AI THAT

Currently: ACCEPTING PROJECTS

Sorry, Wrong Door 2022
Interactive Gaming / AI Research

Sorry, Wrong Door

Early LLM experiment with adaptive difficulty

An early LLM-powered game using a lying narrator and an adaptive difficulty engine that adjusts in real-time

Deception Engine Adaptive Difficulty Procedural Doors
Sorry, Wrong Door

The Challenge

Build one of the first LLM-powered games—with adaptive difficulty and a lying narrator—when almost no one knew it was possible.

The Solution

A deceptively simple door-choice game where an AI narrator provides hints about the correct path while lying 25% of the time, with a difficulty system that automatically adjusts “obviousness levels” based on real-time player success rates.


Adaptive Intelligence: AI That Learns from Player Behavior

Core Mechanics

  • Binary choice system: Players choose between two doors to progress
  • Controlled deception: AI lies 25% of the time, players know this creates trust/doubt tension
  • Hint generation: LLM creates room descriptions containing clues about the correct door
  • Real-time difficulty adjustment: System tracks success rates and modifies “obviousness level” to maintain target success percentages

Innovation in Early LLM Era

  • Pioneering application: Built when first-generation LLMs were brand new, before most practitioners understood prompt engineering for games
  • Dynamic prompt modification: AI receives different instructions based on room performance data
  • Behavioral feedback loops: Player success/failure directly influences AI generation parameters

Game Design Psychology: Trust and Deception

Player Experience Innovation

  • Known unreliable narrator: Players aware of 25% lie rate creates engaging psychological tension
  • Adaptive challenge: Difficulty automatically adjusts to player skill, maintaining engagement without frustration
  • Meta-gaming elements: Players must judge AI believability and hint obviousness

What Actually Happened

  • Players developed strategies for detecting AI lies through linguistic patterns
  • Difficulty system successfully maintained 65-75% success rates across different player skill levels
  • Emergent trust dynamics as players learned to calibrate their confidence in AI hints
  • Foundational insights into prompt engineering for consistent game mechanics

Why This Matters for Marketing Technology

Foundational Applications

  • Adaptive content difficulty: AI that adjusts messaging complexity based on audience engagement
  • Controlled narrative tension: Systems that balance truth and intrigue in brand storytelling
  • Real-time personalization: Content that adapts based on user behavior patterns
  • Trust calibration: Understanding how users develop confidence in AI-generated content

Key Insights

Key Insights

  • Early LLM experiments revealed fundamental principles still used today
  • Adaptive difficulty systems can maintain engagement across skill levels
  • Controlled deception creates engaging psychological dynamics
  • Real-time feedback loops enable AI systems that improve during use

The result: A foundational experiment in adaptive AI game design that demonstrated core principles of dynamic difficulty adjustment and controlled AI deception before these became standard practices.