"This is Fake! Shared It by Mistake": Assessing the Intent of Fake News Spreaders

Key claims

Unintentional spreading is a distinct phenomenon. Unlike malicious actors who deliberately propagate falsehoods, ordinary users can share fake news believing it is true, or without evaluating its veracity. This group deserves education, not platform sanctions, and designing interventions requires distinguishing them from intentional spreaders and bots.

Intent can be computed from information about influence. The core insight is that unintentional spreading correlates with high influence—users who receive strong internal signals (alignment with past beliefs) and external signals (repeated exposure, social proof) are more likely to share unintentionally. A quantifiable "effective degree" (incoming edge weights in an influence graph) predicts whether a spread was intentional or unintentional.

Two types of influence drive unintentional spreading:

1. Internal influence: Alignment with a user's pre-existing beliefs, measured by semantic similarity to news they have previously shared (confirmation bias)
2. External influence: Social proof from repeated exposure or similar content; quantified by how many times the fake news (or similar variants) has been shared in the user's network

The influence graph model captures nuanced spreading patterns. The method:

• Treats each post as a node
• Weights edges by (a) content similarity (using pre-trained transformers), (b) temporal decay (using a forgetting curve), and (c) user-attitude similarity (whether the user endorsed the shared content)
• Excludes cases where a user repeatedly shares the same news (a strong signal of intentionality, per prior work)
• Aggregates incoming edge weights as the "effective degree"—a scalar metric of received influence

A semi-automated annotation method enables scalable dataset creation. Manual annotation of intent is labor-intensive and subjective. The researchers discovered that confident intentional labels align with two detectable signals: (a) malicious users (identifiable via bot-detection tools like Botometer), and (b) correction posts (where the user explicitly acknowledges the fake claim's falsity). By automating identification of these signals, they simulate manual annotation at scale with high agreement.

Intent predicts fake news detection performance. The computed intent (effective degree) ranks in the top 5 most important features (out of 100+) when used in XGBoost classifiers for fake news detection. A heterogeneous graph neural network incorporating the influence graph achieves 0.86 AUC on limited training data, outperforming variants without the influence structure.

Psychology and computation inform each other. The model grounds itself in cognitive psychology: confirmation bias and the illusory truth effect (repeated exposure → increased belief) are the primary mechanisms. Conversely, the computational design respects these mechanisms—e.g., using a forgetting curve because recent exposures matter more, excluding repetition-based intentionality signals, and modeling attitude-weighted similarity.

Method

Research questions addressed:

1. Why do users spread fake news unintentionally?
2. How can we compute intent from observable signals?
3. Can we scalably annotate intent in a dataset?
4. Does the computed intent metric improve fake news detection?

Influence graph construction:

• Nodes: all posts from a user
• Directed edges: from previous posts to the target post (the fake news sharing event)
• Edge weight = f(content similarity, attitude similarity, temporal decay)
• Effective degree = sum of incoming edge weights (quantifies received influence)

Similarity computation:

• Content similarity: pre-trained transformers (BERT, etc.)
• Multi-modal features available for images
• Deep ranking used to refine similarity scores

Temporal decay:

• Forgetting curve: e^(1 − time_interval)
• Recent influence weighted more heavily
• Normalizes effective degree (upper bound: e / (e−1))

Semi-automated intent annotation:

1. Detect malicious users: Botometer threshold 0.5
2. Identify correction posts: manual labeling (fast, since few posts are corrections)
3. Label all other posts as unintentional
4. Validates against manual annotation with high agreement (Cohen's kappa substantial/high; area under ROC high)

Dataset:

• 300 posts manually annotated by experts with intent and confidence
• 119 posts with agreement across annotators (high confidence set)
• Semi-automated method extends to thousands

Evaluation:

• Distribution analysis: unintentional spreaders have significantly higher effective degree than intentional spreaders
• Machine learning: XGBoost on 100+ features; effective degree in top 5
• Graph neural network: heterogeneous graph model on influence graph; 0.86 AUC with limited training data

Notes

This work makes a significant contribution by centering the question of intent—a practically important distinction that platform policy and intervention design depend on. Blocking or penalizing unintentional sharers is counterproductive; educating them is. The paper provides both theoretical grounding (confirmation bias, truth effect) and computational operationalization (influence graphs, effective degree) to automatically infer intent from behavior.

The semi-automated annotation mechanism is pragmatically important. Manual annotation of intent is subjective and costly; the observation that confident intentional labels cluster around malicious accounts and correction posts enables a scalable proxy that maintains high agreement with human judgment.

The use of a forgetting curve is methodologically sound and somewhat novel in this context. Misinformation research often treats all historical posts equally, but the psychological finding that older information decays in influence is well-established (Ebbinghaus curve). Incorporating this into the influence graph improves realism.

The graph neural network results (0.86 AUC with limited training data) suggest that structure matters: pairwise features alone underperform when the graph structure is incorporated. This supports the model's assumption that influence is a networked phenomenon, not just a set of individual features.

A limitation worth noting: the method assumes access to a user's full sharing history and their attitudes toward shared content. In practice, deleted posts, private accounts, and incomplete user histories may limit applicability. The paper would benefit from discussing graceful degradation under incomplete data.

The work opens avenues for follow-up research on intent classification (beyond binary intentional/unintentional), cross-platform influence propagation, and interventions tailored to inferred intent (e.g., education for unintentional sharers vs. account restrictions for malicious actors).