Google DeepMind Funds Research into Risks of Millions of AI Agents Interacting

Google DeepMind is providing $10 million to fund research into the potential risks of millions of AI agents interacting.

Dr. Rohin Shah, head of AGI Safety and Alignment Research at Google DeepMind, stated that new types of risks could emerge as large numbers of AI agents capable of performing tasks without human help or following the instructions of other AIs become prevalent. Shah expressed concern about amplified versions of problems currently seen online, such as scams, prompt injection, and other cyberattacks.

This funding initiative aims to encourage research to begin outside of technology companies. Academia has the strength to conduct future-oriented research that industrial labs may not prioritize immediately. Alongside Google DeepMind, partners in this initiative include Schmidt Sciences, the UK government's ARIA, the non-profit research organization Cooperative AI foundation, and Google.org, Google's philanthropic arm.

Rohin Shah and James Fox believe the only way to understand what could happen when millions of AI agents interact is by running realistic simulations. They hope researchers will place AI agents into a 'sandbox' environment and closely observe their behavior. Studying single agents or small numbers of agents in isolation makes it difficult to predict complex future interactions.