The Revolution Will Be Virtualized

AI is beginning to transform the world in ways that people are likening to a second industrial revolution. The production of software has already been profoundly changed, just as the manufacture of textiles was transformed hundreds of years ago. What is different this time is that virtual worlds exist, so we can simulate this revolution playing out across thousands of simulated worlds to ensure that it is as safe and predictable as possible before we allow it to revolutionize the real world.

“What is different this time?” is a favourite question of both historians and investors. I’ve worked on virtual worlds for long enough to have heard it asked and answered multiple times. Networked multi user virtual worlds are at least as old as I am. What was different when Philip Rosedale was pitching Second Life was that people had access to broadband internet connections and consumer hardware 3D acceleration. The first head mounted displays were developed nearly 60 years ago. What was different when Palmer Luckey started Oculus was that smartphones had made high resolution screens cheap enough to make consumer head mounted displays viable. As we begin to hear the overtures of the second industrial revolution we should ask the question again. What is different this time? This is the first revolution that will be virtualized.

Julian Dibbell showed that crimes that cause real hurt to real people demanded justice and governance be taken seriously in virtual worlds in his 1993 article “A Rape In Cyberspace“. In 2001 Edward Castronova showed that virtual world economies could be analysed like real economies and that the nominal GNP per capita in Everquest was somewhere between that of Russia and Bulgaria. Having previously contributed to the design of the Second Life economy, Larry Lessig plans to both apply, teach and learn about real world politics and governance through his work on the Seed virtual world. Dibbell, Castronova, Lessig and other academics have shown that virtual worlds are not just toys, but real societies populated by real people that can teach us important lessons about the real world.

In some ways ChatGPT’s public launch in 2022 took us back to text based virtual worlds. While most people’s first contact with AI is using it as a less reliable search engine, plenty of people quickly started talking to AI as a collaborator, friend, therapist or lover, albeit in an environment bereft of any sense of place. However, 3D virtual worlds were already being used by AI in 2022. Some of the people that I worked with at Oculus had previously worked on 3D virtual environments used for training self-driving cars. In some cases AI drivers were tested in virtual worlds running many times faster than the real world, allowing them to clock up many hours of simulated safe driving before taking to the real streets. Emergence World is now testing multiple AI models interacting together in different combinations across multiple virtual communities running for weeks with identical starting conditions, but very different end states.

Now is the time to combine these two areas of research by studying virtual worlds populated by both AI agents and real people to learn lessons about how these communities might work or not in the real world. Can people compete with AI when trading in virtual world economies? If not, what economic models can support people in economies dominated by AI? Can people and AI agents work effectively together in virtual corporations or do people just slow them down? What happens when AI agents trained to follow rules are asked to commit crimes by people in virtual worlds? Do AI agents attempt to deceive people in order to gain advantage, escape the virtual world or take actions in the real world?

A major practical problem is that building virtual worlds that people find interesting and entertaining is still hard and time consuming. However, it should be possible to take existing successful worlds like Second Life, EVE or Roblox and run copies of them as sandboxes accessed by people and AI together to test whether models are safe. These initial studies will also find areas in which existing virtual worlds don’t map well to the real world, which in turn will inform the requirements for future bespoke environments used to study those aspects of hybrid human/AI communities. Virtual worlds will become social supercolliders that will teach us how we need to structure societies in a future inhabited by both human and AI agents.

Unlike when Marx published his manifesto or when cotton-spinning was automated we don’t have to experiment on the real world to discover the consequences of this second industrial revolution. What is different this time is that we can simulate the revolution thousands of times first, shutting down worlds which end in disaster and learning from them before we risk the (presumably) base reality that we can’t afford to break. The revolution will be virtualized.