Agentic Detection of Online Conspiracies

The 30-second take

• What: The authors treat conspiracy detection as inferring illocutionary force with an agent that issues social-context queries, and on a manually annotated adversarial set it beats text-only classifiers and a non-agentic model given the same contexts.
• Abundance angle: today, reading speaker intent in conspiracy-adjacent talk is scarce expert (and language-specific) labor. Adaptive, tool-using interpretation could become a cheaper default moderation or research layer if it holds beyond this Hebrew corpus (near-term software; validation and abuse risk still decide access).
• Who should care: Trust-and-safety and computational-social-science teams, especially those working outside English, and anyone who has watched keyword filters flatten satire into “conspiracy.”

What the paper actually did

Conspiratorial talk online is not always a slogan with a stable keyword. The same surface text can endorse a theory, raise a legitimate concern, criticize it, satirize it, or mock it. The hard part, the authors argue, is inferring the speaker’s intent — the utterance’s illocutionary force — not only spotting conspiracy-related claims.

They say relevant social contexts can do that work and propose an agentic framework with tools for social queries. They demonstrate it on a dataset of Hebrew tweets covering an estimated 80–90% of public Hebrew tweets from late 2018 to early 2023, spanning several election cycles, the COVID years, and vaccination campaigns — coverage they say helps recover different social contexts.

On a manually annotated adversarial dataset, context-aware workflows beat text-only classification, and the agentic framework beats other settings, including a non-agentic model that sees the same contexts the agent can access. They analyze errors and token-economy tradeoffs. The conclusion they draw is that conspiracy detection is a socially embedded interpretation task: classification depends on access to context and on adaptive, per-case tool use that asks only for evidence needed at the current reasoning step.

What makes this disruptive

The scarce capability is intent-sensitive reading of conspiracy-adjacent speech, especially in a language that is not the usual English benchmark. Keyword and claim classifiers collapse satire and worry into the same bin.

The paper’s move is agentic tool use over a nearly complete multi-year national tweet stream, plus a comparison that isolates agency: same contexts, worse performance if the model cannot choose queries. That is a sharper claim than “add more context features.”

It is still one language, one platform, one adversarial test set. Treat the win as a methods signal for socially situated classification, not a finished global conspiracy detector.

Why it matters (outside the lab)

Abundance lens: careful interpretation of political and health talk is still an elite human (or English-centric model) service. If tool-using agents can recover intent from social context, more of that reading can sit in ordinary research and safety stacks — a cheaper default, with serious misuse risk if deployed as blunt takedown software.

Near-term, the preprint updates how teams think about context versus adaptive querying. Medium-term, other languages, platforms, and independent annotation decide whether this becomes default tooling.

No calendar. Better classifiers do not automatically make healthier public spheres.

Limitations & open questions

Preprint; we have not rerun the agent or re-annotated the adversarial set. Hebrew Twitter 2018–2023 is unusually complete in their description but is not the live, multilingual, multi-platform web. Manual adversarial labels can encode the annotators’ theory of “endorsement.”

The abstract does not list the tools, the base model, or exact metrics. Token-economy analysis is mentioned, not quantified here. An agent that queries social context can also leak or over-collect social data.

Abundance is not automatic, and a stronger detector can be used to suppress legitimate concern as easily as to study discourse.