This article argues that human–AI friendships are possible and, in some current and many future contexts, can be valuable—both for the humans participating in them and for the wider society in which such friendships exist. To begin, the article discusses some weaknesses in existing conceptions of friendship, including the theories of friendship commonly applied to human–AI relationships. It proposes a flexible conception of friendship that avoids these weaknesses. The article then analyzes the main objections to human–AI friendships being either possible or good—that humans cannot see AI as their friends; that these friendships are not good because they are unequal; that AI lack the capacity to be a friend; and that human–AI friendships displace (more valuable) human–human friendships. It argues that each of these objections fails to withstand critical scrutiny. In some cases, the analyses also amount to a rejection of existing views of friendship that require what are taken to be unnecessary features. Having dealt with these objections, the article shows that human–AI friendship is possible on the more flexible conception of friendship. It then provides some evidence that human–AI friendships can be good, such as by providing human users with the experience of being emotionally supported. The article also notes some practical concerns about the ways human–AI friendships might proliferate and the moral responsibilities that might accrue to governments and corporations accordingly. The article concludes that the new view of friendship means the value human–AI friendship should be re-evaluated as possible and potentially valuable today and in the future.