How Humans and AI Shape the Web: A Scalable Model for Sustainable Knowledge Growth

Table of Contents

• The Five-Variable Engine Behind Human-AI Knowledge
• Growth Regimes: From Healthy Growth to Oscillations and Collapse
• Real-World Case Studies: PubMed, GitHub Copilot, and Wikipedia
• Design Levers for Sustainable Growth
• Key Takeaways
• Sources & Further Reading

Introduction

Imagine the web’s shared knowledge as a living, breathing ecosystem where humans and AI models like large language models (LLMs) both shape and learn from what’s on the internet. That’s the central idea of the new research on the Dynamics of Human-AI Collective Knowledge on the Web: A Scalable Model and Insights for Sustainable Growth. The authors lay out a compact, interpretable model that tracks five pieces at once: how big the archive is, how good it is, how capable the AI is, how skilled the human community is, and how many questions people ask. The big takeaway? Small design choices—like how strictly platforms gate AI content or how we train AI through human feedback—can push this system toward healthy growth or toward dangerous cycles of quality loss and skill erosion. If you want a quick read on why this matters now, this work is a timely lens on our era of rapid AI integration, and you can explore the full paper here: Dynamics of Human-AI Collective Knowledge on the Web: A Scalable Model and Insights for Sustainable Growth.

This blog distills the core ideas in plain language, with practical takeaways you can apply to real platforms today. We’ll walk through what the model actually tracks, what the different growth regimes mean for the web you use, a few real-world snapshots (PubMed, GitHub Copilot, and Wikipedia), and what platform designers and policymakers can do to keep knowledge robust, diverse, and useful.

Why This Matters

Three things make this research feel especially urgent right now.

• First, AI systems are no longer “tools” tucked away behind a product. They’re part of the web’s knowledge cycle: people read AI outputs, cite them, and those outputs flow back into the archive and influence future AI training. That feedback loop can accelerate growth, but it can also magnify mistakes, bias, and quality erosion if not managed.

• Second, we’re starting to see real-world consequences. The paper frames systemic risks like quality dilution (where AI-generated content crowds out high-quality human work) and human competence inversion (where humans become less capable because they rely too much on AI). These aren’t isolated problems—they’re intertwined and can spiral if left unchecked.

• Third, the work provides a practical, testable framework. It’s not just a theoretical model; it’s calibrated with real data (Wikipedia, PubMed, GitHub Copilot) and shows how different platforms live in different growth regimes. That makes it a useful tool for audits and policy design today.

You can read the original research for the full math and experiments here: Dynamics of Human-AI Collective Knowledge on the Web: A Scalable Model and Insights for Sustainable Growth.

The Five-Variable Engine Behind Human-AI Knowledge

Think of the knowledge ecosystem as a small, five-part machine that evolves over time. The model tracks:

• K(t): Archive Size

  • The total number of knowledge items in the shared archive (papers, code, wiki pages, etc.). Content can enter the archive from humans (their contributions) or from AI, but AI entries pass through a gate that moderates quality.

• q(t): Archive Quality

  • A quality score between 0 and 1 that reflects how faithful and reliable the items are. It’s nudged up by high-quality human contributions and by accurate AI content, but it can drift down if AI content comes in with low accuracy or if knowledge becomes obsolete.

• θ(t): LLM Skill

  • A proxy for the model’s capacity and accuracy. It grows via two learning channels: corpus-driven scaling (training on the archive’s content) and learning from human feedback (RLHF). The more and better the data and feedback, the higher θ climbs—though there are diminishing returns.

• H(t): Aggregate Human Skill

  • The overall capability of the human community in using, creating, and evaluating knowledge. It grows by self-study of the archive and by learning from AI explanations, while it also decays over time if new information inflow slows or if overreliance on AI clouds humans’ own skills.

• Q(t): Query Volume

  • How many questions users ask LLMs, i.e., demand for AI-generated information. This isn’t just a count; it’s shaped by how hard a user finds a problem and how much they have already learned.

Two quick visuals to anchor the idea:
- Content inflows: Humans contribute at a steady pace, while LLMs contribute content when the system’s gate admits AI material. A looser gate means more AI in the archive; a stricter gate preserves quality but slows growth.
- Learning pathways: Humans improve by studying the archive (self-contained learning) or by taking AI-derived explanations (assisted learning). The relative appeal of these paths affects the growth of H(t) and, through K(t) and q(t), the long-term health of the ecosystem.

Key point from the authors: you can push the system toward healthy growth by balancing the three actually active levers—how much AI content you admit, how you train the AI, and how people learn (from the archive vs. from AI).

For readers who like a mental model, picture a seesaw with three pins at play: the rate of human contributions, the rate of AI contributions, and the gate that decides what gets kept. Move any pin and the whole balance shifts, sometimes toward robust growth, other times into dangerous oscillations or collapse.

If you’d like to see the exact formal structure (without going full heavy math), the paper lays out the dynamics as a compact set of equations, and the authors emphasize that the model is designed to be modular—so you can swap in different training curves or gate shapes to test new scenarios. For those who want to dive deeper, the original work is a great place to start: Dynamics of Human-AI Collective Knowledge on the Web.

Growth Regimes: From Healthy Growth to Oscillations and Collapse

The model isn’t just a black-box forecast; it maps out different regimes the knowledge ecosystem can inhabit, depending on how you tune the levers.

• Healthy Growth

  • A balanced regime: human learning and AI enhancement reinforce each other. The archive grows, its quality stays high, and both humans and the model improve. This is the sweet spot where the two-way flow—humans learning from the archive and AI helping humans learn—co-evolves constructively.

• Inverted Flow

  • Here, AI dominates content creation, often with a looser gate. The archive swells with AI-generated material that may be lower in quality, and the model starts to degrade because it’s trained on this weaker foundation. Human skill can stall or decay since the archive no longer offers high-quality signals for learning.

• Inverted Learning

  • People lean too heavily on AI explanations, rather than self-studying the archive. Even if AI content is decent, humans don’t build enough internal expertise, so the archive’s quality and future usefulness can suffer. It’s a subtle version of the same problem: over-reliance on AI erodes the human backbone of the knowledge ecosystem.

• Oscillations

  • If you let the system swing too loosely (e.g., high AI content, strong incentives for AI-driven learning, variable gate behavior), you can get boom-and-bust cycles: a spike in queries leads to more AI training, which boosts model skill, which fuels more questions, and so on—until quality and human skill slip and the cycle reverses.

• Collapse

  • The darkest scenario: a self-reinforcing loop where low-quality AI content crowds out human expertise, the model loses performance, and even more low-quality content keeps flooding the archive. This can degrade both the archive and the AI, creating a downward spiral.

The authors run numerical experiments to show how small changes in gate strictness, training mixes, or learning pathways can push the system across these boundaries. They also present two domain-style case studies—PubMed and GitHub Copilot—that illustrate how real platforms can land in different regimes under a shared framework.

A notable real-world takeaway from their simulations: even with inverted flows (AI-heavy content), a strong RLHF emphasis (more feedback-driven training) can rescue model performance and prevent outright collapse. The balancing act is delicate, but the model highlights where policy and platform design can intervene to maintain resilience.

If you want to see how these ideas map to real data, the authors calibrate the model with Wikipedia’s flow in two eras—pre-ChatGPT and post-ChatGPT—showing a clear rise in AI additions and a concurrent drop in human inflow after ChatGPT entered the scene. The calibration suggests tangible shifts in the ecosystem’s balance when major AI players enter the arena. For readers who love concrete numbers and plots, the paper provides those details and you can explore them in depth in the original work.

For a sense of how this matches up with practical platforms, check the PubMed and GitHub Copilot case studies in the paper. They demonstrate two different steady states: biomedical research tends to converge toward higher quality with more modest AI input, while open-source software environments often see larger archives and more active human skill growth but stabilize at a somewhat lower quality. The key is that the same model can illuminate different regimes depending on context and moderation norms.

If you’d like a direct read while you’re thinking about policy implications, the paper’s discussion about the role of the admission gate (the quality gate for AI content) and the mix of training signals (corpus-driven vs. RLHF) provides a practical blueprint for platform designers: tighten gating to protect quality, invest in RLHF to keep models honest, and support human learning pathways that emphasize archive study.

Real-World Case Studies: PubMed, GitHub Copilot, and Wikipedia

The authors don’t stop at theory; they tie the model to concrete platforms to show how its regimes actually play out.

• PubMed (Biomedicine)

  • In their PubMed configuration, the model assumes a relatively conservative gate and a slower knowledge-decay rate. This leads to a scenario where the archive quality can rise quickly when you tune for quality, while AI inflows are present but not overwhelming. The result is a healthy convergence of AI-assisted growth with strong human scholarly contributions. This aligns with expectations in medicine, where high accuracy and verifiability are critical and moderation norms tend to be strict.

• GitHub & Copilot (Open-Source Software)

  • In a software-development context, the model uses a larger AI inflow and a faster knowledge turnover. Copilot-like AI content can dominate the archive, but software knowledge decays faster, so quality can stabilize at a reasonable level if gate thresholds and training emphasis are tuned. The takeaway here is that in fast-moving technical domains, you can still preserve usefulness and growth by balancing AI content with robust human curation and by prioritizing training signals that keep the model aligned with human coding norms.

• Wikipedia (Pre-ChatGPT vs. Post-ChatGPT)

  • The calibration on Wikipedia reveals a shift: pre-ChatGPT era showed almost no AI inflow (AI content effectively off), with growth driven by human contributions. Post-ChatGPT era shows a substantial AI inflow and a reduced marginal impact of human input. This mirrors a broader real-world trend: AI is now a more active player in knowledge production, but the system’s health depends on how we manage gatekeeping and learning pathways to keep quality from eroding.

Across these snapshots, a common thread emerges: the same five-variable model can explain why different domains drift toward distinct steady states. The design choices—admission gate strictness, the balance of training methods, and how people learn—shape whether a platform stands at a high-quality, high-growth equilibrium or slips into riskier territory.

If you want to explore the data-fitting side, the authors present a robust calibration approach for Wikipedia (and robustness checks for the other domains) showing how the model’s parameters can be estimated from observed monthly changes in new pages, editors, and user views. It’s a nice example of turning a theoretical dynamical system into a practical auditing tool for real-world platforms.

For more, you can revisit the original paper to see the detailed parameter values and the exact fit results. And as always, the broader point remains: the model is a framework for policy-and-design experiments, not a one-size-fits-all forecast.

Design Levers for Sustainable Growth

One of the most valuable parts of this work is its actionable guidance for anyone responsible for a platform or policy that touches the web’s knowledge commons.

• Gatekeeping the archive (the a0 gate)

  • A stricter gate helps maintain quality by filtering out low-accuracy AI content before it becomes entrenched in the archive. The authors show that, even in inverted-flow scenarios, a strong gate can prevent a downward spiral and stabilize quality while still allowing meaningful AI benefits.

• Training mix: corpus-driven vs. RLHF

  • The model highlights a practical lever: biasing training toward human feedback (RLHF) can protect the system from model quality collapse when AI content is growing too fast. This is particularly relevant as more platforms push for rapid AI iteration. The key takeaway: keep a healthy share of RLHF in the training mix, especially when data quality or human signal strength is uncertain.

• Human learning pathways

  • Balance matters. If humans rely too much on AI explanations (over-reliance on LLM answers) without sufficient self-study, human skill can stagnate. Encouraging learning from the archive (self-study) in tandem with AI-assisted learning creates a more resilient loop.

• Domain-specific tailoring

  • The PubMed vs. GitHub Copilot contrasts show that different knowledge domains will settle into different regimes. Platforms should consider field-specific decay rates, domain norms, and the value placed on human-led vs. AI-assisted learning. A one-size-fits-all policy is unlikely to achieve long-term resilience.

• Platform-level audits

  • The calibration approach used for Wikipedia demonstrates that the model can act as an auditing tool. Regularly fitting the model to current data could help platforms anticipate tipping points and adjust gates, training emphasis, or learning pathways before problems become critical.

In short: sustainable growth isn’t about maximizing AI content or maximizing human effort in isolation. It’s about finding the right balance among gate quality, the training mix, and the ways people learn, all tuned to the specific domain and the platform’s goals.

For a concise pointer back to the source as you think through policy or product decisions, the original work remains the best reference: Dynamics of Human-AI Collective Knowledge on the Web.

Key Takeaways

• The web’s knowledge system can be understood through five interacting pieces: archive size (K), archive quality (q), AI model skill (θ), human skill (H), and query volume (Q). Their interactions drive the system’s growth or decline.

• Healthy growth arises when AI and humans feed the archive in a balanced way, with learning pathways that combine self-study and AI-assisted insights, and with thoughtful training that leverages human feedback.

• Risks include inverted flow (AI dominating content) and inverted learning (over-reliance on AI), which can trigger quality erosion, skill degradation, and even model collapse. The model shows how small policy changes can avert or trigger these regimes.

• Gatekeeping the archive (quality gates) is a powerful lever. Stricter gates can stabilize quality and prevent collapse, even when AI content inflows are high.

• Training strategy matters. Increasing RLHF and tuning the corpus-driven vs. feedback-driven learning mix can mitigate risks when AI content grows rapidly.

• Real-world case studies (PubMed, GitHub Copilot, and Wikipedia) illustrate that different domains settle into different steady states. The same framework helps compare platforms and design targeted interventions.

• The approach offers a practical auditing tool for platform designers and policymakers to forecast tipping points and test “what-if” scenarios before implementing changes.

• This is a timely framework for the AI era, providing a shared lens to discuss model reliability, content quality, and human expertise as the web evolves.

Sources & Further Reading

• Original Research Paper: Dynamics of Human-AI Collective Knowledge on the Web: A Scalable Model and Insights for Sustainable Growth

  • https://arxiv.org/abs/2601.20099

• Authors:

  • Buddhika Nettasinghe
  • Kang Zhao

If you’re curious to dive deeper into the math, the paper also includes detailed equations, calibration methods, and robustness checks that ground these ideas in data. But the big-picture takeaway is clear: the future of knowledge on the web depends on thoughtful balance—between what we admit, how we train intelligent systems, and how people keep learning.