Reproducible builds, not copyleft, may govern AGI — paper argues

A new academic paper by Masayuki Hatta, accepted at the AGI-26 conference, argues that traditional copyleft licensing — the legal hack behind the GPL that ties source code availability to distribution — is fundamentally broken for AGI systems. Large language models and future AGIs can rewrite licensed source code into functionally equivalent derivatives that strip away original obligations, a form of “code laundering” that copyleft cannot defend against. The paper systematically shows that the artifacts needed to reconstruct a model — code, data, weights, hyperparameters, toolchain, and hardware — are each subject to independent legal and technical constraints that no current open-source framework resolves.

Hatta proposes that a functional analogue of copyleft for AGI must be grounded in reproducible builds: a practice guaranteeing bit-exact reconstructability from declared inputs. Drawing on the Open Source AI Definition (OSAID), the Model Openness Framework (MOF), OpenMDW, and deterministic-inference research, he defines seven requirements for AGI-oriented reproducible builds. The paper also critiques Maffulli's Second Liberation thesis, which claims AI fulfills Stallman's dream — arguing it collapses unless AGI systems are themselves reproducible. Finally, Hatta suggests that Model Context Protocol (MCP) and similar AI-to-AI coupling mechanisms form a new dynamic linking layer for which copyleft is ill-suited, advocating Masnick's "protocols, not platforms" framework as a better governance template.