The “Epstein Files” and the anatomy of hidden social networks
How secrecy reshapes structure and why its analysis is conditional on the data
If you find value in #ComplexityThoughts, consider helping it grow by subscribing and sharing it with friends, colleagues or on social media. Your support makes a real difference.
→ Don’t miss the podcast version of this post: click on “Spotify/Apple Podcast” above!
Disclaimer. Sometimes it’s worth stepping into the public discussion, especially when oversimplifications and unsupported claims start spreading faster than the evidence.
That’s the situation with the Epstein Files and the many “networks” circulating after their release. This post is not about speculating beyond the records and not about naming “the” key players. It’s about a methodological point from network science: how secrecy reshapes what becomes observable, why centrality is conditional on the data we actually have and, overall, why it is important to understand these simple facts to avoid misleading conclusions.
A network of power
The Epstein case pushed a disturbing question into the open: what does a “network of power” look like when it spans money, politics and influence?
And — crucially — how much of that network can any set of files actually reveal?
In complex systems, visibility is never neutral: and this is true especially for social networks. When relationships are risky to document, the information ecosystem becomes biased toward what is safe to record. Structurally, this creates partial observations of a larger network; functionally, it creates feedback loops that reward discretion, compartmentalization and deniability. The two images below illustrate a simple network-science point.
The above image shows a network of individuals (nodes) connected by relationships (edges). Clusters indicate densely connected groups, while bridges are nodes that link otherwise separate clusters. The red circles highlight nodes that appear “important” in the observed graph.
Alongside the solid nodes, we can also see a faint, blurry layer. Those nodes are not “less real”: they represent parts of the same system that are less visible in the data, connections that are harder to document, less frequently recorded or simply outside the boundary of what was observed (i.e., what the archive contains). In my schematic, the faint layer is a visual proxy for low-visibility parts of a system (ties and nodes that are under-recorded or outside the observed boundary). So the point is visual: even in one network, visibility can be uneven, and that unevenness shapes what we infer.
This second image “zooms” on one red node. In the context of the Epstein Files, you can think of this node as a reference anchor (for instance, Epstein in the released records) without implying that this captures the entire underlying network. The shading by distance (1–3 steps) highlights the methodological point: reconstructions from documents tend to reveal local neighborhoods around what is observable, not the full system.
Why observed graphs can be misleading?
First: “missingness” (i.e., what’s missing from the record) can be strategic, and often it is not random.
In settings where exposure is costly, the most sensitive ties are less likely to appear in documents or traces, so the data you can extract is systematically incomplete.
Second: secrecy trades off with coordination.
A useful lens is what network scientists call covert-like structures: networks operating under secrecy incentives, often balancing operational efficiency against secrecy by using intermediaries, buffers and compartmentalized subgroups. This can make the observed network look fragmented even if the underlying system is robust.
Third: centrality (i.e., who looks like a “hub” or any other measure of “influence”) is conditional on what you can see.
Measures like betweenness (i.e., who sits on many paths between groups) or “most connected” are properties of the measured graph. If key nodes/edges are missing, rankings can shift dramatically. So “central in the leak-derived network” is not the same as “central in the real system”. This is basic network science.
Beyond Epstein?
Large leaks can reveal huge structures and still be partial. The Panama Papers (I have even rebuilt and analyzed its network in the past) and the FinCEN Files are reminders that document-based networks illuminate major interfaces of systems, not necessarily their full extent. Those cases are methodologically similar in one respect: document-defined boundaries.
Network reconstructions built via human coding or AI can be extremely valuable for organizing evidence and generating hypotheses. The caution is about over-interpretation:
Good → “This person is central in this extracted dataset”
Risky → “This person is central in the whole network”
Similarly for groups (i.e., “clusters”):
Good → “These individuals form a group in this extracted dataset”
Risky → “These individuals form a group in the whole network”
So, how to do this responsibly (and still powerfully)? If you cannot refrain from making a network analysis, or if you cannot refrain from reading someone else’s network analysis, at pay attention to this:
Report centrality with uncertainty: run sensitivity checks (e.g., how rankings change if you add/remove plausible missing ties).
Triangulate across data layers: combine documents with public records, timelines, institutional roles and independent investigative sources (don’t treat one archive as the boundary of the system).
Use claims calibrated to the sample: “within-records” language avoids turning an inference tool into an accusation.
Take-home message
Leaks can meaningfully erode secrecy and expose parts of hidden networks. But in systems incentivized to avoid detection, what we reconstruct is often a biased slice. The scientific stance is neither cynicism nor certainty: it should be disciplined inference strong enough to guide investigation, humble enough to admit what the data cannot prove.
→ Please, remind that if you find value in #ComplexityThoughts, you might consider helping it grow by subscribing, or by sharing it with friends, colleagues or on social media. See also this post to learn more about this space.
I really appreciate the careful epistemology here, and the conclusion of "neither cynicism nor certainty". It's at the heart (or brain?) of scientific endeavours as a whole, but could be seen more often in scientific writing.