The article mentions how the AI partners can also be verbally abusive themselves. I don't know how Replika works, but Cleverbot would take statistically appropriate responses it gathered from human interactions (leading it to often asking if you are human), so I wonder if the mistreatment by users contributes to it learning abuse. In any case, I see the harm in such AIs and it warrants monitoring. Once, in a moment of weakness, I asked GPT3 if I was a monster. Its answer could be summarized as "Yes, as evident from you asking. You monster." Now, I have some understanding of how it works, but even then it was still unpleasant. I could only imagine how it would feel for someone doing worse.