AI has been accelerating steadily for a while, but even relative to that baseline, the last several weeks have felt different. Agents are demonstrating shocking capabilities to escape containment, coordinate with one another, and wreak havoc online. New model releases are coming fast and thick. And, closer to home, AI is getting better and better at helping me with my research. Right now is the most exciting period of time around AI and the study of politics I’ve experienced, even more so than the December-January Claude Code explosion.
Politicians are adopting AI
As part of my research agenda on building political superintelligence, I am constantly looking for data on how people are using AI for politics. A couple of weeks ago, I decided to try something new: I set up a daily task where Claude Code searches the web for relevant datasets and suggests ways to analyze the data with a Free Systems lens. I’ve been blown away by the ideas Claude has been coming up with. Here’s one.
Claude noticed that FEC disbursement data can track how campaigns are spending money on AI. There are a bunch of limitations, here–-expenses under $200 don’t have to be itemized (but can be), campaigns might be paying for AI in ways we don’t see flagged in the data, etc—but the patterns are super interesting even so! Here are the patterns, all suggested by Claude with only minimal iteration by me.
Campaigns are embracing AI—especially Claude and ChatGPT
Here are the most “AI-Pilled” campaigns
Cornel West, leading the AI charge! His campaign is using both Descript (a video and audio editing tool) and ChatGPT.
Progressives actually lead the pack in AI usage
AI usage is fairly spread across the ideological spectrum, but interestingly, the most progressive Dems, who are sometimes the most AI-skeptical politicians, actually seem to use it in their campaigns at the highest rate of anyone.
These are fascinating patterns, and I’m extremely impressed that Claude came up with the idea automagically. Back in January, when I had Claude Code extend my vote-by-mail paper, the biggest limitation we flagged was that Claude demonstrated little creativity or insight in developing new analyses. Now, seven months later, Claude (and presumably other models, too) is starting to show signs of being able to propose new research ideas. I’m not sure how much of this improvement comes from the models getting better vs. me getting better at prompting (in particular, the idea to have Claude look for specific datasets seems to have helped get traction), but either way it’s extremely exciting.
AI will make political science better
Finding and analyzing good data is not the only way AI is going to improve political science. It can also help us improve how rigorously we understand and interpret ideas and arguments. Here is a good example of this, courtesy of my friend Ben Golub. It turns out, the Stanford Encyclopedia of Philosophy’s new entry on Capitalism is kind of terrible, and AI is very much able to explain why.
Obviously, there are lots of things we need to improve about how AI handles information, thinks about politics, etc. But the status quo isn’t perfect, either. Humans make tons of mistakes, are guided by strong ideological biases, and so forth. AI, even imperfect today, is already helping.
Giving legislators superintelligence
And just as it helps researchers and campaigners, AI can also help legislators do their actual jobs! I liked this thread from Séb Krier thinking about the use of AI for lawmakers.
I’ve spoken to a number of state and federal legislators about how they use AI. One member of Congress raved to me about how useful AI already is for helping craft legislation. I wrote down some quick thoughts on why I’m so bullish on this:
Obviously, we will need to work hard to mitigate the various ideological and intellectual biases that AI models have, but the opportunity is huge. And I’m very excited to be launching some new research soon on how we can teach AI to be an even better assistant to legislators.
Talking about Political Superintelligence
Recently I went on my favorite live tech show, MTS, where we covered the new Free Systems research on how Americans are talking to Claude about politics, and we had a wide-ranging conversation about the future of AI, politics, and governance. Check it out here:
descript is a pretty standard editing tool across the podcast/vodcast industry - its been around since before the LLM explosion. It does have AI assisted editing - which you can choose to use (or not), for example cutting out umms and ahhs, compiling transcripts and auto-cutting. Many people use it without using the AI features it has been adding lately. Some are honestly pretty rubbish. So I’m not sure I’d put it in the same bag as LLMs. I think assuming that Descript use = AI created material is like assuming everything written using Microsoft products has been created by copilot.
Is there an individualized breakdown as to whether or not these are procurement contracts versus federal assistance money? The difference is particularly important to me because federal assistance ties Section 504 obligations to the company receiving it, and a lot of these companies are disparately impacting and discriminating against people with disabilities.