Practitioners should test AI models by sending different data contexts for the same prompt