At the same time, these productivity gains might not be distributed equally across the economy if generative AI is adopted unevenly across income and education levels, even within industries and occupations.
Notes on the tracker and its data
The Generative AI Adoption Tracker visualizes data from the first nationally representative U.S. survey of genAI usage at work and at home. Data come from the Real-Time Population Survey (RPS), a national online labor market survey of working-age adults aged 18-64 that has run since 2020.
RPS is designed and weighted to be nationally representative and to complement existing government surveys, such as the Current Population Survey or the American Community Survey, by carefully replicating core sections of those surveys while still leaving room for novel questions.
The current version of the tracker shows the combination of six survey waves, run quarterly from August 2024 through November 2025, including 25,000 respondents.
The Project on Workforce is an interdisciplinary, collaborative project between the Harvard Kennedy School's Mossavar-Rahmani Center for Business and Government, the Harvard Business School’s Managing the Future of Work Project, and the Harvard Graduate School of Education.
—
Bruce Schneier: On AI and elections
AI is poised to affect every aspect of the upcoming election. We’re going to see avatars, both authorized and unauthorized. Candidates are going to have authorized avatars and get them to speak to voters in ways that an email just can’t.
Candidates are going to use AIs for all aspects of running the election. There are already AI companies that help candidates file papers to run for office, gather signatures, get on ballots, put up websites. And here don’t think Congress—think local elections, where there’s no time, there’s no money, there’s no expertise, there’s no staff.
There are applications to help candidates connect with donors who can fund their campaigns. There are applications to help with door-to-door get-out-the-vote campaigns. All of this is going to be reenergized through AI.
What’s interesting is that I don’t think we are seeing yet a candidate who knows how to harness AI. If you remember back in 2008, Obama was the first social media president. He figured out how to use social media for campaigning in a way no one else did before. There’s going to be something similar for AI. Someone will figure out how to use that technology in new ways, for new types of politics, new types of electioneering, in a way we haven’t thought of before. No candidate is there yet.
“AI is poised to affect every aspect of the upcoming election. We’re going to see avatars, both authorized and unauthorized.”
We are seeing, at the edges, AI seeping into politics at the local level, and it is not yet partisan. Republicans and Democrats alike are organizing against big AI data centers in their communities. It will be interesting to watch whether that issue becomes partisan or remains a local issue.
I think a lot about AI in democracy. In the fall, I published a book called “Rewiring Democracy” that looks at AI politics, AI legislating, AI administration, AI courts, and AI citizens. This is a fundamental technology that is going to change all sorts of things at every level. And we’re in the middle of watching it change.
Some of it will be bad; some of it will be good. AI enhances power. The changes are going to be real. And I think this upcoming election, we’re going to see a lot of them.
Bruce Schneier is a security technologist, an adjunct lecturer in public policy at Harvard Kennedy School, and a fellow at Harvard’s Berkman Klein Center for Internet and Society.
—
Teddy Svoronos: On teaching generative AI
The challenges are significant in terms of teaching generative AI, and the reason is twofold. One, these things are changing incredibly quickly. We teach a course on generative AI. Every semester, we have to substantially revamp particular classes to just keep up with what’s happening. There has been a categorical shift in what’s possible with these tools. The other reason is that, from a learning perspective, it is quite easy to have AI short circuit your learning. People can use AI to help them learn. When you’re doing an assignment, it’ll accomplish that assignment as quickly as possible, but it won’t help you learn as quickly as possible. This tension with using these tools well and internalizing the difficult stuff that is in coursework is a challenge.
“No matter what area students go in their careers, they should try to think of what it might look like when the core skills, the core competencies, become something that’s easy to automate.”
Even with new, very advanced tools, the need for very specific narrow-domain expertise is really important, and being able to use these tools well is also really important. Those two things now trade off against each other. The more I use AI tools, the less I rely on my own intuition, my own judgment. That process might be undoing some of the stuff that I’m trying to learn. I need to be careful about how and when I rely on them because, in the process, I might be hurting myself in the long run.
No matter what area students go in their careers, they should try to think of what it might look like when the core skills, the core competencies, become something that’s easy to automate. How could you think about your value-add, and how can you make sure that you’re bringing the judgment that you’ve developed in your life into the way in which you use those tools?
Teddy Svoronos is a senior lecturer in public policy. Along with HKS faculty members Sharad Goel and Dan Levy, Svoronos teaches degree program and executive education courses on using generative AI.
—
Illustration by Stuart Kinlough / Ikon Images; headshots by Martha Stewart