Proxies are a promising and potentially-risky way to test the safety of AI systemsHow can scientists (and communities) make general claims about the effects of AI systems that adapt to humans in different ways?