It’s too easy to make AI chatbots lie about health information, study finds