Artificial Improvisation is the subfield of Artificial Intelligence devoted to designing musical agents capable of participating in collectively improvised musical performances with other agents (humans or artificial). It is well-known that the understanding of artificial agents is plagued by the opacity surrounding their inner processing, especially when it results from machine learning rather than human programming. This opacity questions the possibility of understanding those agents and their music. The central hypothesis behind the project “Inside Artificial Improvisation” is that the opacity of Artificial Improvisation may be reduced by adopting an ethnomusicological perspective on the music it produces and the socio-technological arrangements that enable it.

The project is structured around four working packages, each targeting a specific aspect of Artificial Improvisation associated with a distinct opacity issue: opaque notations provided by algorithms and graphical interfaces; the opaque emergence of musical form; the opaque distribution of agency between algorithms, human operators and human performers; the opaque representation of identity in the construction of digital clones of individual improvisers, made possible by this technology.

Each aspect is addressed by a specific theoretical framing and investigated empirically by ethnographic fieldwork in the research teams at Ircam currently developing Artificial Improvisation software, and empirical-musicological studies involving the recording and the analysis of solo and collective improvisations including artificial agents.

The main ambition of the project is thus to articulate a theoretically-founded and empirically-informed strategy for getting "inside" artificial improvisation.