Defamation has never been primarily about truth in the abstract, and it has never depended on whether harm exists in some general sense. It has always depended on something much more operational and much less visible: the ability to point to a speaker and say that this person made this statement in a way that can be examined, challenged, and, if necessary, sanctioned. The entire structure of defamation law rests on that anchor, because without it there is no stable way to transform reputational harm into a legal claim.
AI-generated outputs do not simply complicate that structure. They undermine the condition that makes it possible. The problem is not that harmful statements are becoming harder to evaluate or that verification requires more effort. The problem is that statements are increasingly detached from any origin that the system can recognize as a speaker in the first place, which means that the legal logic built around attribution begins to lose its point of application.
What emerges is not a more difficult version of defamation. It is an environment in which reputational harm continues to circulate with increasing efficiency while the mechanism designed to address it struggles to locate something it can meaningfully act upon.
The disappearance of the speaker is not an edge case, it is the default
Traditional information environments, even when fragmented and fast-moving, still produced identifiable points of authorship. A journalist publishes an article, a user posts a claim, a platform hosts content that can be traced back through accounts and timestamps, and even when the chain is long or obscured, it ultimately converges on an actor whose role can be defined. The system tolerates complexity because it still produces endpoints.
AI systems do not behave in that way, and more importantly, they do not need to. An output generated by a model is assembled through the interaction of training data, probabilistic inference, prompt structure, and platform constraints, yet none of these elements alone can be isolated as the author of the resulting statement. The output looks like speech, it reads like speech, and it functions like speech in its effects, but it is not anchored to a speaker in a way that survives legal scrutiny.
This is not a temporary limitation that can be resolved with better tooling or clearer disclosures. It is a structural property of generative systems. The output exists as a surface of the system. The speaker does not exist in a form that the law can reliably engage with.
