Power concentrates.
Whoever controls the systems gains leverage over everyone else, and needs very few people’s cooperation to use it.
About us
Why we study the concentration of AI power, and what we are doing about it.
About IDAI
The Institute for Decentralized AI is a nonprofit research institute. We study how AI concentrates power, how this can be countered, and what decentralized alternatives need to compete. This includes secure protocols for AI agents, mechanisms that hold up against strategic and adversarial behavior, and evidence that plural systems can be safe. We work between Oxford and Stanford, bringing together researchers from AI safety, security, and economics.
The problem with centralized AI
AI pushes the world toward one where a few actors control the systems everyone depends on. The effect comes from feedback loops: a lead brings more users, data, revenue, and compute, which bring a bigger lead. AI makes large hierarchies cheaper to run than open networks. Open systems that cannot defend themselves lose to walled gardens. And people come to depend on the systems they learn from, confide in, and get attached to.
Failures become systemic.
One outage, one compromised model, or one bad update reaches everyone at once. When everyone relies on the same few models, they share blind spots and values: errors stop canceling out and start compounding.
Oversight gets brittle.
A few overseers can be captured, pressured, or simply wrong, and nobody is positioned to check them.
Exit disappears.
Once people’s work, data, and relationships live inside one system, leaving costs too much, for individuals and institutions alike.
Theory of change
Incentives matter more than appeals.
Advocacy has its place, but the most effective path against concentration is to weaken the incentives that drive it and strengthen the ones that favor decentralization.
Decentralized systems can be made as capable as centralized ones.
We focus on scaling decentralized systems and organizations, so real alternatives to a handful of dominant providers can emerge. Acting quickly matters: setting standards now can have powerful downstream effects.
Decentralized systems can be made safe at scale.
Decentralization introduces its own security and safety risks: collusion, cascading failures, and attacks that spread across many agents at once. Solving them early makes decentralized systems more attractive, and keeps them from becoming a systemic risk of their own as they grow.
The people who can solve this do not yet work together.
AI safety, security, economics, and mechanism design each understand part of the problem but rarely talk. Most funders and policymakers do not yet see centralization as a safety problem. Field-building closes the first gap; making the case closes the second.
What we do
Everything we do should make the open option more capable, more secure, or better understood.
Research
We test whether decentralized systems can be made as capable and as safe as centralized ones. Most of this is technical work others can adopt without asking us: verified protocols and mechanisms, attacks that force fixes, benchmarks that define what “safe” means when agents interact, and tools that spread capability instead of concentrating it. We also study how people come to depend on AI.
Field-building
We build connections across AI safety, AI security, economics, mechanism design, decentralized systems, and sociology. We bring these communities together through workshops, a shared map of the field, and visiting positions.
Shaping the defaults
We contribute to the standards and protocols agentic systems will run on. We make the case to researchers, funders, and policymakers that centralization is a safety problem, and decentralization can be a safety mechanism.