Elevation over automation

Everywhere I look, somebody in the political sector is building something with AI.

More content. More analysis. More email variants, tested against each other. More segmentation, and then more reports about the segmentation. Faster asks, at higher volume, personalized in a way that is somehow more industrial than personal. Most of it is being built by people I like and respect, and almost all of it is pitched on the same promise, which is that it will save you money and time.

Just what our democracy needed: More email. More texts. More optimization. More more more….of everything but people. 

Despite the claims, nobody is going to save any money. What we are going to get is an overwhelm of junk. You can already see the leading edge of it on the internet, where the cost of producing plausible text fell to nothing and the volume rose to meet it. In the coming election we will see it at full strength: more emails nobody asked for, written by nobody, and arriving from campaigns people used to trust.

Organizers should be more worried about this than anybody.

If you point the AI Big Tech has developed haphazardly at political work, inside the systemic forces of the day, and it will do three things very efficiently:

It replaces relationships.

It replaces human and community voice.

It replaces participation.

Those are not just three features of good organizing. They are the three things organizing is made of. And they were already in a rapid state of decline in our political culture, even before generative AI arrived on the scene. 

In the third chapter I wrote about transactional contact replacing the relationships that actually produce power. Optimization got there first, at human speed, one A/B test at a time. Big Tech’s AI does the same thing at machine speed and reports it as scale.

In the fifth I wrote about mass-produced politics that persuades nobody, and about what is actually scarce, which is the sense that a real person who knows something meant this. Big Tech’s AI is a machine for producing the opposite. Political slop, strategically hollow, indistinguishable from everything else in the inbox, erasing the specificity of voice and place.

Then participation. In the first chapter I argued that politics had become something you consume rather than something you do. In the eleventh, that the work we hand people is chosen precisely because nothing depends on it. There was not a great deal of participation left to remove.

AI is not designed with ill intent. It is predictive. One of its key potentials is to accelerate the processes that might take human craft longer to accomplish. The market is anti-organizing by design, and AI will accelerate the power of the commercial market. 

This isn’t an argument against the technology. It is an argument about what it is transforming and how to utilize technology now.

Human + Machine describes three waves of business transformation across the last century and a half, each one set off by a technology that allowed a rethinking of process rather than a speeding up of the old one.

Standardization came first, with the industrial revolution. Ford, the assembly line, work broken into identical steps so each one could be measured and improved.

Automation came second, with computing. Arriving in the 1970s, peaking in the nineties, databases and personal computers taking over the back office. The advantage was straightforward: replace human activity with machine activity and pay for it once.

The third wave is adaptive process, and it is where the actual potential of this technology sits. Tools that are flexible, fast, and responsive to the person using them at the moment they are using them.

Most of the AI integration happening in our sector right now is set to fail from the beginning, and the technology is not the reason. The people directing it are still reasoning in second-wave logic. They have not changed their paradigm. They have changed vendors.

The difference between the second wave and the third can be described using arithmetic.

Automation works like addition. Technology plus human equals outcome. Your technology is worth 5, your people are worth 5, your outcome is 10. The technology improves and it is worth 8. Under this logic you can cut your people to 2 and still get 10. Same result, fewer salaries. Inside that model it works, which is exactly why the instinct is so hard to argue anybody out of.

Adaptive collaboration works like multiplication. Technology times human equals outcome. Same starting point, technology at 5 and people at 5, and now your outcome is 25 rather than 10. The relationship between them produces more than either one does alone.

Run the old playbook on the new equation. The technology improves to 8, so you cut your people to 2, and your outcome is 16. You bought better technology and ended up worse off than you started, because you cut the thing that was doing the multiplying.

Keep the people at 5 and let the better technology amplify them, and it is 40.

Two organizations can make the same investment, with the same tools, in the same month, and one of them ends at 16 while the other ends at 40. They’re using the same software. The results speak to the value assigned to humans and technology.

This keeps going wrong because of a category error.

A general purpose technology is not a better tool. The steam engine and electricity did not make existing work cheaper. They made different work possible, and the people who understood that first took the returns.

Electricity was more dangerous at first, not less. A great many people were electrocuted. Houses burned. The early years were worse in measurable ways than the era of lamps and candles had been. And then, eventually, there were far fewer house fires than there had ever been with an open flame in every room.

Nobody's careful personal habits made the grid safe. Voltage standards did. Insulation codes. Inspection. Liability rules that put the cost of failure on the people who built the thing.

Electricity is a system, so the answer had to be systemic. This is a system too, which is why the useful fights here are the boring ones. Alternatives that actually exist, so that the ethical version is available to be chosen. Accountability for what these companies produce rather than for what they promise. Standards written in rooms our sector is actually in.

A third wave makes one thing possible that no second-wave technology ever could. The process itself can be built into the tool.

What gets built in is the thinking around the work. The questions somebody asks before they start. The order they ask them in. The moment an experienced person stops a workflow and informs you the process your taking won’t produce the desired outcome. Work has always required a person who knows the work, sitting with somebody who is doing it, at the moment they are doing it.

Chapter thirteen called that the one input that never scaled. It is why the cycle stayed shut, and tomorrow I will get into what we encoded and how.

It’s funny to frame it this way, but treating this as a faster way to send emails is like buying a power station to run a slightly larger candle factory.

Domain expertise matters more in this age, not less, which should reassure anybody in our sector currently worried about their job.

When I use an image generator my images come out cliché, generic, immediately recognizable as machine-made. The model is fine. I cannot name the style I want, or reference the period or the movement, or say what to avoid, and I have no criteria for judging the third iteration against the first. My designers can do every one of those things, and given the same tool they produce something not remotely comparable, because the tool is amplifying something they have and I do not.

That is what specialization is, and it does not become useless when the machine gets good. It becomes the thing that decides what the machine produces. The designer becomes a strategic designer. The organizer becomes a strategic organizer. Their judgment is worth more than it was, because it now directs a great deal more output.

It also means the tools that result in the best results, won’t be the smoothest experience for users. 

Most commercial AI is designed for maximum adoption through minimum friction. Say the word and the machine obliges. That makes the results worse.

A joint study by MIT Sloan and Accenture found that adding moderate, targeted friction to AI workflows, in the form of prompts that pushed people to scrutinize what they had been handed, improved accuracy significantly without meaningfully slowing the work down. Renée Richardson Gosline, who led that research, calls it positive friction, and her finding on anchoring is the one to sit up for. When people are handed AI-generated content, between 60 and 80 percent of their final work ends up mirroring what the machine suggested.

Once you have seen it, you cannot unsee its potential.

The tools worth having are the ones that know when to push back, when to ask a clarifying question, and when to say are you sure that is what you actually want here. Good friction hands the work back to the human at the points where the machine should not be deciding. You cannot design it without knowing the work well enough to know where its blind spots are, which is the whole reason this cannot be built from outside our sector by people whose domain is software.

Collaboration with this technology is unlike collaboration with any tool that came before it. With each advance it mimics humanness more closely, and our relationship to it starts to feel less like operating something and more like talking to someone. Close collaboration with a machine that is not guided by human values carries an influence, and as we work alongside it we have to stay aware of the way it changes us. These are the big questions we will need to keep asking as we learn to live with this technology. Just like the industrial revolution or the automation of the prior decades, this will change us.

My team has ended up with one principle underneath everything we build.

Elevation over automation.

The logic of automation says take what people do, have the machine do it instead, and pay fewer people. The logic of elevation says take what people know, which is their expertise and their strategic judgment and their understanding of their own communities, and amplify it, so that more of them can create at a higher level, more independently, and more in line with what the group is actually trying to do.

Automation replaces the chef with a recipe generator. Elevation gives every cook in the kitchen the instincts of the best chef on the team, and keeps the chef in the room, directing and curating and evolving the craft.

The third wave is here whether our sector engages with it or not. What is still open is whether we use it to manufacture junk politics faster and cheaper than we ever could before, or to make democracy work better than it currently does.

I know which side I am on. We will lose this one by default if we spend the next two years arguing about whether to use this technology at all.