This article addresses learning in desktop virtual worlds where students role play for professional education. When students role play in such virtual worlds, they can learn some knowledge and skills that are useful in the physical world. However, existing learning theories do not provide a plausible explanation of how performing non-verbal virtual-world actions (e.g. performing a virtual chest examination in a virtual hospital) can lead to the learning of the physical world equivalent. Some theories are particularly implausible because they claim that students learn to perform physical world actions by acting on the virtual world in an embodied way. This is implausible because learning requires a high degree of correspondence between the learning performance and the target performance, and there is insufficient physical correspondence between the performance of a virtual-world action where students click on a mouse to make the avatar take actions and the physical-world equivalent where students perform the action with their own body. In this article, we use Austin's speech act theory to provide a more plausible theory of learning in virtual worlds. We show how non-verbal virtual-world actions performed by avatars can function as performatives and as performatives, they can correspond sufficiently to physical world actions to explain how performing non-verbal virtual-world actions can lead to physical world learning.