The Cognitive Divergence: AI Context Windows, Human Attention Decline, and the Delegation Feedback Loop
AI context windows grew 3,906x since 2017 while human attention span shrank 90%, creating a massive cognitive gap.
A new research paper by Netanel Eliav from the Machine Human Intelligence Lab presents a stark, data-driven analysis of a growing asymmetry between artificial and human intelligence. The study, titled 'The Cognitive Divergence: AI Context Windows, Human Attention Decline, and the Delegation Feedback Loop,' tracks two opposing trends. On one side, large language model (LLM) context windows have exploded exponentially, growing by a factor of ~3,906 from 512 tokens in 2017 to 2,000,000 tokens by 2026, with a doubling time of roughly 14 months.
On the other side, human sustained-attention capacity is in secular decline. The paper introduces a token-equivalent metric called the 'Effective Context Span' (ECS), derived from reading-rate meta-analyses. It estimates that the human ECS has contracted from approximately 16,000 tokens (a 2004 baseline) to just 1,800 tokens in 2026. This creates a raw AI-to-human context ratio of 556–1,111x, which adjusts to 56–111x when accounting for known AI retrieval degradation issues.
Beyond documenting this divergence, the paper's core contribution is the 'Delegation Feedback Loop' hypothesis. It posits that as AI capabilities grow, the cognitive threshold at which humans choose to delegate tasks to AI drops, extending even to tasks of negligible mental demand. This reduced cognitive practice may then further attenuate the very human capacities already in decline, creating a self-reinforcing negative cycle. The paper reviews supporting neurobiological evidence from eight imaging studies and argues neither trend reverses spontaneously.
Finally, the work proposes a concrete research agenda to address this issue. It calls for the development of a validated psychometric instrument to measure the Effective Context Span and advocates for longitudinal studies to directly track AI-mediated cognitive change over time, framing this as a critical area for future human-computer interaction and societal research.
- AI context windows grew from 512 tokens (2017) to 2 million tokens (2026), a ~3,906x increase.
- Human 'Effective Context Span' declined from ~16,000 tokens (2004) to ~1,800 tokens (2026), a ~90% contraction.
- Proposes the 'Delegation Feedback Loop': easier AI delegation may further reduce human cognitive practice and capacity.
Why It Matters
Warns of a self-reinforcing cycle where reliance on powerful AI could permanently degrade fundamental human cognitive skills.