The Cognitive Divergence: AI Context Windows, Human Attention Decline, and the Delegation Feedback Loop
Summary: This article explores the self-reinforcing dynamic between the exponential expansion of large language model (LLM) context windows and the secular contraction of human sustained-attention capacity, terming this phenomenon the Cognitive Divergence.
The research documented in the paper, arXiv:2603.26707v1, highlights a stark contrast in the evolution of artificial intelligence capabilities compared to human cognitive abilities. Over the past few years, AI context windows have experienced extraordinary growth, with their capacity surging from 512 tokens in 2017 to a staggering 2,000,000 tokens projected by 2026. This represents a factor increase of approximately 3,906, with a fitted lambda rate of 0.59 per year and a doubling time of about 14 months.
In contrast, human Effective Context Span (ECS) has been on a steady decline. Derived from a validated reading-rate meta-analysis by Brysbaert in 2019, the ECS has dropped from an estimated 16,000 tokens in 2004 to a projected 1,800 tokens by 2026. This estimation is based on longitudinal behavioral data that concluded in 2020 as noted by Mark in 2023. The implications of this decline are significant, as they highlight the diminishing capacity of human attention in a world increasingly influenced by advanced AI technologies.
- AI Context Window Growth:
- 2017: 512 tokens
- 2026: 2,000,000 tokens (projected)
- Human Effective Context Span Decline:
- 2004: 16,000 tokens
- 2026: 1,800 tokens (estimated)
The research also discusses the escalation of the AI-to-human ratio, which has dramatically shifted since the launch of ChatGPT in November 2022. The ratio increased from near parity to an astonishing 556–1,111 times raw output and 56–111 times quality-adjusted output, after accounting for retrieval degradation as noted by Liu et al. in 2024 and Chroma in 2025. This discrepancy raises questions about the long-term implications for human cognitive skills and the nature of human-AI interaction.
In addition to documenting this divergence, the paper introduces the Delegation Feedback Loop hypothesis. This concept posits that as AI capabilities continue to expand, the cognitive threshold at which humans choose to delegate tasks to AI decreases. This phenomenon extends even to tasks that require minimal cognitive demand. Consequently, the reduction in cognitive practice may further exacerbate the decline in human attention capacities, as explored in the works of Gerlich (2025), Kim et al. (2026), and Kosmyna et al. (2025).
Neither the expansion of AI capabilities nor the decline in human attention appears to reverse spontaneously. The paper characterizes this divergence statistically and reviews neurobiological mechanisms across eight peer-reviewed neuroimaging studies. It also presents empirical evidence that informs the delegation threshold and proposes a research agenda focused on developing a validated ECS psychometric instrument along with a longitudinal study examining AI-mediated cognitive change.
This comprehensive analysis of the Cognitive Divergence not only sheds light on the current state of AI and human interaction but also raises critical questions about the future of cognition in an increasingly automated world.