Your News Source's Friends Reveal Its Reliability: How Network Structure Exposes Unreliable Information
When Your News Source Hangs Out With Bad Actors
In the summer of 2024, a particular corner of Telegram was sharing articles from three dozen different websites. Most people scrolling past wouldn't have thought twice. But researchers looking at the same data saw something striking: nearly all of those sites were unreliable. Not coincidentally unreliable — structurally, mathematically unreliable. They clustered together in a web of shared audiences the same way reliable outlets clustered with other reliable outlets. The websites a person reads, it turns out, reveal more about source quality than the words on the page.
This is the central insight of new research from a team at the University of Konstanz, the Max Planck Institute for Dynamics and Self-Organization, and the Complexity Science Hub Vienna. Rather than analyzing what news articles say, they analyzed where news articles travel — specifically, how often domains appear together in the same Telegram conversations. Their findings, published on arXiv, suggest that the company your news source keeps may be the most reliable signal of all.
The researchers built a network of 45,574 news domains connected by their co-sharing patterns across 24,658 Telegram chat groups. They then asked a simple question: could the structure of that network — the patterns of who links to whom — help identify unreliable sources? The answer, according to their analysis, is yes. Graph Neural Networks, which are algorithms designed to learn from relational data, consistently outperformed traditional machine learning approaches when given the same information about each domain. GraphSAGE, one of the architectures they tested, achieved 63% accuracy in classifying domains as reliable, questionable, or unreliable. That's a 14% relative improvement over a network-unaware baseline running on identical features.
But perhaps more provocatively, the network worked even when stripped of all content. When the researchers removed the actual text of articles and kept only metadata about how articles spread — how viral they became, how many chats shared them, how bursty their sharing patterns were — the Graph Neural Network still achieved 53% accuracy. The baseline dropped to 47%. The network structure alone added a meaningful edge, even without reading a single headline.
The Landscape of Digital Information
To understand why this matters, you need to appreciate how broken the landscape has become. The Reuters Digital News Report 2024 found that more than a third of the world's population now consumes news through social media. For younger demographics, that number is substantially higher. These platforms have become primary information sources rather than supplementary ones, and they operate with a fundamental tension at their core: they're designed to maximize engagement, not accuracy.
Algorithms optimized for engagement prioritize content that provokes strong emotional responses, that confirms existing beliefs, that keeps people scrolling. Quality is incidental. A fabricated story that generates outrage will outcompete a carefully reported piece that challenges assumptions. This structural incentive means that unreliable information doesn't just survive online — it often thrives.
The problem isn't just quantity. It's that the visual and linguistic markers that once distinguished credible journalism from partisan screed have eroded. In 2016, "fake news" was Oxford's word of the year. In 2017, "post-truth" followed. Since then, the sophistication of low-quality content has only increased. Modern unreliable sources often mimic the formatting, tone, and structure of professional journalism so closely that even trained fact-checkers struggle to distinguish them. And with the rise of large language models, the cost of producing plausible but fabricated content has dropped to nearly zero. A bad actor can now generate thousands of convincing news articles per hour, each one indistinguishable by style alone from legitimate reporting.
This has put traditional content-based detection methods in an increasingly untenable position. Machine learning classifiers that analyze text for signs of unreliability — sensationalist language, lack of sources, unusual writing patterns — face a moving target. The moment they learn to flag a pattern, unreliable producers adjust. The arms race between detectors and generators may, in the long run, favor the generators.
Telegram exemplifies these dynamics. With over a billion active users, it's one of the world's largest messaging applications. Unlike platforms with algorithmic feeds and follower-based networks, Telegram centers on public channels and group chats where administrators broadcast to large audiences. The platform's emphasis on privacy and encrypted communication has made it attractive to users who distrust mainstream authorities. In Russia, after Telegram was blocked in 2018, founder Pavel Durov framed the resistance as "Digital Resistance," cementing the app's reputation as an alternative information ecosystem where unverified or unreliable content can circulate freely.
Previous research has shown that URLs from lower-quality sources are shared more frequently on Telegram than content from professional news media. Users who spread lower-quality information tend to be more active contributors than those sharing higher-quality content. The platform has become, in effect, a petri dish for understanding how unreliable information propagates through social networks.
Building the Network
The researchers began with data: nearly 100 million URL-sharing events across Telegram chat groups, with about 95% of the activity occurring between 2017 and 2023. This was collected through snowball sampling, starting from known Telegram communities and expanding outward to capture their connections.
From this raw material, they constructed something more interesting. Each shared URL represents a signal — a user, in a particular chat, thought that piece of content was worth spreading. Individually, these signals are noisy. But in aggregate, they encode something meaningful about the relationships between news sources.
The process worked like this: two domains are connected if they've been shared in the same Telegram chat at a statistically significant rate. To determine statistical significance, the researchers used the Bipartite Configuration Model (BiCM), which generates random versions of the network while preserving the overall activity patterns of chats and domains. If two domains appear together more often than would be expected by chance, given how popular each one is, then the connection is kept. If their co-occurrence could easily be explained by popularity alone, it's discarded.
This filtering is crucial. A naive approach would connect any domains that ever appear in the same chat, but that would create a misleading picture. A domain that publishes hundreds of articles per day will inevitably overlap with many other domains by sheer volume. The BiCM correction strips away these spurious connections, leaving only relationships that reflect genuine audience overlap.
Of the 4,884 domains in the training set, 93.7% survived this validation — 4,574 domains remained in the final network. These nodes were connected by 350,883 validated edges, giving an average degree of about 153. Each domain, on average, is significantly co-shared with roughly 153 other domains. The network is sparse, with a density of only 0.03 — meaning that any given pair of domains has only a 3% chance of being directly connected — but highly structured.
That structure reveals itself in a property called assortative mixing. In social networks, assortativity refers to the tendency of similar nodes to connect. Homophily, the principle that "birds of a feather flock together," is a form of assortativity. The researchers measured this tendency with respect to reliability and found a coefficient of 0.22 — statistically significant by any reasonable standard (p < 0.001 in a permutation test with 5,000 iterations).
What this means practically: if you know that a given domain is unreliable, you can predict with reasonable confidence that its neighbors in the network will also be unreliable. The same holds for reliable domains. Low-reliability sources cluster together, as do high-reliability ones. The network topology itself encodes reliability signals that have nothing to do with what the articles say.
The network visualization makes this pattern viscerally clear. When you color nodes by reliability score and look at the top 50% by connectivity, the clusters practically organize themselves. Light blue nodes — unreliable sources — huddle in distinct regions. Darker nodes — reliable sources — occupy different territories. The graph isn't randomly colored. It has structure, and that structure corresponds to quality.
Features: What the Models Saw
To train their classifiers, the researchers needed to describe each domain in the network. They developed two types of features: content-aware and content-agnostic.
Content-aware features came from the actual text of articles. The researchers scraped articles from each domain and processed them through a multilingual embedding model — paraphrase-multilingual-MiniLM-L12-v2 — which converts text into numerical vectors that capture semantic meaning. Each domain's content feature was the average of its article embeddings. This approach captures what domains are actually saying, in a form that machine learning models can process.
Content-agnostic features, by contrast, describe how content spreads rather than what it says. These include metrics like the magnitude of sharing activity, the "burstiness" of temporal patterns (how much sharing activity comes in sharp spikes versus steady trickles), and the total number of messages and unique chats that shared each piece of content. These features are proxies for reach and visibility — how far did this content travel, and how did it travel?
Both feature sets were aggregated at the domain level and standardized using z-score normalization, meaning each feature was scaled to have a mean of zero and a standard deviation of one. This ensures that features on different scales (like raw sharing counts versus normalized burstiness scores) contribute appropriately to the models.
The Models: Graph Neural Networks vs. Baselines
The core question the research addresses is whether network structure adds predictive value. To answer this rigorously, the researchers needed to isolate the contribution of network topology from the contribution of node features. Their solution was elegant: train the same features on different model architectures, some of which use the network and some of which don't.
The baseline was a Multi-Layer Perceptron (MLP) — a standard neural network that processes each node's features independently. Given the features for domain X, the MLP predicts whether X is reliable, questionable, or unreliable. It never sees the network. It never knows what other domains X is connected to. This is essentially what any content-based classifier does: make predictions based on what you can extract from the item itself.
Against this baseline, the researchers tested three Graph Neural Network architectures: Graph Convolutional Network (GCN), Graph Attention Network (GAT), and GraphSAGE. These models share a common intuition — that a node's neighborhood contains information — but implement it differently.
GCNs propagate information through the network by averaging features from neighboring nodes. Each layer of the network allows information to travel one hop further, so a three-layer GCN can integrate information from nodes up to three connections away. GATs add attention mechanisms, learning to weight neighbors by their importance — a domain connected to 100 reliable sources matters more than one connected to 100 unreliable ones. GraphSAGE takes a different approach entirely: rather than averaging neighbor features, it samples a fixed number of neighbors and aggregates their information, which makes it better suited for generalizing to unseen nodes.
All models were implemented using PyTorch Geometric, with three layers, ReLU or ELU activations, dropout regularization, and Adam optimization. Hyperparameters were tuned per model to maximize classification accuracy.
What the Models Found
The results were consistent and striking. Across both feature settings — with content and without — the Graph Neural Networks outperformed the MLP baseline. And within the GNN family, GraphSAGE was the clear winner.
Classification Accuracy With Content Features
| Label | Value |
|---|---|
| GraphSAGE | 63 |
| GCN | 57 |
| GAT | 57 |
| MLP Baseline | 55 |
| Random | 36 |
With content features available, GraphSAGE achieved 63% accuracy and a 63% F1-score. The MLP baseline, running on identical features, reached only 55% accuracy and 55% F1-score. That's an 8 percentage point absolute improvement, or about 14% relative — a substantial margin in a three-class classification problem. GraphSAGE correctly identified 101 out of 171 unreliable domains in the test set. The other GNN architectures, GCN and GAT, tied at 57% accuracy.
The advantage persisted even without content. When the researchers removed the semantic text embeddings and kept only spreading dynamics, GraphSAGE achieved 53% accuracy versus 47% for the MLP baseline. The relative improvement was similar — roughly 13% — even though absolute performance dropped for both. GCN and GAT fell to 50% accuracy, still outperforming the content-free MLP but trailing GraphSAGE.
A random baseline, which would simply guess uniformly across the three classes, would achieve 36% accuracy by chance. All models substantially exceed this floor. A simple network classifier, which predicted the reliability class based only on the majority class among a node's neighbors, achieved 38% — barely above random. This illustrates that exploiting the network's structure through GNNs adds substantial value over naive neighbor-based heuristics.
Classification Accuracy Without Content Features
| Label | Value |
|---|---|
| GraphSAGE | 53 |
| GCN | 50 |
| GAT | 50 |
| MLP Baseline | 47 |
| Random | 36 |
The performance drop between content-based and content-agnostic settings tells an important story. GraphSAGE's accuracy fell from 63% to 53% — a 10 percentage point drop — when text was removed. But the MLP's drop was steeper: from 55% to 47%, an 8 percentage point fall. The GNN's content-agnostic performance (53%) substantially exceeds the MLP's content-based performance (55%) — the difference is not large, but it's suggestive. The structural signal is powerful enough that a model without content nearly matches one with content but no network awareness.
This matters because content-agnostic features are scalable in ways that content analysis is not. Scraping, cleaning, and embedding article text requires resources and faces legal and technical obstacles. Spreading dynamics can be computed from sharing metadata alone, which is often more readily available. And spreading dynamics are harder to manipulate than content. A bad actor can easily publish credible-sounding text; it's much harder to manufacture the organic sharing patterns of a legitimate news source.
The Assortativity Signal
The key theoretical finding of this paper is the assortativity — the tendency of similar-reliability domains to cluster. The coefficient of 0.22 is high by the standards of network science. Real-world networks often show weak assortativity or even disassortativity (where high-degree nodes connect to low-degree nodes, for example). An assortativity of 0.22 for a continuous variable like reliability suggests strong homophily in sharing behavior.
Why does this pattern exist? The researchers offer a behavioral explanation: users tend to consume and share from sources aligned with their existing worldviews. Someone who follows one conspiracy-adjacent site likely follows others. Someone who trusts mainstream outlets tends to stick to that ecosystem. These preferences manifest in sharing behavior, which creates the co-sharing network.
This isn't just correlation. The researchers' interpretation is causal: sharing behavior causes the clustering, not the other way around. It's not that unreliable sources deliberately cluster to avoid scrutiny (though they sometimes do). It's that users self-sort into information ecosystems, and the network captures that sorting.
The assortativity is robust. It survived permutation tests at p < 0.001, with a z-score of 149.4 — meaning the observed coefficient was 149 standard deviations above the mean of permuted networks. This is not a marginal finding.
The network visualization in Figure 3 reinforces this. When you restrict to the top 50% of nodes by degree — the most connected, and thus most influential, domains — the clustering is visually apparent. Light blue (unreliable) domains form distinct communities. Darker (reliable) domains form others. The clusters are not perfectly separated — there are bridges between them, and some domains sit in mixed neighborhoods — but the overall structure is clear.
This structure is exploitable precisely because it encodes the behavioral signal. If unreliable domains are more likely to be neighbors of other unreliable domains, then a model that knows your neighbors can make better predictions about you than one that knows only your own features.
Why This Changes Things
The implications of this research extend beyond the specific empirical findings. The study demonstrates a principle: that relational structure can compensate for the weaknesses of content-based detection.
Content-based methods have an inherent vulnerability: they assume that unreliable content looks different from reliable content. This assumption is increasingly false. Modern unreliable sources can produce text that is grammatically correct, stylistically consistent, and superficially sourced. Large language models make this trivially easy. A classifier trained to detect the linguistic signatures of low-quality journalism will increasingly find those signatures absent.
Network-based methods don't rely on content quality. They rely on social behavior — who shares what, with whom, how often. These patterns are harder to fabricate. A bad actor can create a website and populate it with convincing text. They cannot as easily manufacture the organic sharing behavior of a community of believers. The network structure is a byproduct of human behavior, and human behavior is harder to spoof than text.
This doesn't mean content analysis is obsolete. The study shows that combining content features with network structure produces the best results (63% accuracy with content versus 53% without). But it means that network structure can serve as a fallback when content analysis fails or is infeasible. For languages where embedding models perform poorly, for domains where scraping is blocked, for cases where generative AI has made content indistinguishable — the network still works.
The content-agnostic approach also offers practical advantages. Text embedding requires crawling, scraping, and processing — operations that are resource-intensive, legally contested, and easily disrupted. Spreading dynamics can be computed from sharing metadata alone, which is more readily available and less intrusive. A platform like Telegram could, in principle, compute spreading metrics for domains shared in its chats without ever analyzing article content. This makes network-based approaches more deployable in real-world systems.
The research also validates a shift in focus from articles to domains. Individual articles are hard to assess; they may be one-off fabrications, satirical pieces, or outlier opinions that don't reflect a source's overall reliability. Domains represent aggregate reliability — the reputation of an outlet built up over years of publication. A single article from a reliable outlet might be wrong; the outlet's overall track record is more informative. By classifying domains rather than articles, the researchers capture this aggregate signal.
The Telegram context is also significant. While Telegram is not the only platform where unreliable information spreads, it is a prominent one, and its structure — public channels, group chats, focus on privacy — makes it a useful laboratory. The findings may not generalize perfectly to Twitter/X, Facebook, or TikTok, each of which has different architectures and user behaviors. But the underlying principle — that sharing behavior encodes reliability signals — likely extends beyond any single platform.
The Limits of the Signal
A 63% accuracy figure requires context. In a three-class problem, random guessing would achieve 36% accuracy. The GraphSAGE model more than doubles the random baseline, which suggests substantial predictive power. But it also means that roughly 37% of predictions are wrong, and we don't know how errors are distributed.
The confusion matrices in the paper (referenced in supplementary materials) would reveal whether errors are symmetric or biased. If the model systematically misclassifies unreliable domains as questionable rather than reliable, that might be acceptable — the harm of a false positive (marking a reliable source as unreliable) may be different from the harm of a false negative (missing an unreliable source). The paper notes that false negatives are particularly costly in misinformation detection, since incorrectly labeling fake news as true can facilitate its spread. GraphSAGE generates fewer false negatives than the MLP, which is encouraging.
The class distribution is imbalanced. About 42% of domains in the training set are reliable, 43% are questionable, and only 15% are unreliable. This imbalance means that baseline accuracy is inflated — a model that simply guessed "questionable" for everything would achieve 43% accuracy. GraphSAGE's 63% accuracy represents a genuine improvement over this trivial baseline, but the imbalanced distribution may affect the model's calibration across classes.
The ground truth labels come from a dataset compiled by Lin et al., which aggregates six fact-checking databases through principal component analysis. This approach has strengths — it combines multiple assessments to reduce individual biases — but also weaknesses. Fact-checking organizations may disagree, and the PCA aggregation may not perfectly represent the truth. A domain that one organization marks as unreliable and another marks as reliable might receive a middle score that doesn't reflect either assessment accurately. The researchers acknowledge this limitation by treating reliability as a spectrum rather than a binary classification.
The network construction also has limitations. The Bipartite Configuration Model validation is conservative — it may remove genuine connections that appear strong by naive measures. But it may also fail to capture weak but real relationships that the model treats as noise. The choice of statistical threshold affects the network's density and, consequently, the GNN's ability to propagate information through it.
Telegram's specific culture and user base limit generalizability. Telegram users may be systematically different from users of other platforms — more privacy-conscious, more ideologically motivated, more likely to seek alternative information sources. The assortativity pattern might be stronger or weaker on platforms with different user bases.
Finally, the temporal scope matters. About 95% of the data comes from 2017-2023, a period that includes the COVID-19 pandemic, the 2020 US election, and the rise of generative AI. These events may have affected sharing patterns in ways that make historical patterns less predictive of future ones. A model trained on 2017-2023 data might perform differently on data from 2024 onwards.
What's Next
This research opens several avenues for future work. One immediate question is whether the approach generalizes to other platforms. Twitter/X, Facebook, Reddit, and TikTok all have sharing patterns that might encode reliability signals. A cross-platform study could test whether the assortativity observed on Telegram exists elsewhere, and whether GNN-based approaches transfer across platforms.
Another avenue is temporal dynamics. The current study treats the network as static, but sharing patterns evolve over time. A domain that was once reliable might become unreliable (or vice versa), and the network might capture these transitions before content-based methods detect them. Dynamic graph models that track changes over time could potentially predict reliability shifts before they manifest in content.
The performance gap between content-based and content-agnostic approaches suggests room for improvement in network-based methods. Better graph architectures, more sophisticated aggregation functions, or different ways of incorporating temporal information could narrow the gap. If content-agnostic methods could approach content-based performance, they would become highly practical for real-world deployment.
Causal mechanisms also deserve investigation. The current study establishes correlation — reliable domains cluster with reliable, unreliable with unreliable — but doesn't explain why. Understanding the behavioral and psychological drivers of this clustering could inform interventions. If users self-sort based on ideology, then reducing ideological segregation might reduce assortativity. If users are simply following recommendations from similar users, then disrupting recommendation algorithms might help.
The practical deployment question is thorny. Any system for automatically labeling news sources as unreliable faces risks of abuse, gaming, and unintended consequences. A bad actor could use such a system to falsely label competitors as unreliable. A government could use it to suppress legitimate criticism. The researchers acknowledge these risks but don't fully address them. Responsible deployment would require safeguards, appeals processes, and careful consideration of who controls the labels.
Finally, there's the question of what reliability means. The Lin et al. dataset treats it as a continuous score, but the boundaries between reliable, questionable, and unreliable are inherently fuzzy. A satirical website that publishes false stories might score low on reliability but not be malicious. A partisan outlet that publishes accurate information within a narrow ideological frame might score poorly on neutrality while being reliable on facts. The three-class discretization used in this study simplifies a complex reality, and downstream users of such classifications should be aware of what they're losing.
The Broader Significance
Beyond the specific findings, this research reflects a broader shift in how we think about information quality. For decades, the dominant approach to fighting misinformation focused on content: find the fake, expose it, remove it. Fact-checkers read articles and render verdicts. Algorithms look for linguistic patterns associated with unreliable content. This approach has value, but it treats the problem as fundamentally textual — as if the words themselves are the source of the problem.
The network-based approach offers a different framing. Instead of asking "Is this content true?", it asks "Where does this content come from, and who else pays attention to it?" This shifts the unit of analysis from the article to the source, from the statement to the community. It leverages the social context of information consumption rather than analyzing information in isolation.
This shift is, in some ways, a recognition of how humans actually assess credibility. When you encounter an unfamiliar news source, you probably don't read every article and evaluate its claims. You probably check who else reads it, whether reputable outlets cite it, what communities it belongs to. You use social proof as evidence of quality. The network-based approach automates this intuition.
It also acknowledges that information ecosystems are structured. They are not random collections of articles but organized communities of producers, consumers, and sharers. That structure contains information that content-based methods miss. A single article from an unreliable source might look identical to one from a reliable source. But if that source is embedded in a network of unreliable sources, that context reveals something that the article alone cannot.
The emergence of generative AI makes this approach more important, not less. As AI-generated content becomes indistinguishable from human writing, content-based methods will struggle. Network-based methods, which depend on the relational structure of sharing behavior, may prove more robust. A fabricator can mimic the style of credible journalism; they cannot easily mimic the organic sharing patterns of a credible community.
This doesn't mean network-based methods are foolproof. Bad actors could, in principle, create networks of coordinated sharing to manipulate the structure. They could build communities of seemingly reliable sources that actually spread misinformation. The arms race between detectors and manipulators will continue. But the structural signal adds a layer of defense that pure content analysis lacks.
Conclusion
The research by Keßler, Ventzke, Priesemann, and De Marzo demonstrates that network topology systematically improves domain reliability classification, even when content analysis is unavailable. Their approach — building a statistically validated co-sharing network from Telegram data and applying Graph Neural Networks — achieves meaningful gains over content-based baselines.
The key insight is assortativity: reliable and unreliable domains cluster separately in sharing networks, and that clustering is exploitable. GraphSAGE leverages this structure to achieve 63% accuracy with content features and 53% without. Both figures substantially exceed baselines and represent meaningful improvements in a difficult classification problem.
The practical implications are significant. Network-based methods can serve as a complement to content-based approaches, providing robust signals even when content analysis fails. They offer a path forward in a landscape where generative AI has made content increasingly hard to evaluate.
But the research also highlights how much remains unknown. Why do reliable and unreliable domains cluster? Can the approach generalize to other platforms? How should reliability labels be governed to prevent abuse? These questions will drive the field forward.
For now, the finding stands: your news source's friends tell you something important about its quality. The network is the message.