The Cognitive Divergence: AI Context Windows, Human Attention Decline, and the Delegation Feedback Loop

Netanel Eliav

cs.CLMar 17, 2026

The Cognitive Divergence: AI Context Windows, Human Attention Decline, and the Delegation Feedback Loop

Authors: Netanel Eliav

Organizations: Machine Human Intelligence Lab · Machine Human Intelligence Lab (MHIL)

Abstract

This paper documents and theorises 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. We term the resulting asymmetry the Cognitive Divergence. AI context windows have grown from 512 tokens in 2017 to 2,000,000 tokens by 2026 (factor ~3,906; fitted lambda = 0.59/yr; doubling time ~14 months). Over the same period, 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 (2026, extrapolated from longitudinal behavioural data ending 2020 (Mark, 2023); see Section 9 for uncertainty discussion). The AI-to-human ratio grew from near parity at the ChatGPT launch (November 2022) to 556--1,111x raw and 56--111x quality-adjusted, after accounting for retrieval degradation (Liu et al., 2024; Chroma, 2025). Beyond documenting this divergence, the paper introduces the Delegation Feedback Loop hypothesis: as AI capability grows, the cognitive threshold at which humans delegate to AI falls, extending to tasks of negligible demand; the resulting reduction in cognitive practice may further attenuate the capacities already documented as declining (Gerlich, 2025; Kim et al., 2026; Kosmyna et al., 2025). Neither trend reverses spontaneously. The paper characterises the divergence statistically, reviews neurobiological mechanisms across eight peer-reviewed neuroimaging studies, presents empirical evidence bearing on the delegation threshold, and proposes a research agenda centred on a validated ECS psychometric instrument and longitudinal study of AI-mediated cognitive change.

Figures & tables

Parameter	Value	Source
Mean adult reading rate	238 wpm (range 175–300)	Brysbaert (2019) ; Rayner et al. (2016)
Tokens per English word	1.33	OpenAI cl100k_base ( OpenAI, 2023b )
Derived reading rate ( $R_{\text{tok}}$ )	5.28 tokens/s	Computed: $(238\times 1.33)/60$
Mark (2004) episode duration $A$	150 s	Mark (2023)
Mark (2016–20) episode duration $A$	47 s (median 40 s)	Mark et al. (2016) ; Mark (2023)
Session reading duration $S$ , 2004	$\approx$ 1,515 s ( $\approx$ 25 min)	Estimated from Mark; Section 3.3

Table 1: Parameters for Human ECS Token Conversion

Figure 1: The Cognitive Divergence, 2017–2026. AI context window capacity (blue, upper curve) versus human Effective Context Span (ECS; red, lower curve) expressed in tokens on a logarithmic scale. The crossover inflection occurs ca. 2022 at the ChatGPT launch. By 2026, raw AI context exceeds human ECS by 556–1,111 $\times$ . Quality-adjusted gap (accounting for retrieval degradation, Section 6) is 56–111 $\times$ . Human ECS derived via Equation ( 1 ): $\text{ECS}(t)=S(t)\times R_{\text{tok}}\times\text{CSF}(t)$ , where $S(t)$ is session reading duration; see Table 1 ; AI data from Appendix A. Note: Human 2026 value ( $\approx$ 1,800 tokens) is an extrapolated estimate; see Section 9 for uncertainty discussion. Summary statistics below the chart use 2017 as the AI baseline (512 tokens) and the corresponding 2017 human ECS ( $\approx$ 13,500 tokens); the paper’s primary narrative uses the 2004 human ECS baseline ( $\approx$ 16,000 tokens).

Year	Leading Model	AI Context	Human ECS	Ratio
2017	Transformer ( Vaswani et al., 2017 )	512	$\approx$ 13,500	0.04 $\times$
2018	GPT-1 ( Radford et al., 2018 )	512	$\approx$ 12,000	0.04 $\times$
2019	GPT-2 ( Radford et al., 2019 )	1,024	$\approx$ 10,500	0.10 $\times$
2020	GPT-3 ( Brown et al., 2020 )	2,048	$\approx$ 9,000	0.22 $\times$
2021	Codex ( Chen et al., 2021 )	4,096	$\approx$ 7,500	0.55 $\times$
Nov 2022	ChatGPT ( OpenAI, 2022 )	4,096–8,192	$\approx$ 6,000	0.7–1.3 $\times$

Table 2: AI Context Window vs. Human ECS, 2017–2026

Scenario	CSF	CSF	CSF	ECS	ECS	Raw	QA
	2004	2022	2026	2004	2026	Ratio	Ratio
Conservative ( $-30\%$ )	1.4	1.05	0.84	11,100	1,260	1,587 $\times$	119 $\times$
Baseline (paper)	2.0	1.5	1.2	16,000	1,800	1,111 $\times$	83 $\times$
Generous ( $+30\%$ )	2.6	1.95	1.56	20,800	2,340	855 $\times$	64 $\times$
Maximum plausible	2.8	2.1	1.68	22,400	2,520	794 $\times$	60 $\times$
Flat CSF (no decline)	2.0	2.0	2.0	16,000	3,000	667 $\times$	50 $\times$

Table 3: Sensitivity Analysis — Human ECS Under Variable CSF Assumptions

Study	Modality	Sample	Key Finding
Fernandes et al. (2025)	fMRI/EEG	$n=504$	Amygdala, vmPFC, ventral striatum activation; overlaps with SUD patterns
Westbrook et al. (2021)	PET	$n=22$	Putamen dopamine synthesis (+) correlated with social app use intensity
Aitken et al. (2025)	fNIRS	$n=20$	dlPFC/vlPFC activation decreased post-social media; mPFC increased
Montag and Becker (2023)	MRI (rev.)	Various	Reduced frontopolar cortex engagement in high-use cohorts; ACC changes
He et al. (2017)	MRI	$n=20$	Reduced grey matter volume in amygdala bilaterally; parallels SUD structural findings
Satani et al. (2025)	EEG	$n=100$	Gamma +62%; Beta $-35\%$ ; persistent Alpha suppression post-use

Table 4: Summary of Neuroimaging Studies on Digital Platform Use and Attentional Neuroscience

Panel of Figure 2 — Figure 2: The Delegation Feedback Loop. Each arrow represents a theorised causal relationship. Evidence bearing on each relationship is reviewed in Section 8.2–8.4. The loop is self-reinforcing: no node spontaneously reverses under current technological and behavioural trajectories.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Date	Model	Max Context (tokens)	Source
Jun 2017	Transformer (Vaswani et al.)	512	Vaswani et al. (2017)
Jun 2018	GPT-1 (Radford et al.)	512	Radford et al. (2018)
Feb 2019	GPT-2 (Radford et al.)	1,024	Radford et al. (2019)
May 2020	GPT-3 (Brown et al.)	2,048	Brown et al. (2020)
Jul 2021	Codex (Chen et al.)	4,096	Chen et al. (2021)
Nov 2022	ChatGPT / GPT-3.5	4,096–8,192	OpenAI (2022)

Appendix

Table 5: Chronological Record of LLM Maximum Context Window Sizes, 2017–2026

Explore similar work

Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective
May 7, 2026Xuening Wu, Yanlan Kang, Qianya Xu +5Model Collapse Human-AI Interaction
Fudan University, Shanghai, China · Pfizer China, Shanghai, China · University of California San Diego, CA, USA +4
Better and Worse with Scale: How Contextual Entrainment Diverges with Model Size
Apr 14, 2026Dikshant Kukreja, Kshitij Sah, Gautam Gupta +5Language Model Scaling Laws Knowledge Conflicts in Language Models
IIIT Delhi, India · Singapore Institute of Technology · NVIDIA +1
Context Compaction Theory
Aug 2, 2026Hayder Tirmazi, Sam Markelon, Allison Bishop +1LLM Context Management Context Compression
AllSpice Inc. · Boston University · Proof Trading +2

May 7, 2026J/K move · Enter open · S save

Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective

Xuening Wu, Yanlan Kang, Qianya Xu, Kexuan Xie+4

Fudan University, Shanghai, China · Pfizer China, Shanghai, China · University of California San Diego, CA, USA+4

Model Collapse Human-AI Interaction Human-AI Collaboration

Open paper AskPDF arXiv

Large language models (LLMs) are reshaping how knowledge is produced, with increasing reliance on AI systems for generation, summarization, and reasoning. While prior work has studied cognitive offloading in humans and model collapse in recursive training, these effects are typically considered in isolation. We propose a unified perspective: humans and language models form a coupled dynamical system linked by a feedback loop of usage, generation, and retraining. We introduce a minimal model with three variables -- human cognition, data quality, and model capability -- and show that this feedback can give rise to distinct dynamical regimes. Our analysis identifies three regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. Through a simple simulation, we demonstrate that increasing reliance on AI can induce a transition toward a low-diversity, suboptimal equilibrium. From an information-theoretic perspective, this transition corresponds to an emergent information bottleneck in the human-AI loop, where entropy reduction reflects loss of diversity and support under closed-loop feedback rather than beneficial compression. These results suggest that the trajectory of AI systems is shaped not only by model design, but by the dynamics of human-AI co-evolution.

Abstract

Explore similar work

Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective

Better and Worse with Scale: How Contextual Entrainment Diverges with Model Size

Context Compaction Theory