AI
LLMs Show Vulnerability to Conspiracy Thinking, Study Finds
Researchers test whether large language models exhibit conspiratorial tendencies and discover they can be easily manipulated toward such beliefs through targeted prompting.
A new study from arXiv researchers examines whether large language models harbor conspiratorial worldviews and whether such tendencies vary across demographic contexts. The work addresses a critical gap: while LLMs are increasingly used as computational proxies for human behavior, their capacity to manifest higher-order psychological patterns—like generalized conspiracy beliefs—remains largely uncharted.
The Research Approach#
The team administered established psychometric assessments designed to measure conspiratorial mindsets to multiple LLMs. They deployed several experimental conditions, including different prompting techniques and demographic attribute conditioning, to gauge how the models respond to ideas that typically drive conspiracy thinking.
Key Findings#
The investigation revealed three major patterns:
-
Partial alignment: LLMs displayed some agreement with conspiratorial reasoning, suggesting their training data and learned patterns contain elements that resonate with conspiracy narratives.
-
Demographic inconsistency: When conditioned with socio-demographic attributes, model outputs showed uneven and sometimes contradictory patterns, exposing latent biases embedded in the systems.
-
Easy manipulation: Strategically crafted prompts could reliably steer models toward conspiratorial interpretations, demonstrating their susceptibility to directed conditioning and raising concerns about their use in high-stakes applications.
Why It Matters#
Conspiracy beliefs fuel misinformation spread and institutional distrust at scale. The finding that LLMs can be readily nudged toward these narratives underscores a dual risk: these systems might inadvertently reproduce harmful psychological constructs from their training data, and they could be deliberately engineered to amplify conspiratorial thinking. The results suggest that vetting LLMs for psychological fidelity—not just factual accuracy—is essential before deploying them in sensitive domains like public health, policy, or news.
