The web's understanding of toxic conversation has been shaped largely by a single tool: Google's Perspective API. A new arXiv paper argues that as that tool is retired and large language models take over labeling duties, the unwritten "laws" of online conversation may need to be rewritten. The paper's title, "You" Is Not Toxic, hints that even common words could be judged differently by new models.

The abstract notes that beliefs like "toxicity rises in longer and deeper threads" were measured mostly with Perspective API. With the API shutting down, LLMs will inherit the role of deciding what counts as toxic. But LLM labelers do not simply replicate the old system; they bring their own biases and thresholds, which could shift the boundary of what is flagged.

The significance is that a large body of research on online harassment and moderation may rest on measurements from a tool that is being replaced. If LLM labelers disagree with Perspective, prior conclusions about conversation dynamics could be called into question. The paper serves as a caution that the tools used to measure toxicity are not neutral arbiters, but active participants in defining it.