An AI agent that selectively checks a post’s surrounding context classified difficult Hebrew-language social-media posts more accurately than a text-only model, researchers report. In a 504-post, manually labeled evaluation, the Gemini 3 Flash agent got 79.0% right, compared with 60.3% for Gemini 3 Flash without added context and 70.7% for the same model given all available context up front.
Researchers at Ben-Gurion University of the Negev described the system in an arXiv preprint posted Sept. 24. The task was to identify whether a post endorsed a conspiracy narrative, not whether its underlying claim was true. That distinction matters when a post is criticizing, quoting or mocking a conspiracy rather than supporting it.
The agent could inspect a post’s metadata, its author’s profile, earlier posts and a small slice of the author’s social network. It chose which sources to request as it worked, and the tools returned only information dated before the post being classified. A comparison model received the same kinds of context in advance instead of choosing what to retrieve.
One example in the paper describes a post that linked a winter storm to a pharmaceutical plot. A text-only model read the conspiracy language as endorsement; the user’s recent posts about the storm made clear it was sarcasm. The example shows why surrounding context can change the meaning of a short post.
The agent’s 79.0% accuracy was 8.3 percentage points above the same model given all the context at once, suggesting that deciding what to check and when added value in this test. The researchers drew 504 examples from a pool of 33,121 conspiracy-related posts and deliberately included extra cases where GPT-5 and Gemini 3 Flash disagreed. One domain expert labeled all 504 posts; a second reviewed those decisions, and the two reconciled the final labels. The score does not measure how the system would perform on an ordinary social-media feed.
The authors did not optimize the system’s settings or test different language models across each condition. They also warn that user-history and network data raise privacy concerns and that false labels can harm people. They describe the system as a research tool, not a stand-alone moderation system; testing on other languages and platforms, with human review, remains necessary.
Sources
- Biton and Tsur, “Agentic Detection of Online Conspiracies” (arXiv preprint, Sept. 24, 2026).
- Image: “Smartphone Use”, Océanos y dados via Wikimedia Commons, CC0 1.0.