Virtual AI agents tested in social good experiment

Nonprofit organisation Sage Future has launched an unusual initiative that puts AI agents to work for philanthropy.

In a recent experiment backed by Open Philanthropy, four AI models, including OpenAI’s GPT-4o and two of Anthropic’s Claude Sonnet models, were tasked with raising money for a charity of their choice. Within a week, they collected $257 for Helen Keller International, which supports global health efforts.

The AI agents were given a virtual workspace where they could browse the internet, send emails, and create documents. They collaborated through group chats and even launched a social media account to promote their campaign.

Though most donations came from human spectators observing the experiment, the exercise revealed the surprising resourcefulness of these AI tools. one Claude model even generated profile pictures using ChatGPT and let viewers vote on their favourite.

Despite occasional missteps, including agents pausing for no reason or becoming distracted by online games, the experiment offered insights into the emerging capabilities of autonomous systems.

Sage’s director, Adam Binksmith, sees this as just the beginning, with future plans to introduce conflicting agent goals, saboteurs, and larger oversight systems to stress-test AI coordination and ethics.

For more information on these topics, visit diplomacy.edu.