Artificial hope

As the first demos of AI-generated games become playable, what's their purpose – and their cost?

- 

Two and half years after the arrival of ChatGPT kicked off the so-called AI revolution, the technology has been inserted with varying levels of grace and utility into every digital platform you can think of, from WhatsApp to eBay, with its stylistic tics becoming increasingly well-known even as the models improve. It remains a fixation of the Silicon Valley set, a disaster for environmentalists, and – according to OpenAI CEO Sam Altman – wildly unprofitable. But while the everyday applications of ChatGPT are widely understood, claims of its potential in game development remain rather less substantiated.

In this respect, Microsoft has led the charge of late, beginning with a full-press debut of a family of generative AI models in February. The first World and Human Action Model (WHAM), named Muse, was trained on seven years' worth of footage from Ninja Theory's otherwise very much neglected multiplayer shooter, Bleeding Edge. April saw the release of WHAMM, a much more efficient take on a similar approach, which was trained on a mere week's worth of Quake II footage and playable – sort of – in a browser.

These aren't the only Microsoft-owned games that have been used to train AI models, however. In November 2024, two companies, Descart and Etched, announced Oasis AI, which presented a gen-AI knockoff of Minecraft based on millions of hours' worth of footage of people playing the original game. The technology was not unauthorised by Microsoft or Mojang, but neither of the IP holders have made any official comment or taken action to date.

Questions of authorisation and legality aside, the experience of actually playing Oasis AI's recreation of Minecraft isn't all that different from the