Anthropic is reportedly making it easier to detect whether text was generated by AI, a technical fix to a social problem that public shaming has apparently been solving faster anyway. The New Yorker's Brady Brickner-Wood notes that few forces have proved as powerful at tempering AI use as the threat of being caught. The watermark is the institutionalization of the cringe. What happens when the cringe gets automated?
Detection as Social Infrastructure
The parallel to consider here is the AI provenance paper currently circulating on arXiv: Adam Mazzocchetti's runtime governance framework proposes trusted provenance as a baseline for any AI agent that can take consequential actions. The framing is technical, about file modifications and fail-closed execution, but the social logic is identical to the watermark debate. Provenance means: who made this, how, and can it be verified. That is exactly what public shaming tries to establish in the absence of technical infrastructure. We built the shame loop because we didn't build the provenance system. Now both are arriving simultaneously, and the question is whether they reinforce or undercut each other.
What Shaming Actually Regulates
The deeper issue is that shaming targets output, the finished text, the polished LinkedIn post, while technical detection targets process. A 2026 arXiv paper by Despoina Giarimpampa et al. on LLMs as surrogate experts in security surveys found systematic bias in how AI models describe practitioner workflows, bias that is invisible in the output and only legible at the process level. Shame cannot see process. Watermarks can. But watermarks are also gameable, strippable, and subject to the same adversarial pressure as every other detection system. Kyle Chayka's thinking on algorithmic homogenization applies here with uncomfortable precision: the watermark may just train the next generation of text to be more convincingly human-sounding, laundering the slop rather than stopping it. The mark of the machine is coming. Whether it marks anything worth marking is the question nobody is asking.