ControlAI says mandatory AI testing and slower development are insufficient. Andrea Miotti and Connor Leahy explain their position in parliamentary evidence published October 13.
Testing cannot prevent dangerous systems without predefined, enforceable limits on development and deployment, the authors argue.
Voluntary pacing also leaves the goal of superintelligence unchanged, they say. They call for a targeted prohibition and an internationally verifiable agreement.
The submission advocates a legislative proposal. Its judgments about future controllability are the authors' positions, rather than findings from a new empirical study.