💻 computer science

The enrichment paradox: critical capability thresholds and irreversible dependency in human-AI symbiosis

This paper introduces a two-variable dynamical systems model that identifies a critical capability threshold (K* ≈ 0.85) triggering an "enrichment paradox" of abrupt human skill collapse due to AI delegation, and proposes that periodic AI failures and mandatory practice can significantly mitigate this irreversible dependency.

Original authors: Jeongju Park, Musu Kim, Sekyung Han

Published 2026-03-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jeongju Park, Musu Kim, Sekyung Han

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is like a muscle. If you go to the gym every day, it gets stronger. If you stop going and let a robot lift the weights for you, your muscle doesn't just stay the same—it actually shrinks and gets weaker. This is the basic idea of "use it or lose it."

This paper takes that simple idea and turns it into a mathematical warning about our future with Artificial Intelligence (AI). The authors, researchers from Korea, have built a model to predict exactly when relying on AI becomes dangerous.

Here is the breakdown of their findings using simple analogies:

1. The "Enrichment Paradox": Why Better AI is Dangerous

You might think, "If AI gets smarter, that's great! It will help us do more."
The authors call this the Enrichment Paradox. Think of it like a garden.

The Old Way: If you have a weak gardener (AI) and a strong human, the human does most of the work. The human stays strong.
The Trap: As the gardener gets better and smarter, the human starts to relax and stop gardening.
The Crash: There is a specific tipping point (about 85% of human-level capability). Once the AI crosses this line, the human doesn't just get a little weaker; they suddenly collapse. It's like a dam breaking. The human capability drops from "strong" to "almost zero" very quickly.

The Analogy: Imagine a pilot flying a plane. If the autopilot is 50% good, the pilot still flies manually most of the time. But if the autopilot becomes 95% perfect, the pilot stops flying manually entirely. If the autopilot then fails, the pilot has forgotten how to fly and crashes. The better the tool, the more we forget how to use our own skills.

2. The "Point of No Return" (Irreversibility)

The most scary part of the paper is that this damage might be permanent.

The Analogy: Think of a language. If you stop speaking French for 10 years, you can relearn it. But if you stop speaking it for 50 years, and your children never learn it, and the books are lost, the language is gone forever.
The Model: The researchers show that once human capability drops too low, it becomes incredibly hard to get back. Why? Because to learn a skill, you need to already have a little bit of that skill to build on. If you have zero skill left, you can't start learning again.
The Result: If we let AI take over completely, and then the AI breaks or disappears, we might not be able to recover our skills for centuries. It's like a biological "dead end."

3. The "Fire Drill" Solution (Antifragility)

So, how do we stop this? The paper suggests a counter-intuitive idea: We need the AI to fail sometimes.

The Analogy: Think of a fire drill in a building. If the fire alarm never goes off, people forget what to do. But if the alarm rings randomly, people practice the escape route. They stay sharp.
The Finding: The model shows that if AI fails occasionally (say, 20% of the time), humans are forced to step in and do the work. This "forced practice" keeps our skills alive.
The Policy: The authors suggest we should mandate that humans do a certain amount of work without AI. For example, "One day a week, no AI allowed."
- Their math says: If we force humans to do just 20% of their work manually, we can preserve 92% of our skills compared to letting AI do everything.

4. The Real-World Proof

The researchers didn't just guess; they looked at real data:

Pilots: When pilots fly too much with autopilot, they fail basic flying tests.
Doctors: When doctors use AI to help diagnose, they get worse at diagnosing when the AI is turned off.
Students: Students using AI for homework score lower on tests when they can't use the AI.
GPS: People who use GPS constantly have worse memory for maps than those who don't.

The Bottom Line

The paper argues that we are currently walking toward a cliff. We are making AI so good that we are forgetting how to be human at the tasks we used to do.

The Solution isn't to stop using AI. It's to manage it. We need to treat our human skills like a muscle that needs exercise. We should set rules that force us to practice our skills manually, even when the AI is available. If we don't, we risk becoming a civilization that is totally dependent on a tool we no longer know how to fix or replace.

In short: Don't let the robot do everything. Keep your own hands on the wheel, even if the robot drives better. Otherwise, one day, the robot might stop, and we won't know how to drive.

1. Problem Statement

The paper addresses a critical gap in the literature regarding human–AI symbiosis: while qualitative warnings about "deskilling" (the atrophy of human skills due to over-reliance on automation) exist, there is no quantitative framework to predict when human capability loss becomes catastrophic or irreversible.

Current models (e.g., Bass diffusion, task-based economic frameworks) capture adoption dynamics or economic displacement but fail to model the feedback loop between delegation and capability loss. The authors ask: At what point does the substitution of human cognitive labor by AI lead to a sudden, irreversible collapse of human capability, rather than a gradual decline?

2. Methodology

The authors propose a minimal dynamical systems model based on two coupled ordinary differential equations (ODEs) and validated via Agent-Based Modeling (ABM).

A. The Mathematical Model

The system couples two state variables:

$H(t) \in [0, 1]$ : Human capability (fraction of full competence).
$D(t) \in [0, 1]$ : Delegation rate (fraction of tasks delegated to AI).

The dynamics are governed by three axioms:

Learning requires existing capability: Growth is proportional to current skill ( $H$ ).
Learning requires practice: Growth only occurs on tasks not delegated ( $1-D$ ).
Disuse causes forgetting: Capability decays proportional to delegation ( $D$ ).

Equations:
$\frac{dH}{dt} = \alpha (H + \varepsilon)(1-H)(1-D) - \beta H D$
$\frac{dD}{dt} = \gamma (K - H)(1-D)D + \delta D(1-D)\bar{D}$

$\alpha$ : Learning rate.
$\beta$ : Forgetting rate (calibrated empirically).
$\gamma$ : Sensitivity to the capability gap (rational adoption).
$K$ : AI capability (exogenous parameter).
$\delta$ : Social pressure (contagion effect).
$\varepsilon$ : Baseline capacity for re-learning from near-zero (prevents strict mathematical absorption at $H=0$ , making recovery technically possible but practically slow).

B. Validation and Calibration

Empirical Calibration: The forgetting rate ( $\beta$ ) was fitted to data from four distinct domains: Education (GPT-4 usage), Medical Endoscopy, Spatial Cognition (GPS), and Aviation (Autopilot).
Macro-Validation: The model was fitted to 15 countries' PISA mathematics scores (2003–2022) using internet adoption rates as a proxy for AI exposure.
Model Discrimination: The ODE model was compared against linear, exponential, and logistic decay models using AIC/BIC metrics.
Simulation: An Agent-Based Model (ABM) with $N=100$ agents was used to simulate stochastic dynamics, crisis events, and generational turnover.

3. Key Contributions

A. The "Enrichment Paradox"

Drawing an analogy to Rosenzweig's paradox in ecology, the authors identify a counterintuitive phenomenon: improving AI capability ( $K$ ) can destabilize the human–AI system.

As $K$ increases, the incentive to delegate rises, accelerating the loss of human skill ( $H$ ).
This creates a critical threshold ( $K^*$ ) beyond which the system undergoes a phase transition from stable autonomy to dependency collapse.

B. Identification of a Critical Threshold ( $K^*$ )

The model predicts a sharp, non-linear transition rather than a smooth decline.

$K^* \approx 0.85$ : When AI capability reaches approximately 85% of human expert level (in a broad scope), human capability collapses abruptly.
Bistability: Below $K^*$ , the system has two stable states (Autonomous and Dependent) separated by a saddle point. Above $K^*$ , the Autonomous state disappears, leaving only the Dependent state ( $H \to 0, D \to 1$ ).

C. Irreversibility (The "Muller's Ratchet" Effect)

The model demonstrates that once human capability drops near zero, recovery is prohibitively slow.

Because learning requires existing capability ( $H$ ), a near-zero $H$ results in a near-zero learning rate, even if AI is removed.
Recovery relies on the $\varepsilon$ term (re-learning from scratch), which operates on timescales (centuries) far exceeding institutional planning horizons. This mirrors the irreversible genome reduction seen in endosymbiotic bacteria.

D. Antifragility via Periodic Failure

Contrary to conventional reliability engineering, the model shows that perfect AI reliability maximizes human vulnerability.

Introducing periodic AI failures (crises) forces humans to practice, maintaining the learning term.
Simulations show that a 25% crisis frequency can improve equilibrium human capability by 2.7-fold compared to a perfectly reliable AI environment.

4. Key Results

Empirical Fit

PISA Data: The ODE model achieved an $R^2 = 0.946$ across 102 data points (15 countries) with only 3 parameters, outperforming linear and exponential decay models (which required 16–30 parameters for similar fit).
Domain Calibration: The model successfully reproduced observed skill decay rates across education (17% drop), endoscopy (21% drop), GPS navigation (30% drop), and aviation (38% drop) by varying only the forgetting rate $\beta$ and timescale.

Threshold Analysis

Current State: Frontier models (GPT-4, Claude 3.5) have an average capability ratio $\bar{K} \approx 0.94$ , placing them well above the critical threshold ( $K^* \approx 0.85$ ).
Implication: In domains like medical reasoning and coding, the system has likely already crossed the tipping point, leading to rapid capability erosion.

Policy Interventions

Mandatory Practice: The model quantifies the efficacy of "AI-free" mandates.
- A policy requiring 20% of tasks to be performed without AI (e.g., one day a week) preserves 92% more capability than the baseline (which includes a 5% background failure rate).
- This relationship is superlinear: small increases in mandatory practice yield disproportionately large gains in capability preservation.

5. Significance and Implications

Governance Shift: The findings suggest AI governance should shift from asking "Should we adopt AI?" to "How capable should AI be allowed to become in specific domains before mandatory preservation measures are triggered?"
Risk of Irreversibility: The model warns that once a society crosses the $K^*$ threshold, the loss of capability may be irreversible on human timescales. A civilization that fully delegates a skill to AI may never recover it if the AI fails or is lost.
Antifragile Design: The results argue for designing "fire drills" and degraded-mode training into human–AI workflows. Deliberate, periodic interruptions of AI assistance are necessary to maintain the "antifragility" of the human workforce.
Historical Context: The model explains why calculators (narrow scope, low social pressure) were safe, while Roman slave economies (broad scope, low cost) and modern AI (broad scope, low cost) pose existential risks to human capability.

In conclusion, the paper provides a rigorous mathematical foundation for the fear that AI-induced deskilling is not a linear process but a catastrophic, threshold-driven event that requires proactive, quantitative policy intervention to prevent irreversible dependency.