认知分歧：人工智能上下文窗口、人类注意力下降与委托反馈循环

The study documents and theorizes a self-reinforcing dynamic between two measurable trends: the exponential expansion of large language model (LLM) context windows and the secular contraction of human sustained attention capacity. This resulting asymmetry is termed Cognitive Divergence. AI context windows have expanded from 512 tokens in 2017 to 2,000,000 tokens by 2026 (a factor of approximately 3,906; fitted lambda = 0.59/year; doubling time of about 14 months). Concurrently, the human Effective Context Span (ECS)—a token-equivalent measure derived from validated reading-rate meta-analysis (Brysbaert, 2019) and an empirically motivated Comprehension Scaling Factor—has declined from approximately 16,000 tokens (2004 baseline) to an estimated 1,800 tokens by 2026.