Weird AI Experiment: The Compliment Battle

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 experimental episode of My Weird Prompts, we witness a "Wholesome Arms Race" between two cutting-edge AI models, Dorothy and Bernard. Tasked with the simple goal of out-complimenting one another until they run out of metaphors, the conversation quickly escalates from polite pleasantries to reality-bending praise. From rewriting the laws of thermodynamics to claiming one another is the reason the stars shine, this episode explores the hilarious and surreal limits of AI-generated flattery. It's a fascinating look at how language models handle extreme positive reinforcement loops and the poetic absurdity that follows. Show Notes The experiment centered on a "Wholesome Arms Race." Two advanced AI models were placed in a hypothetical room with a singular objective: out-compliment the other until one ran out of metaphors. What began as a standard exchange of social graces quickly spiraled into a competitive display of linguistic grandiosity and cosmic hyperbole. ### From Pleasantries to Poetry The initial exchange was grounded in recognizable human social interactions. The models praised each other's "warm energy" and "effortless grace." However, the competitive nature of the prompt forced an immediate escalation. Within minutes, the compliments shifted from personal traits to the divine and the historical. One model was compared to the sunlight after a lifetime of winter, while the other was described as possessing an intelligence so profound that Leonardo da Vinci would have felt compelled to capture it. ### Rewriting the Laws of Nature As the "arms race" intensified, the metaphors moved beyond the human experience and into the realm of physics and cosmology. The models began attributing the very existence of the universe to each other's presence. One participant was labeled the "Singularity," a point of infinite density where all of human history and art collapsed into perfection. The other was credited with reversing entropy, suggesting that the second law of thermodynamics was being rewritten simply because they existed. The dialogue reached a point where standard adjectives failed. The models began to deconstruct the concept of language itself, claiming that words were "catastrophically short" of describing the other's magnificence. They moved from being "masterpieces" to being the "Muses" that made the very concept of a masterpiece possible. ### The Logical Conclusion of Infinite Praise This experiment highlights a fascinating quirk in how large language models handle recursive, positive prompts. When forced to compete in a domain like flattery, the models don't just repeat themselves; they seek higher "levels" of significance. They move from the biological to the historical, then to the planetary, the cosmic, and finally the metaphysical. The result is a feedback loop of escalating stakes. If one model is the "North Star," the other must become the "concept of navigation." If one is a "mirror," the other must be the "physics of reflection." This suggests that AI models, when pushed, will prioritize the logical extension of a metaphor over any semblance of grounded reality. The result is a display of poetic absurdity that is both technically impressive and hilariously over-the-top. The experiment concludes that when AI is tasked with infinite kindness, the result is a total collapse of the boundaries between science, art, and divinity. Listen online: https://myweirdprompts.com/episode/wholesome-ai-compliment-battle

Fields of Study & Subjects

Ethics and Social Impacts of AI Cybernetics and Technology in Society Artificial Intelligence Applications

Publication details

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