What Writing a Constitution Taught Us About AI Governance

Rules People Help Write Are the Rules They Follow

Rules People Help Write Are the Rules They Follow

The first workshop happened in an office, and the room did not believe in it. Charles Hoskinson, co-founder of Cardano, remembers the mood as deep skepticism. As he recalled on the Facilitation Lab podcast, people expected the effort to descend into chaos, get voted down, and get “bike shedded” into paralysis. The goal was a constitution for a global blockchain ecosystem with no executive branch, no incumbent voting system, and no central authority to impose one. Thousands of people who had never met would have to write the rules they would live under. The skeptics had a point. Most organizations cannot get a department to agree on a meeting policy. It worked anyway. The Cardano constitution was approved by community delegates, then ratified on-chain, and it now governs the ecosystem. We were in the room for much of it. Voltage Control certified facilitators helped design the process and led workshops from the community stage through the Constitutional Convention. What we learned there is the most useful thing we know about writing an AI governance policy. Rules written in a back room get routed around. Rules people help write get followed. And when trust is high, you need far less bureaucracy to keep everyone safe.

The Problem Cardano Had Is the Problem You Have

Cardano’s challenge looks exotic, but its shape is familiar. Hoskinson described it as a bootstrap problem: “how do you build consensus and consent when you have no incumbent decision system.” That is exactly where most companies sit with AI. There is no inherited rulebook for agents, copilots, or model access. The existing policies were written for a world where software did what it was told and people did the judging. Someone now has to write new rules, and whoever writes them will shape how hundreds or thousands of people work. The default move is to hand the job to legal, security, and IT. They draft a company AI policy, circulate it for a short comment window, and publish it. The document is usually competent. It is also usually ignored, because the people it governs had no hand in it and cannot see their work in it. Cardano could not take that shortcut. In a decentralized system, a constitution nobody accepts is not a constitution. That constraint forced a different method, and the method is the lesson.

What the Process Actually Looked Like

The community stage ran on local workshops. Intersect, the Cardano member organization that coordinated the effort, reports that 64 workshops ran from early August to mid-October 2024, across 51 countries on six continents, with more than 1,400 participants. Each workshop was a day-long session where community members worked through a small set of specific questions about the draft text, rather than reacting to the whole document at once. The Civics Committee managed which questions went where and kept that distribution flexible, so topics that proved contentious got more attention as the weeks went on. Facilitators were paired with workshops through our global network of certified facilitators and alumni, and where possible we sent facilitators who spoke the participants’ native language. Each workshop elected a delegate to carry its perspective forward. While the community workshops were still running, remote synthesis workshops brought elected delegates together to review the data and feedback flowing in, which let hosts adjust agendas mid-process. No two rooms behaved the same way. Caterina Rodriguez, who facilitated South American workshops in Spanish, described it this way: “There were groups that wanted to move to the process methodically. There were groups that wanted to take no breaks. Then there were groups that were like, we want a full-blown debate. So it was about being adaptable to what was emerging in the space.” Then came the convention. From December 4 to 6, 2024, delegates met simultaneously in Buenos Aires and Nairobi, connected by video link, with about 450 participants and 63 elected delegates representing 51 countries. On the final day, 95 percent of delegates voted to approve the draft. The delegate-approved text went on-chain for a vote on January 30, 2025\. It needed 75 percent approval from delegated representatives and unanimous approval from the interim Constitutional Council. It cleared both and was enacted on February 23, 2025.

Why It Held: People Own What They Build

Hoskinson does not credit the document’s brilliance, and is candid that the first version is imperfect. The credit goes to the process. What changed the room, in Hoskinson’s telling, was a shift in what people thought they were playing for. Once participants understood that “it’s a long game and it’s not about winning every round of the game, but rather it’s about participating and playing the game and learning and growing from the game,” he said, “they got substantially more collaborative. The learned helplessness went away and the cynicism disappeared almost overnight.” The framing of the first draft is worth sitting with. “The very fact we came together and wrote a Constitution, even if it’s not a perfect Constitution, is a monumental achievement because people from 50 countries had to come together and get something done.” Then: “So the very first one is really a minimum viable product and it’s a trust building product.” A trust building product. That is a different goal than most AI policy committees set for themselves. They aim for a complete, airtight document. Cardano aimed for a document a global community could stand behind, then planned to improve it. The ownership effect is not mysterious. Hoskinson described what happens after people build something together: “once people get over that hostility and they start working together, then they actually come up with interesting things and they enjoy working together so you don’t have to twist their arms and force them.” You don’t have to twist their arms. That is the outcome every AI policy owner wants and almost none of them get.

The Price of Distrust

The most practical idea from the Cardano work is about the cost of distrust. Hoskinson explains it with a land sale. Two neighbors who trust each other hit a snag in the paperwork, shake hands, and fix it. Two neighbors who distrust each other hit the same snag and end up with lawyers, years of delay, and enormous cost. Same facts, same outcome, wildly different price. “The only delta between those two realities is trust.” Distrust does not change the outcome. It changes what it costs to get there. Every extra signature, review, and escalation exists to de-risk a relationship that is missing trust. Read that with your AI governance policy in mind. Approval queues, exception forms, and blanket tool bans are all de-risking devices. They exist because leadership does not trust how people will use AI, and often people do not trust leadership’s reasons for the rules. Each side’s distrust justifies the other’s. A participatory process breaks that loop. When the people who use AI every day help write the rules, two things happen. Leaders learn how the work actually gets done, so the rules fit reality. Employees learn why each constraint exists, so the rules stop feeling arbitrary. Trust goes up on both sides, and you can afford fewer gates. This is the case we made in Govern AI Like You Hire: clear, trusted governance is what lets you say yes quickly.

How to Write an AI Governance Policy People Follow

You are not convening delegates from 50 countries. You do not need to. The structure that worked for Cardano scales down cleanly to a company AI policy, and every piece of it is something your organization can do this quarter. Start with a draft, not a blank page. Cardano’s community did not write from nothing. They reacted to a working text. Your legal, security, and IT leaders should still write version one. The difference is that it is labeled a draft and treated as one, with a real commitment to change it. Then take it to the people, in their own rooms. Cardano ran local workshops rather than one global town hall, because people speak more freely among peers and in their own language. The corporate translation is sessions by function or team, where an engineer, a marketer, and a claims processor each examine the policy through the lens of their actual work. Ask specific questions, not “any feedback?” Cardano’s workshops worked through a handful of targeted questions about the text, and organizers moved questions to where they were needed. Ask yours the same way. Which data would you need to put into an AI tool to do your job well? Where would this rule stop you from doing something useful and safe? What would you do if the approved tool could not handle the task? Specific questions produce usable answers. Open comment windows produce silence or complaint. Have each session choose a representative. Cardano’s workshops elected delegates, and synthesis sessions brought them together while workshops were still underway. A small AI policy working group made of people chosen by their peers, not appointed by leadership, carries far more legitimacy than a committee of directors. Bring them together to review what the sessions surfaced and revise the draft in the open. Make the approval visible. Cardano set its thresholds in advance and published the results. Inside a company, that might mean the working group signs off on the revised policy before leadership adopts it, and everyone can see what changed and why. People accept rules more readily when they can trace their input into the final text. Finally, ship it as version one. Hoskinson’s “minimum viable product” framing matters here. AI capability changes faster than any policy cycle. Commit to a revision cadence, quarterly or twice a year, and run the same participatory loop each time. That last step is not optional. Hoskinson described a force the Cardano community calls relationship entropy: without regular contact, people in a decentralized system grow suspicious of each other and stop extending the benefit of the doubt. “Workshops pull people together and they restore the relationships,” Hoskinson said, “so you have to do that on a recurring and regular basis.” An AI policy written once and left alone decays the same way.

Facilitation Is the Hidden Infrastructure

None of this happens by convening people and hoping. Hoskinson called the Cardano process “one of the most challenging things I ever did in my lifetime” and said a large share of it was education, teaching people collaboration, iteration, and negotiation skills so disagreement stayed productive. That is the honest cost of participatory governance. A poorly run listening session does more damage than no session at all, because it teaches people their input is theater. The skeptics’ prediction, chaos and endless bike shedding, is exactly what happens when a large group deliberates without structure. The structure is the facilitation. Specific questions. Neutral hosts who can hold a room through real disagreement. A synthesis method that turns hundreds of comments into a revised draft people recognize as theirs. Representatives who carry context between rooms. These are learnable skills, and they are the difference between collaborative governance and a long, frustrating meeting. The friction in AI adoption has moved. Execution is getting cheap. Trust, alignment, and governance are where organizations now stall, which is the shift we call the New Friction. Participatory governance is not a soft add-on to that problem. It is the most direct way through it, and as we argued in Involve Before You Mandate, involvement has to come before the rule, not after it.

The Rules You Write Together

Cardano’s skeptics were not wrong about the difficulty. They were wrong about what would make it work. They assumed the hard part was the document. The hard part was the trust, and the process built it. Your AI governance policy faces the same test. You can write it in a back room, publish it, and spend the next year enforcing rules people quietly route around. Or you can write it with the people who must follow it and spend that year improving rules they already own. The first path feels faster. It is not. Every rule people do not believe in needs a gate, a review, and an audit behind it. Every rule they helped write mostly enforces itself. If you are about to write or rewrite your company’s AI policy, start by deciding who is in the room. Want help designing a participatory process your people will actually trust? Talk with us. Helping groups write the rules they will live under is work we have done at a scale few organizations ever attempt.