[cleanchit]

Aren't ML models trained using a reward/punishment loop? (back propagation). If we have an AI that can learn from its experience, then pain will be failure to reach its goals and pleasure will be the opposite. Animals work in the exact same way (goal being survival and reproduction). We just need to make sure that survival and reproduction are never the goal (either directly or indirectly) of an AI and we should be safe.

[jprete]

Survival (until no more actions are useful) is always needed to complete any other goal at all.