Boulder NLP Seminar: Abe Handler

When News Becomes Training Data: The New Paywall Problem in the Market for Content We introduce The New Paywall Problem, which emerges from a core information asymmetry in the nascent market for news text. Technology firms claim to allow news publishers to opt out of data collection starting on specific dates. But publishers and policymakers cannot verify such promises. As publishers choose between signing deals with technology firms and litigating to shape the future of news distribution, we draw on theory from information economics to explain how this core asymmetry undermines basic conditions required for trade. We then explore a potential solution based on observable token probabilities from large language models (LLMs). We use controlled experiments across multiple empirical settings to show that when technology firms exclude documents from training data, it causes a detectable change in LLM membership inference (MI) scores that quantify the chance that an individual document was...