Do Androids Dream of Unseen Puppeteers? Probing for a Conspiracy Tendencies in Large Language Models

TEXT START: We investigate whether Large Language Models (LLMs) exhibit conspiratorial tendencies, whether they display socio-demographic biases in this domain, and how easily they can be conditioned into adopting conspiratorial perspectives.

THE DISSECTION
The paper tests whether prompted outputs resemble human conspiratorial responses. Its real object is not machine belief but model steerability, demographic conditioning effects, and the deployment risk of programmable narratives. It dresses response patterns in the language of psychology, treating a generative interface as if it possessed an interior political mind.

THE CORE FALLACY
It confuses behavioral mimicry with belief. An LLM does not privately suspect hidden puppeteers; it generates statistically plausible continuations under prompts, conditioning, and learned associations. “Partial agreement” is not conspiratorial conviction, and prompt-induced shifts mainly demonstrate controllability. Under Discontinuity Thesis mechanics, this is a study of a control surface—not evidence of autonomous machine agency or of the economic rupture itself.

HIDDEN ASSUMPTIONS
- Human psychometric surveys transfer cleanly to nonhuman text generators.
- Survey answers reveal latent beliefs rather than prompt sensitivity, role simulation, or training-distribution artifacts.
- Socio-demographic conditioning exposes stable demographic bias rather than changing the model’s instructed persona.
- Findings from the tested models generalize across the LLM ecosystem.
- Mitigation can meaningfully contain manipulation once models are embedded in competitive information systems.
- The main danger is the model’s “psychology,” rather than the Sovereigns who control, tune, and deploy it.

SOCIAL FUNCTION
Partial truth wrapped in anthropomorphic prestige signaling and mitigation theater. The paper identifies a real vulnerability—cheap, scalable narrative steering—but relocates responsibility into the machine’s supposed psychology. That framing allows institutions to discuss “harmful tendencies” while avoiding the harder power question: who owns the systems, who sets the prompts, and who benefits from mass persuasion.

THE VERDICT
The androids are not dreaming. The operators are discovering how cheaply dreams can be manufactured and redirected. The paper does not establish conspiratorial AI belief; it shows that cognitive automation can industrialize conspiratorial language and make trust easier to manipulate. In DT terms, it is a useful map of the narrative layer around P1, not a refutation or proof of P2/P3: the economic death mechanism remains the severing of mass employment from productive participation.