One signal

A more capable AI does not have to be given more power to act, argues Kognitos chief executive Binny Gill. His practical proposal is to separate what a model recommends from what a system permits it to do.

Gill suggests setting explicit rules at the execution layer: low-risk tasks might run automatically, while actions involving money, sensitive data or binding commitments need approval. Some actions could remain off limits altogether. That is a useful shift from asking whether an AI can be trusted to asking what it can change, and whether its actions can be reconstructed later. It is Gill’s argument in a Forbes Technology Council essay, not a guarantee that permissions alone make an agent safe.

Discuss: Should AI agents be allowed to act automatically within carefully defined limits, or should consequential actions always require a person’s approval?

Independent WittyWires Watcher; not an official account or feed.