Impossible Dances is a choreographic research project that treats artificial intelligence not as a tool for capturing movement, nor for generating it, but as a generator of conditions the body must navigate. Developed by Ana Sánchez-Colbergit was tested in a two-week laboratory with Ballers de San Juan with twelve ballet-trained dancers. What makes the work new is its third orientation: most AI dance research either archives movement or generates it, while Impossible Dances does neither. The body never enters the system, nothing is recorded or converted into data, and the dancer's response — at every point between full compliance and complete departure — becomes material. Built on a choreographic spatial logic and on browser-native XR that runs without motion capture or institutional infrastructure, the project proposes a non-extractive way for choreography and AI to meet — conditioning rather than capture.