🧬 biology

Learning epidemics: a compartmental and network approach to how artificial intelligence literacy spreads inside organizations

This paper reframes the spread of artificial intelligence literacy within organizations as a contagion process, demonstrating through network epidemiology and machine learning that learning retention is primarily driven by structural communication network properties and an individual's position within them rather than just personal attributes.

Original authors: Onicio Batista Leal Neto, Julia Nicolau, Alexandre Santilli

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Onicio Batista Leal Neto, Julia Nicolau, Alexandre Santilli

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your company as a giant, bustling city. Usually, when a new tool like Artificial Intelligence (AI) arrives, management thinks of learning as a formal school: they send out a memo, hold a training session, and expect everyone to learn at the same time.

But this paper argues that's not how learning actually happens. Instead, learning spreads like a contagion or a rumor through the city's streets. It moves from person to person during coffee breaks, quick chats, and "hey, look at this" moments.

The researchers treated the spread of AI skills like an epidemic. They asked: If we treat "knowing how to use AI" like a virus, how fast does it spread through the office, and why does it get stuck in some places but fly in others?

Here is the story of what they found, broken down into simple concepts:

1. The Three Cities (Organizations)

The study looked at three different companies (let's call them City A, City B, and City C) over three different times. They mapped out who talked to whom about AI.

City A was like a well-connected metropolis. Most people knew each other, and there was a huge "downtown" where almost everyone was connected.
City B was a bit more scattered, with a few busy hubs and many quiet neighborhoods.
City C was a collection of isolated islands. People mostly only talked to one or two specific colleagues, and those groups didn't talk to anyone else.

2. The "Reproduction Number" (R0)

In epidemiology, scientists use a number called R0 to predict if a disease will spread.

If R0 is below 1, the disease dies out (one sick person infects fewer than one other person).
If R0 is above 1, the disease spreads (one sick person infects more than one other person).

The researchers calculated a "Learning R0" for each city. This number told them how good the company's communication network was at spreading new skills.

City A had a high R0 (2.92). The network was perfect for spreading ideas.
City B was in the middle (1.93).
City C was barely above the danger line (1.14). The network was so broken that ideas struggled to travel.

The Result: The company with the best network (City A) kept the most people engaged in learning. The company with the broken network (City C) lost almost everyone. The structure of the network predicted the success of the learning program better than the training content itself.

3. The Shrinking, Densifying Group

Here is the surprising part. As time went on, everyone lost people. People stopped responding to the surveys or stopped talking about AI. The networks got smaller.

But, the people who stayed started talking to each other more often.

Imagine a big, loose party where people drift away. The people who stay end up huddled in a tight circle, talking intensely.
The "learning network" didn't grow; it consolidated. It shrank into a tight-knit core.
In the broken city (City C), the network shrank so much that it became a few tiny, isolated huddles. In the connected city (City A), the core remained strong and dense.

4. Who Stays? It's Not Just About the Person

The researchers used computer models to figure out who stayed engaged and who left. They expected that maybe the smartest people, the most senior people, or the most educated people would stay.

They found something different:

It's about your "address" in the network: The most important factor was Closeness Centrality. This is a fancy way of saying: How close are you to everyone else in the office? If you are in the middle of the conversation flow, you are more likely to keep learning.
It's about your "neighborhood": The Department you work in mattered more than your job title. Some teams had a culture of sharing; others didn't.
The "Tenure" Puzzle: How long you've worked there (tenure) didn't have a simple "the longer, the better" relationship. It was non-linear. This means both very new employees and very old employees behaved differently than those in the middle. The computer models could see this complex pattern, but a simple math line could not.

5. The Big Takeaway

The paper concludes that learning is a structural property.

You can't just buy a new technology and expect it to stick. If your office is like City C (isolated islands), no amount of training will help because the "roads" between people are broken. You need to build bridges first.

If your office is like City A (a connected city), the learning will spread naturally, but you need to protect the "downtown" (the central connectors) because if they leave, the whole system could collapse.

In short: To make your organization learn, don't just focus on the individual employee. Focus on the map of who talks to whom. If the map is broken, the learning won't travel. If the map is connected, the learning will spread like a good rumor.

Technical Summary: Learning Epidemics – A Compartmental and Network Approach to AI Literacy Diffusion

Problem Statement
Organizations increasingly rely on the rapid diffusion of new competencies, yet traditional models of workplace learning often treat the individual as the unit of change, neglecting the communication structures that facilitate knowledge transfer. This gap is particularly acute with generative Artificial Intelligence (AI), where literacy is not a fixed curriculum but a moving target acquired through informal, sideways interactions. The authors argue that the spread of AI literacy functions as a "complex contagion," requiring reinforcement from multiple sources rather than a single exposure. Existing literature lacks an integrated, empirical treatment that measures real communication networks, fits explicit epidemic models to competency spread, and derives interpretable reproduction numbers ( $R_0$ ) across multiple organizations to compare structural outcomes.

Methodology
The study employs a longitudinal, multi-organization design involving three institutions (Institutions 1, 2, and 3) of varying sizes (1,000 to 7,124 employees). Data was collected at three time points: baseline (T0), midline (T1), and endline (T2).

Network Construction: Participants completed an egocentric name-generator instrument, listing up to ten colleagues with whom they discussed AI in the previous 30 days. Directed ties were weighted by conversation frequency.
Social Network Analysis (SNA): Using Python (NetworkX), the authors computed standard descriptors (density, clustering, reciprocity, giant component size) and node-level metrics (degree, centrality measures).
Compartmental Modeling (SEIR): The authors adapted the Susceptible–Exposed–Infectious–Recovered (SEIR) model to learning:
- Susceptible: Unaware/not engaging.
- Exposed: Encountered but not using.
- Infectious: Actively applying and transmitting.
- Recovered: Habitual/consolidated use.
- Transmission ( $\beta$ ) and recovery ( $\gamma$ ) parameters were fixed ( $\beta=0.3, \gamma=0.2$ ) to ensure comparability across institutions.
Reproduction Number ( $R_0$ ) Estimation: Instead of fitting the model to adoption curves, $R_0$ was estimated analytically from network topology using the configuration-model expression: $R_0 = \tau(\langle k^2 \rangle - \langle k \rangle)/\langle k \rangle$ , where $\tau$ is transmissibility and $\langle k \rangle$ represents the moments of the degree distribution.
Machine Learning for Retention: To predict which individuals remained engaged across all waves, the authors pooled data ( $N=770$ ) and trained Random Forest and Gradient Boosting classifiers. Predictors included socio-demographics (tenure, department, role) and baseline network position. Feature importance was analyzed alongside linear point-biserial correlations to detect non-linear relationships.
Simulation and Visualization: A browser-based research console was developed to allow stakeholders to visualize network dynamics, run SEIR simulations on real and synthetic topologies, and explore counterfactual scenarios (e.g., turnover, onboarding).

Key Results

Structural Heterogeneity: The three organizations exhibited distinct baseline network topologies. Institution 1 formed a large, connected giant component (79.6% of nodes); Institution 2 was more distributed with a smaller core; Institution 3 was highly fragmented, dominated by isolated reciprocated pairs.
Network Contraction and Consolidation: Across all three organizations, the discussion networks contracted in size (fewer nodes and edges) over time but became denser. This suggests that while the broader population disengaged, the remaining participants formed tighter, more cohesive clusters.
$R_0$ and Retention Correlation: The baseline $R_0$ values were 2.92 (Inst. 1), 1.93 (Inst. 2), and 1.14 (Inst. 3). This ordering perfectly matched the retention rates of participants across all three waves (44.4%, 28.3%, and 11.4%, respectively). Organizations with higher structural capacity to spread learning retained more learners.
Predictors of Retention: Machine learning models identified Department and Closeness Centrality as the strongest predictors of retention. Notably, Tenure showed a near-zero linear correlation with retention ( $r = -0.001$ ) but was a top feature in non-linear tree models, indicating a complex, non-monotonic relationship where both very new and very experienced employees may behave differently from the middle.
Centrality Metrics: Closeness centrality was the most informative network metric. Eigenvector centrality was uninformative due to network fragmentation, collapsing toward zero in disconnected components.

Key Contributions

Reframing Learning as Contagion: The study successfully applies network epidemiology tools to organizational learning, treating AI literacy as a complex contagion dependent on network cohesion and redundancy.
Topological $R_0$ for Learning: The authors provide a method to estimate a basic reproduction number for learning directly from network topology without needing to fit adoption curves, offering a structural metric for an organization's learning capacity.
Non-Linear Insights: By combining tree-based machine learning with linear correlations, the study reveals that individual attributes like tenure influence learning retention in non-linear ways that traditional statistical methods might miss.
Interactive Tooling: The development of a self-contained research console allows non-technical stakeholders to interact with network data, run simulations, and visualize diffusion dynamics without executing the underlying analytical pipeline.

Significance and Claims
The paper claims that whether an organization learns to use new technology is, to a substantial degree, a structural property of its communication network rather than merely an aggregate of individual efforts. The alignment between the baseline topological $R_0$ and eventual retention suggests that network structure dictates the ceiling of learning diffusion.

The authors argue that this framework allows organizations to move from diffuse intuitions about culture to measurable quantities. Specifically, the $R_0$ can serve as a diagnostic tool: in fragmented networks (low $R_0$ ), the priority should be building bridges between isolated groups before adding content; in cohesive networks (high $R_0$ ), the focus should be on protecting the connected core. The study concludes that while individual attributes matter, they often operate through interactions with network position and local context, necessitating approaches that capture non-linear structural relationships. The authors remain modest, noting that the $R_0$ -to-retention alignment is based on three organizations and should be treated as a hypothesis for further testing, particularly given the limitations of survey coverage differences across the sample.