Aviv Ovadya: Democratic Infrastructure for the Age of AI | Intelligence at the Frontier 2026

I know, I am I am as mentioned the CEO of the AI and Democracy Foundations and we work to accelerate governance innovation and particularly to address a wide variety of risks impacting democracy at all scales. We're thinking about power and sensor structures around the world being reshaped by AI advances and ultimately our goal is for the love of democracy to play out. We realized on innovative democratic processes to use governance around the world across our team and have piloted democratic systems innovations with Frontier AI in the area. And we can think about applying these democratic innovations directly to AI at the AI system level. Uh the the real world and um any better sort of governance regulatory level and and even actually across like us.

So you're thinking about multiple organizations, industry bodies, international agreements, construction of of actual AI agents and so on. But like practically what does this look like? So primarily we focus on accelerating AI and what we call innovative democratic systems and it's a little hard part to do this in cellways. Heard from many people here already, but the precise shape of the future, sorry, is like pretty uncertain, right? We know there's some risks and we know our concentrations and like step away for a year or so.

I call these coordination challenges is pressure toward disempowerment a little bit space. There you go. Yeah, I think that's probably Um yeah, I saw. Um Uh and and you know, our efficient brain is doing great all the time. Uh maybe also know uh that there's some building blocks that are really helpful.

We've heard about maybe that here already. If we go a little deeper at some areas and help navigate those risks. So one of the things that we're thinking about is in addition I mean there's lots of detail implementations, but we also really want flexible building blocks that can be applied as needed. And we want to think about what are the properties that enable them to be used across so many possible future institutions that are directly testable and usable today. So in particular we focus as well as you'll see why on deliberative processes.

Like that's that's where most of our our our main bets are at. And and for simplicity in this talk I'm going to use deliberation and deliberate process here to mean this sort of pretty high level thing. It's it's a combination of a population microcosm and information microcosm in order to make high quality decisions. And just about everything that sort of people have talked about and or many of the many of the examples that I've heard already are sort of from that that arena um and not every favorite. And and there are varying levels of quality uh democraticness and answer to the output.

Uh like the just the how good that is. And so you know, concrete example that you and refers already for those who are here for the year we call our citizen assemblies. And I the basic idea here is pretty simple. You have a representative of the citizens, people um you know, they're the baby picked to be there. You know, they're they're just a few participants in some kind of sample mechanism.

And then in a judicial citizen assembly they may get a week or more of a dedicated compensated time of these spaced out over a month of weekends or two to dig into a issue of the solicitation. It's like it's actually scaffolded uh uh scaffolding workflows that enable them to do some representation they get with input from expert stakeholders in the AI systems now. That ends in a decision that that shows sort of where the common ground is and that might go to a decision maker or it might in some narrow set of cases so far be the decision maker. But across my gosh like this ecosystem of people like doing this, running these kinds of processes is like is a is an exponential. It's unfortunately the exponential is a little smaller than the one for AI but it is an exponential and and there've been now thousands of processes run like this around the world.

And just to make this a little visceral um this is the European um have a cycle. This is where members of parliament where where instead of members of parliament you use a random citizens um selected from across the EU uh you know, across every EU language and this is something that feeds into EU policy making several times a year. So they've got what they call citizen panels or the citizen assembly version um uh that that feeds into policy making three or so times a year. And like okay, why is this important? Why does this matter?

There's a like some really really nice properties here. Right? Deliberative processes are enormously easier to spin up and iteratively test than elections, wars, legislative bodies. Even when they're directly integrating and collaborating those established structures. It can be designed to appropriately deep expertise, diverse perspectives and as we we've heard also with they they can have a definite outcome graphs.

And most importantly they can eliminate any of the perverse incentives and adversarial attacks that you might worry about through random selection including things like you would like civil counts to some extent. Um And and so many of the other promising building blocks here also rely on many of the same underlying capabilities that you need for what we're what you know, defining the sort of broad arena of deliberative processes. Deliberation provides critical tools for addressing all of these challenges and core infrastructure for many of the necessary institutions. So if we can use this well there's so many things that that the ability to to um make decisions in wise ways through representative input. There's so many problems that you can you can actually move significantly toward solving.

It's like it's a new sort of action you can take on the world. And so okay, so what to say deliberation is not just goal of better democratic systems. You know, what is a better actually would mean? What is a high quality decision? So general challenge for democratic systems like AI systems is they also have what I would call a jagged frontier.

Um sir. Uh many of the same analogs apply. You have capability jumps that require more human computes. Or if you want something to be a certain level of quality you need more time for people. Um you can distill a complex process like a citizen assembly it might be a week, two weeks into something that might be only a few hours.

But you've got to be really careful to ensure like is your system actually still good enough? Are your outputs going to be good enough? Is is it going to be something that is actually wise at the end of the day? And and so one of the things that we've done is sort of a essentially a first principles analysis building on lots of practical experience from running processes like citizen assemblies and lighter weight deliberative processes and other kinds of mechanisms for democratic decision making. And developing a set of dimensions that sort of span the core of that practice of design.

I'm not going to walk through them directly here, but but like if you want to you know, dig deeper is the framework is part of a larger body of work and it's I see my position paper uh uh that we we put together on that. Um and particularly how it applies to AI. So these dimensions rather than just answering this being one that can be broken down into more measurable democratic capabilities, right? Like how how implementable are the outputs? How adaptable are they to different contexts?

We can assess each of those capability gaps. We can build a map of the critical research questions needed to address them. We can identify concrete products that need to be built that directly improve those capabilities. And we've sourced these from the democratic practitioners around the world and they tackle concrete process bottlenecks and quality challenges. So you're zooming back out building on all of this work we can now holistically assess capabilities for both an ecosystem as a whole of democratic um systems design and development.

And for particular democratic processes and systems. And so now we have a way of talking about sort of what better actually means. Well, what does it look like to improve on these um like commissions and to build these as good as possible. So we see there being four critical ingredients. So first we want to target the democratic capabilities directly, right?

We need uh we have like closed this it's like very concrete gaps that you know, you know, I had on the screen a moment ago. The frontier doesn't need to be as jagged as it currently is. Um the if you can sort of find the right research uh the right system design, the right inputs like enterprises there that enables again drop probability that was sort of what one of the things I kicked off at Metagov uh uh before I founded the end of the Open Philanthropy like that that's a core core infrastructure that we for human always develop with. And so part of that is like you know, not just directly vote with this, but if you can create well structured advanced market commitments and prizes for where for things that for um uh capability jumps on particular domains that is incredibly valuable and it can drive massive third party investment in the space. If we have the right enterprise grade infrastructure that actually handles all these really challenging compliance, I mean duplicate identity issues like all of these sort of underlying challenges, that helps enable the a thousand flowers to bloom effectively and the level of experimentation that you need if you really want to get the kind of iterative feedback loops that would be ideal.

There's so many like this is just like a small smattering from our our map of concrete fundable projects and components that needed to get built, needed to get built quickly and it it wish it was easy as just like asking an agent to do your thing. There's there's a lot of complexity and iteration of using these things in practice. But it's hard to even know if our capabilities are actually bringing more out measurement. And so right now you can think about sort of democratic system design. Like there is no science of it this system I mean there are pieces of arts of political science, arts of other fields, but you can think about it as sort of like pre-benchmark, pre-ImageNet, pre- like there's no like RCT type thing.

Like none of that is is is is really viable in a in a significant sense. There there there little pockets throughout. And so we need good evals and benchmarks and data. Um if we can get better at simulating processes, we can test different approaches. Right now simulations are not really trustworthy.

But they could be maybe. Um we can evaluate where representative agents might be sufficient and where we really want humans in the driver's seat. And like there's lots of normative components to that, but there's also just like is it even good enough to make effective decisions or to represent someone at all. And we need to like ensure rapidly growing field doesn't repeat costly mistakes. Like I see this over and over again people falling into the same sort of dead ends, the graveyards of projects.

We also can't learn about doing. Right? If we're accelerating if we're going to have a accelerating innovation, you need to actually do iterations. You need to fund portfolio processes. You need to do lots of low risk high volume pilots that rapidly informs R&D.

And then you need these high signal sort of low volume processes that help ensure that this approach is is actually taken seriously and that it's institutionalized in a meaningful sense. And I guess the final piece in this is sort of where we're focusing the most is ecosystem wide acceleration and scaffolding. This is where you think about like any new paradigm and this is you know yes this has been used around the world in various level, you know, across governments as as I sort of illustrated, but this is still something especially from the US context within a number of other jurisdictional contexts, definitely within a corporate context, it's still pretty new. Like you know, I've helped you know, observed and and supported some work actually using these processes of public utilities. Um across the team, but but again this is still new and it can still be delegitimized as an approach if you if people mess up too badly.

And so there there's a need for standards. Um a high profile failure could significantly um slow adoption. And and so the so there there's a piece around standards. There's around um the the the way in which advocacy and storytelling can actually help drive adoption. Like what are the best ways to do that?

How can we accelerate those pieces? Um there's there's um levers with things like ballot measures that people have been exploring where you can essentially institutionalize these without going around existing power structures. And and then of course interoperability um is absolutely critical to reduce replication and sort of fragmentation. One look underlying piece so is that you need you need like the you need to be having actors who are able to identify allies and direct resources toward this as much as possible. That's sort of a lot of what we're we're doing sort of mapping out pieces of the space using tools like uh this this democratic capabilities map that that we're building out.

So assume that um I'm sorry this is Yeah, so I'm not going to go into into this in detail. Um but one of the things that we're particularly focused on is building out all those ingredients with an epistyle model for rapidly work toward each part of those this ecosystem. And there just isn't a lot of time if we want democracy to keep pace. We feel pretty confident there is enormous amounts of funds that could be deployed rapidly to make significant progress on each of these key ingredients. Like each one of those is like three to ten million dollar bucket and in many of those places we know already who are the people who are well positioned to do that.

And the entire democracy space and innovation space is relatively under resourced. It's extremely under resourced compared to many of the other other domains um that uh where you like across AI, crypto, etc. Anyway, you can right now see an early version of this democratic capabilities gap map is at democracybuilder.org. Um that that framework I mentioned part of larger body of work.

If you want to dig deeper there's so much more we could talk about about around like the actual work around adoption within sort of institutions relating to AI and the deliberative workflows that enable all of this to be effective. Happy to discuss that offline or in Q&A. And to to bring it all just together we believe that investment in better democratic systems can dramatically mitigate the key challenges posed by AI advances even under extreme uncertainty. There isn't a lot of time if we want democracy to keep pace. Right?

We feel confident that there is an opportunity to deploy enormous amounts of funds quickly to make progress on many of these key ingredients. We need to make democratic systems and processes better. Our current systems as we've seen even with recent events are not up to the task of the sort of the challenges that we're facing. We know what needs what needs to happen to to do that. And there's simply a matter of execution and resourcing.

And that's it.

[music]