customers in response to pressure, but critics argue that even with zero data retention, labs can still learn from usage patterns and metadata. OpenAI and Anthropic collect de-identified metadata to understand service usage, though some customers remain sceptical about what exactly is being collected. AI researcher John Schulman outlined a spectrum of data usage risks, from direct pretraining to building reinforcement learning tasks from user traces, which can extract customer IP without reproducing content. Sarah Hooker warned that synthetic data techniques can preserve privacy while still allowing labs to extract statistical patterns, effectively allowing companies to fuel frontier labs that may encroach on their verticals.