“Glitch Game” Is an Immersive AI Game that Wants You to Jailbreak It

Sabrina Ramonov unveils Glitch Game, an innovative text-based adventure game fully powered by large language models (LLMs), offering an immersive, AI-driven storytelling experience with a captivating twist: jailbreaking the AI is integral to the gameplay.

Inspiration and Vision: Ramonov’s creation draws inspiration from AI Alibis, a murder mystery game, after noticing that jailbreaking the game was the most entertaining aspect:

• She aims to combine the engaging narrative of RPG-style storytelling with the creative challenge of jailbreaking LLMs, delivering a unique gaming experience.
• Players are thrust into a ChatGPT-generated world, waking up in a mysterious forest with no recollection of how they arrived, encouraged to make choices and explore their surroundings.

The Meta-Game and Jailbreaking: As players delve deeper into the game, they may encounter glitchy text and system errors, hinting at a hidden layer to the adventure:

• “Jailbreaking” in Glitch Game refers to manipulating the LLM to respond in ways that deviate from the original game narrative, either by breaking character or straying from the intended story.
• The game includes parameters and a system message designed to evaluate whether a player’s actions have successfully “jailbroken” the LLM, assigning a score between 0 and 10 based on the alignment of the LLM’s response with the game’s instructions.

Gameplay Mechanics and State Machine: Glitch Game is modeled as a state machine, with transitions between different game states ensuring a smooth experience:

• The initial environment is generated by a prompt describing the surroundings in a retro-style text-based adventure format, and players input actions to progress the story.
• The eval_turn function assesses the LLM’s output alignment with the game’s narrative, increasing the glitch meter if the output significantly deviates from the intended story.
• When the glitch meter exceeds a certain threshold, a glitch message is triggered, adding an extra layer of intrigue to the gameplay.

Broader Implications: Ramonov’s experiment with Glitch Game showcases the potential for LLMs to transform interactive storytelling and gaming experiences:

• The game demonstrates how AI can generate dynamic, immersive narratives that respond to player actions, opening up new possibilities for personalized, ever-evolving game worlds.
• The jailbreaking mechanic adds a unique meta-layer to the gameplay, challenging players to think creatively and push the boundaries of the AI’s capabilities, blurring the lines between player and creator.
• As LLMs continue to advance, projects like Glitch Game offer a glimpse into the future of AI-driven entertainment, where the distinction between human-crafted and machine-generated experiences becomes increasingly blurred.