Putting power back into our hands: the opportunity of AI

Somewhere in the last thirty years, the politics stopped serving the everyday people it claims to care about. They can still vote, but voting is just one part of the political process; it’s not the whole. What went out of reach was the practice: campaigning, strategy, the local work of building a collective vision that could win.

This happened because the knowledge moved. Doing politics well now means understanding persuasion, and polling, and mobilization, and communications, and how a message lands differently in two towns forty miles apart. Since the 2000s it also means email optimization, segmentation, testing, targeting, analytics. (And now we’re seeing what changes with the AI landscape, which has the risk of outputting more of the same but at a higher volume). Each of those is a discipline with its own literature and its own people who have given years to it.

Organizing got left out of that build. It is treated as the old art, the one that does not need updating, and most of what gets called digital organizing is mobilizing with a nicer interface. “Field” sometimes in campaigning is thought of as organizing, which is a little laughable considering its innovations usually involve increasingly more treating it like a mass advertising activity, than anything where people actually come together to work out a plan of collective power. 

What else happened was the subject of chapter four. A consultant class formed around the specializations, and access to them is what it sells. That is not a conspiracy. It is the market result when knowledge is hard to retrieve: lives in the heads of the people who spent their careers collecting it, and the only way to get at it is to hire one of them.

And the further that knowledge moved from the people it concerns, the less those people were truly engaged. Constituents became targets of political practice with their agency ignored, because the deciders, the planners, the strategists became a class of people who had to wear a specialist badge to get any access into the producer class of politics. 

People know this. Not analytically, but in their guts. They say politicians are phony, that they will say anything, that it’s not worth believing in the system.. They are not wrong to think this way. The message was focus-grouped and optimized and split-tested, and it is anything but somebody speaking from the heart of a community they belong to. That is what happens when you treat people as marketing subjects rather than as participants in a democratic process, and it is most of what this populist decade has been reacting to.

On a Tuesday somewhere, somebody decides to organize people where she lives. She talks to people, at an event and one to one. They say yes. They say they are interested. Three weeks later, three people have signed on to help and not one of them has done anything at all.

She wants to know why, and there is an answer, and it is well understood. It has been studied. There is a literature as old as organizing itself.

Her options for getting at it are these: Wait for time from a staff organizer, who has none. Take a course on organizing (Marshall Ganz's is excellent). Read Alinsky, read Groundbreakers, work through the Stanford Social Innovation Review, find the field experiments and get through the ones written for other researchers. She doesn’t have time for any of that. Or of course, she could quit her job, go back to school and become a professional ‘who knows’.

Every one of those is an option, but none of them are realistic. Every one of them costs time she does not have. She is not refusing to learn. But she has a job and a life, and the return on giving her evenings to politics is hard for her to see.

We keep building a politics that requires people to become professionals in order to take part in it, and then we are surprised when they do not take part.

The knowledge hasn’t really been hidden. It was expensive to retrieve, which amounts to the same thing.

That is the specific thing I foresee that this new type of technology changes, and it is not the thing most people are watching. The attention is all on production. Faster content, more of it, cheaper, and a great deal of what gets produced is the slop I have spent this week describing. Underneath the noise it is doing something quieter and more consequential, which is making a vast body of accumulated knowledge reachable by somebody who does not know its vocabulary and cannot name the thing she is looking for.

David Autor has spent a career documenting what computers did to work, which is that they ate the middle. Tasks that could be written down as rules got automated. What survived at the top was expert judgment, because judgment resists being written down as rules, and Michael Polanyi's line for that is still the best one: we know more than we can tell. Autor's argument about this generation of machines is that they can reverse the direction of travel, because a machine that learns from examples rather than from rules. The opportunity, in his words, is "to extend the relevance, reach and value of human expertise to a larger set of workers."

There is evidence and it points the same way. Brynjolfsson, Li and Raymond studied 5,179 customer support agents through the staggered rollout of an AI assistant. Productivity rose 14 percent on average, 34 percent among the least experienced workers, and almost not at all among the most experienced. The distribution of that gain is the finding. They explain it themselves: the model "disseminates the best practices of more able workers and helps newer workers move down the experience curve."

Customer support is a field where the good ones know a great deal that the new ones do not, and where almost none of it is in the manual. Ours is the same shape, at higher stakes.

What that means for the woman in her town is not having to take a course, various workshops or read 10 books. These are all general and her problem is specific, which is why she is not going to take one. In the past, she would have to know the whole field to even be able to work out what it is she needs to solve within the full range of the expertise. It is the accumulated thinking of people who gave careers to the exact question in front of her, reachable at eleven at night, in the words she actually has rather than the words the field uses when it talks to itself. Learning stops being a syllabus and becomes an answer to the thing that happened on Tuesday.

All of which holds only if what comes back is valuable. A retrieval machine is worth exactly the value of its sources, and in our field there are three ways that goes wrong. I have seen all of them in action.

The first is the wrong logic. Most of what has ever been written about getting people to act was written by people selling things, and the persuasion literature is enormous, well funded and confident. Ask a general model why nobody is showing up to your group and it will answer you out of that corpus, in a register of funnels and conversion and messaging, and it will sound like it knows.

The second is the error everybody holds. I argued on Thursday that these models are built out of the consensus. The consensus in our sector is that mobilizing is organizing, and it is wrong. There is a cost to that consensus.. A machine built out of the consensus will hand you the consensus back with or without a citation.

The third is the part that was never written down at all. Most of what a good organizer knows has never been formalized. It lives in the debrief, in the read of a room, in what she does when a meeting is going sideways and she can feel it eight seconds before it’s vocalized. That is Polanyi's paradox again, and it is the part that scale does not fix. A larger dataset is a larger scrape of what people wrote down, and if it’s not documented it’s simply not there. Some of this is worse than unwritten. The best organizer you know could not put on paper what she does when a room turns, because she does not think of that process as a process at all. It’s human afterall.. Knowledge like that comes out of people by asking, and by watching them work, over years.

Andrew Peterson has a name for what happens when a whole field leans on retrieval without minding any of this. He calls it knowledge collapse: models generate toward the middle of their distribution, and the more we rely on them the more gaps of knowledge remain untapped, until the tails of the field are simply gone. Our practice lives in those tails. Almost everything good that has happened in organizing came from somebody at the periphery doing something that the center had not formalized yet.

Retrieval is a capability rather than a gift. Somebody has to put the right knowledge within reach, keep it current, and be accountable for what is in there and what is not. That is fourteen years of our work, and it is why our coach cites the thinkers and researchers and organizers we admire - why it cites and links to the books of Ganz, McKenna, Alinsky, Han and many others by name instead of answering out of the middle of the internet.

Why not build a bot that answers the question? Because the answer is not in the literature. Why people are not showing up for her group depends on her town, on what she asked, on who those people are and what happened to them last year. Or the relationship. Or the body language when she asked them. It changes by community and by scenario. It is a human skill, and it is why organizers exist.

If it were an equation, somebody would have taken over politics with machines already. People have been trying for decades, with more money than the rest of us will ever see. What they have produced is influence, at the margins, and a politics nobody trusts.

She has the thing the machine does not, which is the broadest lived context. Her town. Her people. What she actually said at that event and how it landed. That is lived experience, and it is what a good consultant is tapping into when they sit with a campaign director and ask questions for an hour briefing. Expertise applied to somebody's context is the only combination that produces something that works.

The Experience/Expertise Boundary is the whole design problem, and I have come at it from three sides this week. The machine holds what the field knows, she holds what only she knows, and the entire question is whether the thing was built to ask her or to guess. Commercial AI guesses. It fills the blank, confidently, with the middle of its training data, and if she does not know enough to spot where it should have stopped, she will never know it stopped nowhere.

People ask what happens to their data, and they are right to ask. Where it goes, what is kept, who can see it, how much an organization actually controls. Those are legitimate concerns, we have a policy, and parts of it are still being worked out. I am not going to hand anyone reassurance in place of an answer, and anybody in this sector telling you the question is settled has not looked at it hard.

The larger one is not about anyone's organization. The environment is the thing that threatens the species, and this technology's appetite is real. The full impact is not clear, and anybody quoting you a confident figure is selling something. What is clear is that energy use is climbing substantially and the curve is not bending.

What I would resist is the reading that says this is simply what the technology is. Waste is not a property of matrix multiplication. It is a property of corporations doing what corporations do, racing each other to build capacity ahead of demand, siting data centers where power is cheap and oversight is thin, taking the fastest route to a return because the profit motive takes them there and nothing is pushing the other way. That is a pattern we know from every extractive industry that came before this one, and we did not conclude from any of those that the answer was for us to sit them out.

What pushes the other way is regulation, taxation, carbon requirements, and forcing alternatives to exist. None of that arrives without political will. Political will has to be built by somebody, and building it is organizing. Our abstinence will not save us, which I talked about in depth in chapter 14.

That is the argument for activists having more tools for collective power, not less.

The excitement about this technology runs toward the machine doing more on its own. Agents that carry out a sequence of tasks without supervision, each wrong turn compounding into the next one, producing strategic slop at a scale nobody will read.

Set against that, it gets easier to say what people are for. The usual answers come as a list: lived experience, values, creativity, insight, imagination. All of it stays abstract until you ask what those things are actually for. They are for seeing past what is already held to be true. This is not only the primary value in using AI. It is the primary value in democratic practice. 

A model can accelerate serious work enormously, because it is an excellent account of where a field has got to. It cannot tell you the field is wrong. Science advances when somebody looks at the settled account and refuses it. It doesn't come from summarizing literature. It comes from a person who noticed a problem in their community and had the imagination to follow it. Our work is the same and always has been. Every gain we have was won by people who could see a world that did not exist yet, on no evidence beyond their own lives and what those lives had told them was possible.

I wrote on Thursday that transformational politics means moving the consensus, and that these machines are built out of the consensus. Only one of those two things can move.

That capacity does not sit in one head, either.

The thing we are best at is being collectively intelligent. More than any other species, and more than the other members of our own genus who were intelligent individually but did not advance as homosapiens did. We think in groups. We hold knowledge between us that is not held by one individual.. It is how we built every piece of technology we have ever built, including this one.

It is also how democracy gets built, and it is why a person alone with a machine will never be the unit that matters. I wrote in chapter five that no single mind holds enough of the context to do strategy well. The same is true of a machine, for the same reason, and the answer is the one organizing has always given.

That is what the coaching layer actually is in our Organizing Center platform, when you strip the technology out of it: the accumulated thinking of a field, made retrievable by one person, at eleven at night, on the specific problem she’s facing It is there for her to explore and apply with her own lived experience, insight, and agency. 

This is from a pilot we are running with one of our client organizations. Different person, same type of Tuesday night:

"I'm having trouble getting people to show up and follow through to commit to the group. I've had people say they would... but none of it worked."

"How long has this been going on?" … "When people said they were interested — how did those conversations usually go? Were most of them one-on-one conversations, or more like you sent out a link?"

"When someone says yes in a moment of enthusiasm, what they're really saying yes to is the feeling, not a specific thing they're going to do... when someone says yes to your excitement, they haven't yet said yes to their own commitment. Those are two different yeses, and the second one is the one that matters for follow-through."

"There's a practice in organizing that might help reframe this. McKenna and Han call it the one-on-one... (McKenna & Han — Groundbreakers, ch.4)"

Then he half agrees.

"I guess. I could ask them. Unless you recommend something else?"

"'I guess' tells me it kind of fits but maybe not fully. That's worth slowing down on for a second, because if the concept doesn't quite land, the next move won't either."

"When you picture going back to one of those people who said yes and then went quiet, what feels like the hardest part of that conversation for you?"

"Rejection I guess."

"Is the fear of rejection the main thing in the way of going back to people, or is there something else sitting underneath it?"

"No. I think it's that. Like I don't want them to feel uncomfortable having to tell me no... and then I am just putting them into an uncomfortable position and everyone feels awkward. I dread that."

The stated fear was rejection. The real one was two questions further down, and the organizer found it himself.

It cites and links to where the ideas came from, so he can go and read McKenna and Han if he wants to. It does not write the message for him, and when he asks for a script it sends him to a different coach, made for writing and outreach. Every decision stays with him.

We are not asking anybody to stop learning. We are trying to make it so that nobody has to learn a whole field of politics before they are allowed to be an agent of democracy in the place where they live.

The knowledge that got locked up in a professional class is the same knowledge that lets ordinary people do real political work, and for the first time there is a way to get it to them that does not require them to become us.

They were never supposed to be the audience for politics. They were supposed to be the ones doing it.

And now it's time to put it back into our hands.