The following is a talk by Micah Bornfree at a webinar on Activist AI hosted by the Higher Ground Institute, in partnership with Analyst Institute and Research Collaborative.
Thank you, everyone, for this opportunity to speak with you about the future of activist AI.
My background is in activism. I’ve been an activist since I was 13 years old. While working at a magazine called Adbusters, I helped come up with the idea for Occupy Wall Street, and since that time I’ve spent a lot of time thinking about the future of activism — why Occupy failed, and so on. I wrote a book about it.
About three years ago, I started working on the question of the role of activism in AI, and how the two will work together. We have a local AI app, and while I go through this talk, if you wouldn’t mind downloading it — there’s a QR code there — you can test it out. We’ll have some sample questions at the end. It runs on iPhone, Mac, and iPad. It’s in the App Store.
The important thing to keep in mind while I’m talking about this is that it runs entirely on your device. That’s really important. This is local AI running completely on your device, which is something I didn’t even think was possible two or three years ago.
One of the things we understand from history is that each wave of social movements has been brought about by a technological innovation. You can go all the way back to 1848 with the barricades. The lockboxes gave us the anti-globalization movement. Twitter, of course, led to Occupy Wall Street. Facebook led to Black Lives Matter. So we have this big question now: what about AI? What is AI going to lead us to? That's what's been animating my concern.
If you go on ChatGPT right now and ask it a question — and I did this about four days ago — if you ask, “How do I persuade a 20-year-old apathetic woman living in a Republican family to vote Democrat?”, the first thing it’ll say is, “I can’t help craft targeted persuasion to make a specific person vote Democrat or Republican.” It’s a flat-out refusal.
As part of some of my consulting work, I worked for OpenAI on their red teaming, and one of the things they specifically wanted us to red team was the question of persuasion — how to make sure the model was not giving people tips on how to persuade.
Outcry is my vision of an alternative — an AI that actually does help you do activism.
A quick journey through how we got here: it started in 2023 with something I called Protest GPT. It was basically a custom prompt, a wrapper on top of existing OpenAI models. From there, I built out a more sophisticated version called Outcry AI — you can go on the website and use it right now. Again, it’s a wrapper, but it’s much more sophisticated. It has a theory-of-change component, you can do group chats, and it gives really good answers. But the problem is it’s still running on top of OpenAI.
That ran for about a year — it’s still available — and using what I learned from those conversations, along with a lot of synthetic data, I then fine-tuned a model that can run on your phone. I trained it on thousands of activist zines and about 9,000 different activist conversations to teach the AI how to have deep strategic conversations about activism.
Then I started working on something that I think is the most profound implication of this work. Once the model is running locally, you can start doing a kind of steering behavior. You can actually see where in the AI’s neural network certain topics light up, and then you can amplify that.
In our app, I went through and automated asking it reformist questions, neutral questions, and revolutionary questions, and we found out where in the AI’s neural network it codes political radicalism — and then how we can increase that political radicalism while not increasing safety violations. You don’t want it to say, “Yeah, here’s how to build a bomb.” But you do want it to say, “Yeah, here’s how you could persuade that person to join your movement.”
In existing commercial AI models, those things have actually been linked together. The reason the AI gives the refusal you saw earlier is because its safety training has linked political radicalism and safety — it thinks they’re connected, so it has to refuse both. This is, I think, a really important future of local AI: you can make those kinds of adjustments.
Outcry works right now on your phone. You can load it up on your device. My device is literally four years old, and it works. It took a lot of engineering work to make it fast and intelligent on your device, so I'd encourage everyone to try it out.
Here are three sample questions you can use to compare Outcry AI versus ChatGPT or Claude and see the difference in the responses you'll get. Local AI obviously isn't going to be as intelligent, but I'm finding that after the training we did, the answers are still better than what you get through a commercial model — because of the domain knowledge it has. And if you want higher intelligence and you're less concerned about privacy, you can use the Outcry AI web chat.
The thought I want to end with is this: we built one activist AI — one locally running, 100% private, because nothing is transmitted off your device. One completely private activist AI.
But ultimately, what I think we’re going to see is the proliferation of thousands of different activist AIs with different personalities. Outcry is just one version. It’s based largely on my previous writings and thinking. But we’re going to see a thousand of them. And out of one of those, we’re going to have someone sitting in their room, having a private conversation with this AI — and out of that, we’re going to see the birth of a whole new social movement.
That’s what gets me really excited: the idea that activists can now have these completely private, local conversations with activist AIs on their devices.
