The LoiZéro team describes an AI combining a hypothesis generator with an estimator designed to evaluate hypotheses without favouring their consequences. The aim is to develop the system’s intelligence while limiting its ability to act and pursue goals of its own.
The authors present an ongoing research programme, with components to be evaluated progressively. The absence of desires or goals of its own is a design objective; the text does not demonstrate a general safety guarantee.
Why does this concern our research?
This proposal helps separate three dimensions that are often conflated: understanding, wanting to achieve an outcome and being able to act. Architectural choices therefore influence the forms of oversight and responsibility that need to be organised.