Making AI more decentralized
Technologies that spread capability faster than a lead can compound, such as distillation, open models, and shared compute.
AI concentrates power.
We study why this happens and build safe, competitive alternatives.
Our approach ↗
About IDAI
We are a nonprofit research institute working between Oxford and Stanford. We study how AI concentrates power, how this can be countered, and what decentralized alternatives need to compete.
Our thesis and approach ↗Members and affiliates
Samuele Marro
University of Oxford
Emanuele La Malfa
University of Oxford
Brendan McCord
Cosmos Institute
Philip Torr
University of Oxford
Michael Wooldridge
University of Oxford
Tobin South
Stanford HAI
Adel Bibi
University of Oxford
Weifeng Tao
University of Oxford
Alasdair Paren
Chalmers University of Technology
Jialin Yu
University of Oxford
Francesco Pinto
Google DeepMind
Guohao Li
Eigent AI
Jesse Wright
University of Oxford
Jiaxin Pei
UT Austin / Stanford HAI
Tiziano Piccardi
Johns Hopkins University / Stanford HCI
Michele Lombardi
University of Bologna
X. "Angelo" Huang
ETH Zurich
Fellows and visitors
Projects
Research areas
Map of Decentralized AI ↗AI for decentralized systems
AI that makes markets, protocols, commons, and decentralized organizations as easy to coordinate as a single organization.
Multi-agent security
Verified inter-agent protocols, work on collusion and manipulation, and attacks and defenses that help open agent economies defend themselves.
Loyal agents
AI that acts on behalf of the people it serves, rather than its provider, and goes with them when they leave. We also study how people come to depend on AI.
Selected publications
All publications ↗Questions about IDAI
What is decentralized AI?
We define decentralized AI as any AI system where compute, data, governance, and oversight are distributed among multiple independent parties.
Why does it matter?
Centralized AI can fail in ways that are hard to recover from: power concentrates, failures go global, mistakes become correlated, and leaving becomes difficult. A plural ecosystem can help keep failures local and correctable.
Isn’t decentralized AI worse for safety?
It can be, if decentralization means anyone gets any capability with no oversight. That is not what we argue for. We want to distribute control, including ownership, oversight, and governance rights. We do not argue for handing out dangerous capabilities unconditionally. Instead, we support conducting research jointly into decentralization and safety, so that solving power concentration doesn't come at the expense of other risks.
Why not just regulate?
Regulation can help. But it moves slowly, stops at borders, and is shaped by the actors it is meant to constrain. Our work is meant to help whether or not regulation succeeds.
How can I work with you?
For collaborations, email collaborations@decentralized-ai.org.
Contact
Get in touch about research collaborations and the work of the institute.
Research collaborations
collaborations@decentralized-ai.orgGeneral enquiries
hello@decentralized-ai.orgSupported by
