You only need to know how to type

A few months ago I wanted to know which model I should be using for a particular job, how the pricing actually worked, and which features I had on the plan I was already paying for. I asked ChatGPT. Then I asked Claude. Both of them went away and searched the internet, came back with an answer. The answer was out of date. The model didn’t know what features I could access or what plan I was on.  

These are systems built ‘to know everything’. (After all - isn’t that the purpose that AGI - Artificial General Intelligence is meant for?) Clearly they do not know themselves, or at least nobody built them to explain themselves to the people using them.

A system carrying no model of its own process cannot walk anybody through what it’s doing, including the process of how to use it. It holds information. That is why it opens on an empty box and waits for someone to know what to ask.

At the beginning of the generative AI boom there was a job title for the problem this creates. Prompt engineer. It was a high paying post, and it got treated as a sign of how powerful the technology had become rather than a sign of how badly it had been built.

That job title has mostly gone now. Most people who use these tools can write a serviceable prompt, and the industry took that for the access problem solving itself.

But if you ask somebody who does not work in this world where they would start with AI a lot of them cannot tell you. They think it is coding. They think it is a development skill, something adjacent to building software, something you would want to learn properly before touching. A lot of them use it to ask for love and relationship advice. That's fair enough.

The people who are fluent with these tools, the ones who know how to use them well, did not get there by acquiring a vocabulary. They got there by opening something they did not understand, typing something bad into it, looking at what came back, and going again. Fluency is a habit, picked up by being allowed to sign up and try. They learn when to call out it’s responses, like when it sounds like  a high schooler faking the assignment they never did the reading on- fancy language and little substance. When to trust it and when to check it. Why prepping data and checking work, can be as laborious as writing it from scratch, if you don’t want to become a meat proxy. (I’ve made mistakes myself plenty of times, when changing just one last thing and copy-pasting trying to get a last email out, before closing shop for the day). 

A great many of the people this book is about have never been given even that chance. For example: the volunteer handed work that required no trust, the leader carrying responsibility with no decision space, the member whose contribution was edited into meaninglessness. Nobody develops an appetite for experimenting in the open by being handed envelopes, as we are currently treating most supporter volunteers. 

The empty prompt box makes it worse than it needs to be, because it assumes four things before you have typed a character. First, that you know what to ask for. Second that you know what the machine is capable of. Third, that you know what good looks like when it arrives. And lastly, that you know how to say it in the register the thing answers well.

The machine never tells you that you asked the wrong question. It answers the one you asked, fluently, at length, in a confident voice. Nothing in what comes back indicates that a better question would have produced something twice as good. The person who asked badly has no way of finding out. The gap between them and the person who asked well stays invisible to both of them.

An interface like that is a mirror of prior advantage. It rewards the people who already had permission to experiment, who already had somebody to compare notes with, who already knew what a good brief looked like because they had commissioned quality AI work before.

That is one of the real costs of this technology and we don’t have enough people tracking it. We argue about what generative AI does to democracy, about the environmental damage behind it, about what happens to our collective political imagination once a machine has proposed the options. This issue belongs on that list. A tool that rewards the already-capable, dropped into a sector already sorted by who had access to training and time and money, accelerates a power differential that was indefensible before the automation. Of the three things this technology strips out of organizing, participation is the one that goes quietly, and the empty box is how it goes.

We built the opposite, starting with the part that sounds trivial and is not. The system knows its own instructions. Ask it how to do something inside it and it tells you, because we put best practices  in rather than leaving it to be searched for on the internet by a machine that would get it wrong. There are videos for the basics, videos for the advanced work, and videos for the things people wouldn’t think to try on their own. The platform is built to help you use the platform.

Underneath these decisions sits an ethos that we built around. You do not need to know AI. You do not need to know politics. You only need to know how to type. The structure of the process is built into the system rather than expected from the person, so the questions it would have been your job to know to ask are the questions it asks you. That is the same design decision I wrote about two days ago, turned around: friction pushes back on somebody who knows too much, guidance asks the first question of somebody who does not yet know enough, and both are the tool holding a process instead of performing a task.

To test this out, I gave what we were building to my mother.

She is about to turn eighty. She is creative and wanted to study art when she was seventeen and never got the chance to study it, which is a story for another time about how much less was on offer to women then. She never got on with computers. Most of our FaceTime calls I spend looking at her coffee mug, her elbow, or the top of her forehead. She is savvy about a great many things in life, and technology has never been one of them.

She could not make something in Canva. Canva assumes a design literacy she was never given any route to acquiring.

She asked to try the thing her son had been spending all his time working on and would not stop talking about. She made a poster for a pretend LGBTQ fundraiser, as a trial, to see what would happen. It was good. (I will go into more detail about why it was good and show it to you in an upcoming chapter).

This was her first time trying anything AI - including creating an image. 

I showed her nothing. No walkthrough, no instructions, no documents. She did not even watch the videos.

She is my mother and she wanted me to have built something worth stress it takes to build something new, there’s bias there for sure. But the part she could not have supplied out of loyalty is how she actually used the technology. On her very first try.

She wanted to make things at seventeen and did not get to study it. Then the computer age arrived and shut the same door a second time, quietly, by assuming a set of skills she had no way to pick up. AI built for someone to simply type did not open a door that had recently closed. It opened one that had been shut her whole life.

There are three places to supply a tool like this. Nearly everything being built focuses on just the first two..

Point it at staff and you have made the bottleneck faster. Staff were never the slow part. Everything routes through them, and a quicker funnel is still a funnel.

Point it at supporters directly and you have removed the organization from the picture. Every person alone with a machine, producing their own content, accountable to nobody, connected to nobody. That gives you junk politics with better graphics at higher volume, and it dissolves the very thing that made any of it powerful in the first place.

Point it at the organizer, the person whose actual job is bringing other people together, and every capability you hand over gets spent on relationships. Mapping who holds power locally. Working out who to invite and how to ask them. Planning the meeting. Making the poster for the meeting. The output is a group.

Organizing's strength is the connection between people, and the ability of those people to stand together through difficulty and over time. The tools should be about bringing people together and elevating what somebody needs in order to do that, rather than replacing any component of it.

Building for organizers rather than for staff is the more difficult choice from a commercial angle and we know it. As an organization, we know we’ve picked a harder road, but one that will lead to more systemic change.. Staff budgets are where the money is. Software sold to a communications department has a purchase order behind it. Software handed to a volunteer group leader has a theory of change behind it, which is a much harder thing to invoice.

Every technological tool answers the question of who gets to do the work, whether or not anybody in the room thought to ask it. Leave it undecided and the tool decides for you, the way the market decides everything, in favor of whoever already has the most.