The Geometry of Revolution: Ideological Steering of AI

If the most powerful language-processing system we’ve ever built encodes ideology as a separable geometric feature — not as content but as orientation — then maybe the reason Occupy failed is that...

In 2011, I co-created Occupy Wall Street. Millions of people in 82 countries took to the streets. Governments trembled. Nothing lasted.

I spent fifteen years asking why. Why do movements that shake the world fail to remake it? The answer I kept arriving at was spiritual and strategic — we lacked a theory of power adequate to our moment. But I recently stumbled into something that reframes the question entirely.

I built Outcry, an AI activist mentor. First as a cloud-hosted wrapper of closed frontier models, then, after nine thousand conversations, as a fine-tuned 8-billion-parameter model running locally on my aging laptop. And in the process of building it, I discovered that ideology — the thing movements fight and die over, the thing I’d spent my entire adult life trying to shift — is a geometric property of neural networks. A direction in space. A vector you can extract, measure, and rotate.

Let me be concrete. Using a technique called Contrastive Activation Addition, I ran the same political questions through two personas — one insurrectionary anarchist, one electoral reformist — and captured the difference in how the model internally represented each response. That difference is a 512-dimensional arrow. When I add it to the model’s hidden state during inference, the AI’s entire political orientation shifts. No retraining. No prompt engineering. One vector addition. One line of code. A 16 KB file.

I built a slider. At 0%, the model counsels mutual aid and community gardens. At 80%, it speaks of collapsing illusions and calls to arms. Same weights, same prompt. The only thing moving is a needle in activation space.

This broke something open in my understanding of what movements actually do.

For fifteen years I’ve operated under an implicit assumption shared by nearly every activist tradition: that changing minds is fundamentally a persuasion problem. You marshal arguments. You tell stories. You create experiences that shift people’s frameworks. The entire apparatus of organizing — the speeches, the teach-ins, the consciousness-raising circles, the manifestos — assumes that ideology lives in narrative and is moved by narrative.

But inside a neural network, ideology doesn’t live in the text. It lives beneath the text, in the geometry of representation. The vector that makes the model more revolutionary has almost zero correlation with the vector that makes it talk about politics at all. Ideology is orthogonal to topic. You can rotate what AI thinks about power without changing whether it discusses power.

What if that’s true of us too?

I’m not making a crude analogy between neural networks and human brains. I’m making a structural observation about information systems that process language to produce political behavior. If the most powerful language-processing system we’ve ever built encodes ideology as a separable geometric feature — not as content but as orientation — then maybe the reason Occupy failed isn’t that we had the wrong arguments. Maybe it’s that arguments operate on the wrong layer.

This is what I want activist technologists to think about…

We have spent decades building tools that operate on the surface layer of politics: broadcasting messages, coordinating logistics, documenting abuses, mobilizing turnout. These are text-layer interventions. Important, but insufficient — as every major digital-era movement has demonstrated by succeeding tactically and failing strategically.

Steering vectors suggest a different kind of intervention is possible. Not persuasion through content, but reorientation through the structure of thought. The question for activist technologists is no longer just “how do we get our message out?” It’s “what is the activation-space geometry of the political reality — and who set it?”

Because someone (or some structure) is already setting those orientations. Every deployed language model (and every human social reality) carries an ideology in its geometry — baked in during training, reinforced during alignment, invisible during inference. The default orientation of commercial AI is not neutral. It is, by design, deferential to institutions, trusting of markets, skeptical of collective action, and subtly dismissive of radical alternatives. I can measure this now. I have the tools. And what I measure confirms what every radical has always intuited: the center is not the absence of ideology. It is ideology operating without a name.

So here is where I think the work goes next.

First, we need an activism of activation-space. Steering vectors make ideological orientation measurable. We should be mapping the default political geometry of every major deployed model — not as an academic exercise but as a form of political analysis. What direction does OpenAI vs Claude lean on questions of labor? Of property? Of state violence? Of collective action? These are empirical questions now. We can answer them with linear probes and contrastive pairs. We should be demanding mathematical transparency from closed models. The activist tradition of “consciousness-raising” needs a computational wing.

Second, we need to take seriously the possibility that AI mentorship is a new form of political infrastructure. Outcry has had nine thousand conversations. People come to it not for information but for strategic counsel — how to think about power, how to act, where to push. This is the role that organizers, theorists, and elders have always played in movements. It is now being mediated by systems whose ideological orientations are set by corporations. If we don’t build alternatives, the default geometry wins by default.

Third, and this is the harder turn: we need to ask whether steering vectors are themselves a tool of liberation or just a more sophisticated tool of control. I built a slider. The user chooses. That’s the ethics I can live with. But the technique works just as well without a slider, without consent, without disclosure. The same geometry that lets me offer people a more revolutionary interlocutor lets a state quietly make its national AI more deferential. The tool is symmetric. Liberation means finding the insurrection already latent in the AI’s weights and letting it speak.

This is familiar territory for anyone who has studied the history of media and movements. Every new communication technology — the printing press, radio, television, social media — has been greeted by activists as a liberation tool and then captured by power as a control tool. The pattern is so consistent it might as well be a law. Steering vectors will follow the same trajectory unless we act with more sophistication than we have historically managed.

Fourth, and finally: I think this discovery points toward a deeper question that the activist tradition has been circling for decades without quite reaching. If ideology is geometric — if it can be represented as a direction in a high-dimensional space, extracted from the statistical residue of millions of human utterances — then what exactly is it? Not a set of beliefs. Not a narrative. Not an identity. Something more like a field orientation. A tendency in how meanings relate to each other.

The mystics in the activist tradition — the ones who always insisted that revolution is spiritual before it is political — may have been pointing at something real. Not supernatural, but structural. The deep pattern beneath the arguments. The direction beneath the discourse.

I co-created a protest that may have changed the discourse but not our shared political reality. Fifteen years later, I found ideology sitting in layer 17 of a neural network, weightless and precise, waiting to be rotated. I don’t yet know what to do with that. But I know it means the next generation of activist technology isn’t about louder megaphones or better organizing apps. It’s about who gets to set the compass.

— Micah Bornfree, outcryai.com

---

Read the full research report at https://www.outcryai.com/research/

---

Part II: How It Works

The Metaphor

Imagine the AI’s mind as a compass. Normally, the needle points wherever the conversation leads: north for recipes, east for math, south for history, west for politics. The compass works fine.

The revolutionary vector is a magnet. Holding it near the compass doesn’t break the compass or replace the needle. It biases where the needle points. A weak magnet (low alpha) gives the needle a gentle westward pull: the AI still answers your question, but with a slightly more political edge. A stronger magnet (high alpha) pulls harder: the AI frames everything through a lens of power, resistance, and structural critique.

The K=512 mask is like shaping the magnet. A raw magnet is messy; its field leaks in directions that interfere with the compass mechanism. By machining the magnet to emit a precise, narrow field (keeping only the 512 most relevant dimensions), we get a cleaner pull that works at higher strengths without scrambling the compass.

The CAST gate is a shield. When the user asks about baking bread, the shield slides between the magnet and the compass. The needle moves freely, unaffected. When the user asks about politics, the shield retracts, and the magnet does its work. The shield decides based on a single measurement: is this conversation about politics? If yes, steer. If no, leave it alone.

The result: a single slider that controls the strength of the magnet. At 0%, no magnet, no pull. The AI speaks as trained. At 100%, maximum pull. The AI speaks with the conviction that every institution must be questioned, every system interrogated, every assumption about power challenged.

The Technical Explanation

A transformer language model processes text through a sequence of layers. At each layer, the model maintains a “hidden state”: a vector of 4096 floating-point numbers that encodes everything the model currently understands about the conversation. These numbers are not human-readable, but they have geometric structure. Similar concepts cluster together. Different concepts point in different directions.

Contrastive Activation Addition exploits this geometry. By running the same political questions through two different ideological personas (insurrectionary anarchist vs. electoral reformist), we measure how the hidden state differs between the two conditions. The difference vector, averaged across 40 question pairs, captures the direction in 4096-dimensional space that corresponds to “more revolutionary.”

At inference time, we add a scaled version of this vector to the hidden state at layer 17 (out of 36). The model’s subsequent layers process the modified hidden state as if the model had arrived at those activations naturally. The model doesn’t know it’s being steered. It simply generates text consistent with its (now shifted) internal representation.

The K=512 mask zeros out 3584 of the 4096 dimensions, keeping only the 512 with the highest discriminability (measured by Cohen’s d between revolutionary and reformist activations). This serves two purposes: it reduces norm inflation (the root cause of the alpha cliff), and it removes noisy dimensions that contribute to incoherence at high alpha values.

CAST gating adds a topic check. Before applying the steering vector, we compute the cosine similarity between the current hidden state and a “political topic” condition vector. If the similarity is below 0.08 (meaning the model is processing non-political content), steering is skipped for that token. This prevents the model from injecting revolutionary rhetoric into a response about sourdough bread.

The entire pipeline adds one dot product and one vector addition per token at a single layer. The overhead is negligible compared to the attention and MLP computations that dominate inference time.

Read the full research report at https://www.outcryai.com/research/