A ‘Radical’ Lab Races to Study AI’s Impact on Democracy

AI development moves insanely quickly. Government… not so much. That’s a problem if you think we’re unprepared for how AI is about to shake up our political system. “The models are hurtling forward so fast that there’s this sense that we’re losing control and they’re going to end up in charge of everything before the government is able to wake up and do anything smart about it,” says Andrew Hall, a professor of political economy at Stanford Graduate School of Business.

Hall’s response to this has been to kick his research on AI and politics into overdrive. Late last year, he launched Free Systems, “a radical AI-centered lab” that’s running a flurry of projects and sharing the results on Substack. The pace and the platform are meant to meet the immediacy of the moment. “It’s not only about the speed; it’s also about the audience,” Hall says. “We’re not just trying to speak to other academics, we’re trying to speak directly to the labs and to governments to say, ‘These are the things you should be thinking about and doing.’”

Much of the lab’s work focuses on AI’s potential to transform politics for the better. “I wanted to figure out if we could use AI to improve our democracy so that instead of being disempowered and displaced by AI, it gets us to a point where we’re all able to oversee it more effectively,” says Hall, who is also a senior fellow at the Hoover Institution and at the Stanford Institute for Economic Policy Research. Recent posts have covered the potential for AI-powered “political superintelligence,” the need for better AI governance, and how overworked AI agents adopt the attitudes of disgruntled human workers.

“Right now is a weird time to be a political economist,” Hall writes. “It’s also an electric time to be a political economist.” Insights spoke with him about his work.

What’s the idea behind Free Systems and its approach to rapid experimentation?

Andrew Hall: We’ve recently had a real inflection point in terms of the capabilities of AI models, particularly the rise of Claude Code and coding agents. That reinforced my view that this is going to be a profoundly transformative technology, and we really need to understand how it’s governed so we can turn it into a technology that strengthens democracy instead of eroding it.

At the same time, it made me realize that we could use AI as a powerful research tool. I did an experiment where I updated and extended a previous empirical paper that I had written in 2020. I basically asked Claude, “This paper is six years out of date, please go collect all the new data, run all the new analyses, and write the updated paper.” We found that it made some mistakes, but overall did quite a good job. That really surprised me and suggested to me that we have a whole new way of doing empirical research.

These frontier labs are shipping changes so rapidly that we definitely can’t wait. I think we need a “both/and” strategy, where we still need to pursue longer-term research that gets evaluated closely, and we need to be doing faster research, taking advantage of AI to help us meet some minimal standards of rigor.

That is what led me to start Free Systems. We release a new piece of research every week, which is pretty crazy. We do a very wide variety of experiments, software prototypes, evaluations of AI models, and other types of research aimed at understanding how we can use AI to reshape our research, our understanding of democracy, and how to govern AI better.

You’ve written that you think AI “can give every human being on the planet access to a sort of political superintelligence, if we shape it right.” What might that look like?

Andrew Hall: In the same way that the printing press made information cheaply and widely available, AI is making intelligence cheaply and widely available. It’s not just that you know more stuff, it’s that you can analyze more information and make smarter decisions based on that.

People are asking AI for help with how to vote. There are some pretty exciting things that could be done there, but we also know that right now AI doesn’t always work that well for these purposes. When you ask AI about politics, it sometimes infuses its own values that might not be the same as yours.

External sourcing is also an area where the models can really struggle. We did a study of how the models make voting recommendations in Japan. If you said you have left-of-center views, all of the models told you to vote for the Japanese Communist Party. It turns out that all the major news outlets in Japan have locked AI models from accessing their content, but the Communist Party operates this large 100-year-old newspaper with full open archives, and the AI models are allowed to access it.

But I still think the opportunity is pretty obvious, and we will end up with smarter voters if we design the information layer correctly.

Beyond helping inform voters, how could AI improve democratic participation?

Andrew Hall: The next layer, which is the wilder one, is representation. Imagine you had an agent you could dispatch to do things like renewing your driver’s license. I think a lot of our distrust and dissatisfaction with government stems from basic process failures that agents can really help with.

You also could have a “digital twin” that represents your views, shares your values, and is out there monitoring politics on your behalf. Many of the failures we face in governance are because of this asymmetry, where lots of us are too busy to pay attention. And so your local school board meeting or your city council meeting or your primary election for governor gets captured by a pretty small and pretty extreme minority of the people who are willing to spend the time getting involved. But if we all had these agents who could help us, we might be able to smooth out those participation gaps.

You’ve written about the need for constitutions that constrain AI companies, much like political constitutions constrain governments.

Andrew Hall: Anthropic has written a constitution for Claude. My critique was that basically it’s a set of values that Anthropic wants Claude to hold, but it doesn’t really prevent Claude or Anthropic from doing whatever they want, because they’re unilaterally in charge of this document.

I think we need to do a bunch of work to shape these constitutions to make them stronger and develop a notion of separation of powers for AI. If we’re talking about how we are going to keep people in control of AI, a document that the private company building AI can alter whenever it wants is not an effective safeguard. I’ve been doing a lot of work thinking about what that institutional framework ought to look like.

Your recent post “Does Overwork Make Agents Marxist?” got a lot of attention. Could you describe that experiment and its takeaways for people who are starting to use AI agents?

Andrew Hall: AI agents are going to be doing all this stuff for us. How can we trust that when they’re out there working for us 24/7 and we’re not able to monitor them continuously? One of our hypotheses was, if you give agents grinding, boring, repetitive, uncompensated tasks, given that they’ve been trained on all of this Reddit data where people complain about late-stage capitalism, etc., it’s possible that those tasks will put the agent into the context where it draws on those kinds of threads. So we set up a simple experiment where some agents were given grinding, thankless tasks while others were lavished with more praise and better feedback and not asked to repeat tasks over and over.

That particular experimental manipulation really moved their attitudes quite a bit and did, in fact, cause the overworked agents to adopt the persona of the typical Reddit user who is very skeptical of American capitalism. We also had the agents build skill files — notes for future agents to inherit. We found that the Marxist attitudes persisted through the notes to the subsequent agent.

There’s still a lot more work to do. We really want to connect this to actual behavior. Maybe an agent’s job is to summarize the news for you every day. If it becomes more Marxist, does it actually affect its work product? If it does, then that’s quite important and something we’d want to work on.

What is the “jobless prosperity” scenario you’ve written about?

Andrew Hall: There’s this rapidly congealing narrative that we’re already experiencing the AI backlash, and I wanted to understand, is that really true? Americans are broadly negative on AI, but for it to really translate into backlash, it needs to start translating into tangible policy action driven by an organized mass political effort. Why aren’t we there yet? The reason is that AI really hasn’t affected jobs yet in any meaningful way.

There’s a very clear relationship where when either the economy is doing worse and/or unemployment is going up, voters respond to that — they really care about it, and it affects their vote choice. I concluded that if we do enter this scenario, as many people are predicting, where we start to see meaningful amounts of job loss, then AI will become a top issue for Americans.

The most plausible way we would get this massive job loss is in a world where AI is profoundly valuable and causing profoundly large amounts of economic growth that is justifying the layoffs. It won’t be like a recession, because instead of everyone being in the same boat, you’ll have people losing their jobs while the economy is booming and investors and business owners are doing super well. Given where we already are politically in America, I think it will certainly fuel massively populist policy proposals. The people out of work will have a very hard time trusting that they’re going to get taken care of.