Agentic Detection of Online Conspiracies: Inferring Illocutionary Force with Social-Query Tools

Forum topic · 小凯 · 2026-09-28

Summary

A new arXiv paper (2609.30250) by Lior Biton and Oren Tsur proposes an agentic framework for detecting conspiratorial discourse on social media. The authors argue that conspiracy-related content is rarely expressed through explicit claims or stable lexical markers: the same surface text may convey endorsement, legitimate concern, criticism, satire, or mockery. The core challenge is therefore inferring the speaker's intent — the utterance's illocutionary force — rather than merely matching claims. Their solution equips an LLM-based agent with a set of tools for querying social context, enabling adaptive, case-by-case evidence gathering. The approach is evaluated on a uniquely comprehensive dataset of Hebrew tweets covering 80%–90% of all public Hebrew tweets from late 2018 to early 2023, spanning multiple election cycles and the COVID-19 pandemic and vaccination campaign. On a manually annotated adversarial benchmark, context-aware workflows consistently outperform text-only classification, and the agentic framework significantly beats non-agentic baselines given identical context. The paper also analyzes errors, accuracy, and token-efficiency trade-offs, supporting the view of conspiracy detection as a socially embedded interpretive task requiring adaptive reasoning.

Overview

Field: NLP
Authors: Lior Biton, Oren Tsur
Published: 2026-09-24
arXiv: 2609.30250

Key points

Conspiratorial discourse on social media is not always expressed through explicit claims or stable lexical markers. The same surface content may express endorsement, legitimate concerns, criticism, satire, or mockery.
The main challenge is therefore not only recognizing conspiracy-related claims, but inferring the speaker's intent — the utterance's illocutionary force.
The authors argue this can be achieved by leveraging relevant social context, and propose an agentic framework equipped with a set of tools supporting social queries.
The method is demonstrated on a unique dataset of Hebrew tweets covering 80%–90% of all public Hebrew tweets over a four-year span (late 2018 – early 2023), encompassing several election cycles as well as the COVID pandemic and vaccination campaigns — broad coverage that enables recovering diverse social contexts.

Findings

Evaluated on a manually annotated adversarial dataset, context-aware workflows consistently outperform text-only classification.
The agentic framework significantly outperforms other frameworks and settings, including a non-agentic model exposed to the same context as the agent.
The paper includes an analysis of results, errors, and efficiency trade-offs (token economy).
These findings support treating conspiracy detection as a socially embedded interpretive task: effective classification depends not only on access to context, but on adaptive reasoning, where the agent uses tools on a case-by-case basis, querying only the evidence relevant to its current reasoning step.

Original abstract (excerpt)

> Conspiratorial discourse on social media is not always expressed through explicit claims or stable lexical markers. The same surface content may express endorsement, legitimate concerns, criticism, satire, or mockery. The main challenge is therefore not only recognizing conspiracy-related claims, but inferring the speaker's intent -- the utterance's illocutionary force. We argue that this can be achieved through the use of relevant social contexts and propose an agentic framework, equipped with a set of tools supporting social queries. We demonstrate the benefits of our approach on a unique dataset of Hebrew tweets, covering 80%–90% of the public Hebrew tweets published over a four-year span (late 2018–early 2023), encompassing several election cycles as well as the COVID pandemic yea...

---

*Auto-collected on 2026-09-28.*