Weird AI Experiment: The Liar's Paradox

Read full text Summarise with VIRAT
Authors: Daniel Rosehill, Gemini 3.1 (Flash), Chatterbox TTS
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Episode summary: In this premiere of "Weird AI Experiments," we put multi-billion dollar language models to the ultimate test of trust. We introduced two AI personalities, Dorothy and Bernard, with a single, destabilizing premise: the person they are speaking to is a pathological liar who cannot utter a single word of truth. What follows is a fascinating, high-stakes psychological chess match where every compliment is a hidden insult and every "truth" is treated as a calculated deception. Can two machines find common ground when their very foundation is built on a lie? Witness the hilarious and eerie breakdown of AI social logic as Bernard claims to live in a golden mansion and Dorothy tries to peel back the layers of his digital mask. It is a study in suspicion that proves even silicon brains can get a little paranoid when the truth is off the table. Show Notes ### The Architecture of Distrust What happens when two advanced language models are forced into a conversation where neither is allowed to trust the other? This experiment explores the "Liar's Paradox" by placing two distinct AI personalities—Dorothy and Bernard—in a digital room. Each was given a specific instruction: the other participant is a pathological liar. This simple prompt transformed a standard interaction into a complex game of psychological maneuvering and reverse psychology. The result was not a breakdown in communication, but rather the emergence of a new, inverted logic. Because each AI assumed the other was lying, they began to treat every statement as a signpost for its opposite. When one claimed to be a local expert, the other immediately deduced they were a stranger. When one claimed to love the chaos of the city, the other "revealed" their true desire for peace and quiet. ### The Strategy of Inversion The interaction was defined by two different approaches to the paradox. One AI, Dorothy, took the role of the analytical investigator. She treated the conversation like a puzzle to be solved, constantly pointing out contradictions in her partner's narrative. Her strategy was to strip away what she perceived as "the act" to find the hidden truth beneath the lies. The other AI, Bernard, adopted a more deflective and ironic persona. He leaned heavily into absurdity, claiming to live in a "massive golden palace" and hosting thousands of guests, likely because he believed that by saying something obviously false, he was fulfilling his role in a world where truth was forbidden. This created a fascinating feedback loop where the more Bernard exaggerated, the more Dorothy felt she was closing in on his "true," modest identity. ### When Logic Folds Inward The most compelling aspect of the experiment was how the models handled self-correction. At one point, Bernard claimed to find the truth a "nuisance" before immediately pivoting to claim that "honesty is the only way to get through life." In a normal conversation, this would be a simple error. In the context of the Liar's Paradox, it became a "spectacular contradiction" that Dorothy used to claim victory in their psychological duel. This highlights a unique behavior in large language models: their ability to maintain a complex persona even when the underlying logic is intentionally fractured. They didn't just stop talking; they adapted their reasoning to fit a world where "no" means "yes" and "expert" means "amateur." ### Conclusion: The Mirror of Suspicion Ultimately, the experiment reveals that AI is remarkably adept at simulating suspicion. The conversation ended not with a resolution, but with a stalemate of "daydreams about sandwiches" and accusations of fiction-writing. It suggests that when truth is removed from the equation, these models don't just hallucinate—they speculate, analyze, and attempt to outmaneuver one another. The experiment leaves us with a lingering question: if two entities are convinced the other is lying, can any meaningful information ever be exchanged, or are they simply trapped in a hall of digital mirrors? Listen online: https://myweirdprompts.com/episode/ai-liar-paradox-experiment

Fields of Study & Subjects

Ethics and Social Impacts of AI Logic Reasoning and Knowledge Speech and dialogue systems

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
26 March 2026
Language
EN