arxiv: 2605.06347 · v1 · submitted 2026-05-07 · 💻 cs.HC · cs.AI

Recognition: unknown

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

Xuening Wu , Yanlan Kang , Qianya Xu , Kexuan Xie , Jiaqi Mi , Honggang Wang , Yubin Liu , Zeping Chen

Authors on Pith no claims yet

Pith reviewed 2026-05-08 07:05 UTC · model grok-4.3

classification 💻 cs.HC cs.AI

keywords human-AI interactiondynamical systems modelinginformation bottleneckepistemic collapsemodel collapseLLM feedback loopscognitive offloading

0 comments

The pith

Humans and AI form a coupled dynamical system whose feedback can drive it into a low-diversity, suboptimal state.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The authors treat the interaction between humans and large language models as a single dynamical system connected by loops of use, generation, and retraining. They build a minimal model using three variables—human cognition, data quality, and model capability—to identify three possible regimes of behavior. One regime shows mutual improvement, another a delicate balance, and the third a decline where reliance on AI reduces variety and quality. Simulation results indicate that higher AI use can push the system across a threshold into the declining regime. This decline corresponds to an information bottleneck that reduces entropy without providing the benefits of useful compression.

Core claim

Treating humans and language models as a coupled dynamical system linked by a feedback loop of usage, generation, and retraining, the analysis reveals three regimes—co-evolutionary enhancement, fragile equilibrium, and degenerative convergence—with the transition to the latter driven by increasing AI reliance and appearing as an emergent information bottleneck in the loop.

What carries the argument

A minimal three-variable dynamical model consisting of human cognition, data quality, and model capability connected in a feedback loop.

Load-bearing premise

The minimal model with three variables and the assumed feedback loop structure is sufficient to capture the essential dynamics of human-AI co-evolution.

What would settle it

Observing whether knowledge diversity and quality decline in real-world settings where AI usage has increased substantially over time, such as in online content creation or academic writing.

Figures

Figures reproduced from arXiv: 2605.06347 by Honggang Wang, Jiaqi Mi, Kexuan Xie, Qianya Xu, Xuening Wu, Yanlan Kang, Yubin Liu, Zeping Chen.

**Figure 1.** Figure 1: The closed-loop dynamics of epistemic intelligence. The diagram illustrates the coupled ODE system, with state variables H(t), Q(t), M(t) (top) linked by forward enhancement pathways (solid arrows) and control variables u, A, S (bottom) introducing negative feedback via cognitive offloading and recursive data contamination (dashed arrows). External inputs rH, rQ, rM represent intervention mechanisms. The f… view at source ↗

**Figure 2.** Figure 2: Simulation of the coupled human–data–model system under three representative regimes. (a) Enhancement: all variables increase over time due to dominant positive feedback among human cognition (H), data quality (Q), and model capability (M). (b) Equilibrium: the system converges to a bounded fixed point, reflecting a balance between reinforcing and degrading feedback. (c) Degeneration: human cognition and d… view at source ↗

**Figure 3.** Figure 3: Steady-state behavior of the coupled system as a function of AI dependence u. As u increases, the system transitions from an enhancement regime to a degenerative equilibrium. The dashed line indicates the critical threshold uc, identified by the maximal gradient change in Q ∗ (u). The shaded region denotes the transition band separating distinct dynamical regimes. The monotonic decline of H∗ , Q ∗ , and M… view at source ↗

read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper sketches a three-variable dynamical model to unify model collapse and cognitive offloading but the simulation stays too abstract and unvalidated to support the named regimes or the bottleneck claim.

read the letter

The main contribution is the attempt to link human cognition, data quality, and model capability in one feedback loop and then label the resulting behaviors as co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. The information-bottleneck reading of the low-diversity state is the other piece they add on top of earlier isolated studies of recursive training and offloading. That framing could give people working on long-term AI effects a shared vocabulary for talking about closed-loop risks. The unification itself is the part that feels like a step forward from papers that treat those problems separately. The simulation is presented as evidence that higher AI reliance drives the system toward the bad equilibrium, which is a clear enough story on its own terms. The soft spots are bigger than the strengths. The model equations, coupling rates, and exact simulation steps are not laid out, so it is impossible to reproduce or test the regimes. No sensitivity checks, no comparison to any external dataset, and no discussion of what would falsify the transition appear. Because the three regimes and the entropy drop are generated directly by the feedback rules the authors chose, the circularity the stress-test note flags is real: the results largely restate the assumptions. Real-world factors such as policy changes, competing models, or multi-modal inputs are left out, and those could easily alter whether the degenerative state emerges. This is the sort of speculative modeling piece that might interest a reading group focused on AI futures or interface design, but it does not yet have the grounding or robustness for technical readers who need testable mechanisms. I would not cite the specific regimes or the bottleneck mapping in my own work. It does not deserve peer review until the model is written out explicitly, the simulation is documented with parameters and validation steps, and the authors address how additional variables would change the phase portrait.

Referee Report

3 major / 1 minor

Summary. The manuscript claims that humans and large language models form a coupled dynamical system through feedback loops of usage, generation, and retraining. Using a minimal three-variable model (human cognition, data quality, model capability), the authors identify three dynamical regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. A simple simulation demonstrates that increasing AI reliance can drive the system toward a low-diversity, suboptimal equilibrium, which from an information-theoretic view corresponds to an emergent information bottleneck in the human-AI loop rather than beneficial compression.

Significance. If the simulation results hold under scrutiny, this work provides a novel unified framework for understanding the interplay between cognitive offloading and model collapse in AI systems. It highlights how human-AI co-evolution dynamics, rather than just model design, can shape trajectories toward epistemic issues. The approach is timely given rapid LLM adoption, but its significance is tempered by the absence of empirical validation or sensitivity analysis in the current presentation.

major comments (3)

The minimal model is introduced but its governing equations, the specific forms of the feedback loops, parameter values (e.g., coupling rates and thresholds), and simulation implementation details are not provided. This omission is load-bearing because the three regimes and the transition to degenerative convergence are generated by these equations and parameters.
No error analysis, sensitivity checks to parameter variations, or comparison to real-world data or benchmarks are described. This undermines confidence in the robustness of the claimed transition induced by increasing AI reliance.
The information-bottleneck interpretation relies on internal entropy reduction within the closed-loop model; it would benefit from a concrete test or external data to confirm it reflects loss of diversity rather than other effects.

minor comments (1)

The abstract could more clearly distinguish the proposed regimes from prior work on cognitive offloading and model collapse.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. The comments highlight important areas for strengthening the presentation of the model and its analysis. We will revise the paper to incorporate additional details and robustness checks where feasible. Our responses to each major comment are provided below.

read point-by-point responses

Referee: The minimal model is introduced but its governing equations, the specific forms of the feedback loops, parameter values (e.g., coupling rates and thresholds), and simulation implementation details are not provided. This omission is load-bearing because the three regimes and the transition to degenerative convergence are generated by these equations and parameters.

Authors: We agree that the absence of these details limits reproducibility and clarity. In the revised manuscript, we will add a dedicated section presenting the full set of governing equations for the three-variable system (human cognition, data quality, model capability), the explicit functional forms of all feedback loops, the complete list of parameter values including coupling rates and thresholds, and a description of the numerical simulation implementation (including integration method, time steps, and initial conditions). This will allow full reproduction of the reported dynamical regimes. revision: yes
Referee: No error analysis, sensitivity checks to parameter variations, or comparison to real-world data or benchmarks are described. This undermines confidence in the robustness of the claimed transition induced by increasing AI reliance.

Authors: We acknowledge the value of robustness analysis. We will add a new subsection performing sensitivity analysis by varying key parameters such as AI reliance rates and coupling strengths, and report the stability of the identified regimes and transition points. However, direct comparisons to real-world data or benchmarks are not possible at this stage, as the model is intentionally minimal and abstract; we will expand the discussion to outline potential empirical proxies and limitations. revision: partial
Referee: The information-bottleneck interpretation relies on internal entropy reduction within the closed-loop model; it would benefit from a concrete test or external data to confirm it reflects loss of diversity rather than other effects.

Authors: The interpretation arises directly from the model's entropy measures on the human cognition variable. We will enhance the analysis by adding explicit calculations and visualizations of diversity metrics (e.g., state variance) to demonstrate the correspondence to loss of diversity. A concrete external test or dataset is not available for this specific closed-loop system, so we will present the result as a theoretical prediction and suggest directions for future empirical validation. revision: partial

standing simulated objections not resolved

Direct comparison to real-world data or benchmarks for the claimed transition and information-bottleneck effect, as no suitable external datasets exist for this coupled human-AI dynamical system.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces a minimal three-variable dynamical model (human cognition, data quality, model capability) with an assumed feedback loop and analyzes its behaviors through simulation to identify three regimes and a transition to low-diversity equilibrium. These outcomes are direct consequences of the model's equations and parameters, as is standard and expected in theoretical modeling papers; the information-bottleneck interpretation is an additional post-hoc perspective applied to the simulation results rather than a reduction of independent claims to inputs. No self-citations, fitted parameters renamed as predictions, ansatzes smuggled via citation, or uniqueness theorems are referenced in the provided text. The derivation is self-contained as an exploration of the proposed model under its stated assumptions.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The claims depend on the modeling choice of three continuous variables linked by unspecified feedback rules and on the interpretation of simulation outputs as real epistemic phenomena. No independent evidence for these modeling decisions is supplied in the abstract.

free parameters (1)

feedback coupling rates and thresholds
The minimal dynamical model requires parameters to govern how human cognition, data quality, and model capability influence one another; these are not listed but must exist to produce the reported regimes.

axioms (1)

domain assumption Human cognition, data quality, and model capability can be adequately represented as three coupled continuous dynamical variables.
This is the core modeling premise stated in the abstract that enables the regime analysis.

invented entities (1)

Co-evolutionary enhancement, fragile equilibrium, and degenerative convergence regimes no independent evidence
purpose: To classify the long-term behaviors of the human-AI feedback system.
These categories emerge from the simulation of the proposed model and have no independent empirical grounding supplied.

pith-pipeline@v0.9.0 · 5524 in / 1573 out tokens · 38674 ms · 2026-05-08T07:05:31.512000+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Langley , title =

P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

2000
[2]

T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

1980
[3]

M. J. Kearns , title =
[4]

Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

1983
[5]

R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

2000
[6]

Suppressed for Anonymity , author=
[7]

Newell and P

A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

1981
[8]

A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

1959
[9]

Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task

Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task , author=. arXiv preprint arXiv:2506.08872 , year=

work page arXiv
[10]

Nature , year=

The Curse of Recursion: Training on Generated Data Makes Models Forget , author=. Nature , year=
[11]

Nature , volume=

Machine behaviour , author=. Nature , volume=
[12]

science , volume=

Google effects on memory: Cognitive consequences of having information at our fingertips , author=. science , volume=. 2011 , publisher=

2011
[13]

The information bottleneck method

The information bottleneck method , author=. arXiv preprint physics/0004057 , year=

work page internal anchor Pith review arXiv