Powerful AI models have created an existential risk to the field, but researchers can’t stop relying on them because they’re too useful.