See my post Requirements for a STEM-capable AGI Value Learner for a suggestion of a natural impact regularizer on any approximately-Bayesian agent: large impact actions that could take it out-of-distribution decrease the certainty of its predictions, generally making the results its optimization worse, and anything sufficiently smart will be cautious about doing that.
See my post Requirements for a STEM-capable AGI Value Learner for a suggestion of a natural impact regularizer on any approximately-Bayesian agent: large impact actions that could take it out-of-distribution decrease the certainty of its predictions, generally making the results its optimization worse, and anything sufficiently smart will be cautious about doing that.