Without watermarking, nobody will write, and nobody will read (twitter.com)

The recent essay argues that without implementing watermarking for text generated by large language models (LLMs), the landscape of writing and reading could face a severe decline. The author posits that as LLMs become indistinguishable from high-quality human writing, readers will be less inclined to engage with any text due to a growing suspicion that it may be AI-generated. This mistrust could lead to a downward spiral: as more writers disengage from creating content, the proportion of human-generated text will dwindle, leading readers to forgo reading altogether, thereby exacerbating the problem. This scenario draws parallels to George Akerlof's "Market for Lemons" theory, which illustrates how the indistinguishability of good and bad products can create market collapse. The essay emphasizes that just as consumers are willing to pay a premium for natural diamonds over synthetic ones, readers might prefer human-authored writing if they can distinguish it from AI-generated content. Thus, watermarking could serve as a crucial mechanism to restore trust and allocate attention efficiently, allowing human writers to thrive alongside AI-generated text and preserving the diversity in content.