Our goal shouldn't be to build merely 'trustworthy' AI