What happens in a system of a billion AI agents?

Human systems shouldn’t be a living experiment - but we’re not prepared for a world where autonomous AIs are as numerous as humans.

The problem

AIs already inhabit billion-scale systems with humans,¹ and their prominence continues to grow in economies, social networks, and workplaces.² However, AI safety has traditionally focused on individuals and small groups of agents, leaving large-scale systems of many agents relatively unexplored.³

Factors like communication cost and clashing preferences means, “more is different,”⁴ and single-agent safety research doesn’t carry over well to big systems.^{5 6} Major labs often focus on small-systems work or conceptual investigations of large systems, while there is limited safety work on real large systems, or connecting established small systems safety to large systems settings.

Human infrastructure was not designed to handle numerous fast, autonomous, flexible AIs. Social media systems have failed to manage AI-driven misinformation, contributing to polarisation and erosion of democratic institutions.⁷ Simpler fast, autonomous algorithms have led to wealth-erasing financial crashes⁸ and accidental, inefficient, illegal price coordination on digital marketplaces.⁹ These system failures are early examples of how destabilising volatility and unwanted coordination could impact billion-scale AI and hybrid systems.

Furthermore, several lines of research have highlighted the potential for accidental emergence of superintelligence from large multi-agent systems,¹⁰ [^bostrom] ³ carrying potentially catastrophic risks.

What can be done

As systems scale to a billion agents, the problem shifts from the individual to the collective. New theory, methods, and experiments are needed across:

• designing safe systems for many AI
• efficiently simulating large-scale AI and hybrid systems
• efficient monitoring of systems where central monitoring of every agent would be too expensive
• federated interpretability - safety monitoring which uses the agent’s host device to verify its behaviour is safe, preserving privacy and distributing the cost of safety across the system
• understanding and predicting system behaviour from network properties, without looking at any individual agent
• controlling systems of AIs by moderating individual or collective behaviour.

Out-of-scope for this problem are single-agent and small-systems research not in direct service of large-systems research.

Getting involved

Gigascale Laboratories is a research collective founded in 2026 to investigate this research agenda.

The best way to learn about big AI systems is to bridge the gap between AI safety and existing systems sciences, so we’re looking for researchers across:

• AI safety
• network science
• complex systems science
• economics
• physics
• epidemiology
• and other systems disciplines,

as well as software engineers interested in scaling research into useful tooling.

We’re especially enthusiastic to invite students and early-career researchers to join us. If you’re interested, don’t hesitate to get in touch - we’re open to applications from curious, self-driven people from all backgrounds and qualifications.

Positions are currently voluntary as we are pre-funding. We support researchers with grant-writing and provide compute and administrative support.

Please get in touch by filling out this form: link.

References