Skip to content

Back to the catalog

Agentic Detection of Online Conspiracies
arxiv.orgpaper

Agentic Detection of Online Conspiracies

This arXiv paper proposes an agentic framework for detecting conspiratorial discourse on social media, where an LLM agent uses a set of "social query" tools to gather context (e.g., speaker history, reply threads) to infer illocutionary intent rather than relying on surface lexical cues. Evaluated on a large corpus covering 80-90% of Hebrew tweets over 2018-2023 (spanning elections and COVID), the context-aware agentic approach significantly outperforms text-only classification and a non-agentic model given the same contexts, with analysis of errors and token-cost tradeoffs.

TL;DR An LLM agent that queries social context (not just text) to detect conspiratorial intent in Hebrew tweets, beating text-only and non-agentic baselines.
Open original ↗
Document

What it claims

The paper's central methodological point, which the summary compresses: the win is not just from having social context, but from adaptive per-case tool use. An agentic classifier that decides which social queries to issue at each reasoning step beats a non-agentic model handed the same contexts up front. The framing is that conspiracy detection is a "socially embedded interpretation task" — inferring illocutionary force, since identical surface text can be endorsement, legitimate concern, criticism, satire, or mockery.

Evidence

The evaluation is run on a manually-annotated adversarial dataset, not on the full corpus — a detail worth noting, since the reported gains are measured on hard, hand-picked cases rather than the broad Hebrew-tweet coverage used to supply contexts. On that adversarial set: context-aware workflows beat text-only classification, and the agentic framework beats both text-only and the context-matched non-agentic baseline. The abstract gives no metric values, no dataset size, and no token counts.

Limits

This is an abstract-only page — no numbers, no baseline figures, no model names, no code or data link. The comparison that carries the paper (agentic vs. non-agentic with identical context) is asserted as "significantly better" without a stated effect size. If you care about the token-economy tradeoff the summary mentions, you must read the PDF; the abstract only names that such analysis exists.

Techniques named here

  • Agentic framework for conspiracy detection

Related in agents