Uncracked City: A Detective Game Using LLM-Powered Suspects

Overview of Uncracked City

Uncracked City is an investigative detective game that tasks players with solving mysteries by questioning AI-driven suspects who may lie. The game features a series of cases set in real-world cities, including Paris, Tokyo, and New York, where players must search for clues, interrogate characters using natural language, and submit a final report to crack the case. It is currently available on mobile platforms, with a focus on high-difficulty, community-wide mysteries.

Core Gameplay Mechanics

The investigation process in Uncracked City is structured into three primary phases:

  • Environment Investigation: Players search real-world locations to uncover clues and unlock evidence.
  • Dynamic Interrogation: Players can question suspects using typed or spoken input in any of the game's six supported languages. Because the suspects are powered by LLMs, they provide unique, voice-acted responses. A core mechanic involves identifying when a suspect is lying.
  • Reporting: Once a player believes they have caught a lie, they must present evidence against the suspect's statement. Players have a limited number of attempts to submit a final report to solve the mystery.

Community and Development Context

Since its launch, the game has sparked discussion within the developer community regarding the evolution of interactive fiction and the technical challenges of AI-based gaming.

Historical Comparisons

Commenters on Hacker News have compared the concept to classic titles like the 1985 Where in the World Is Carmen Sandiego? and the 1990s Sherlock Holmes: Consulting Detective. Users noted that while L.A. Noire previously pushed the boundaries of facial animation and dialogue trees, the current generation of LLMs allows for the ad-hoc, open-ended interrogation that fans of the genre have long anticipated.

Technical and Design Feedback

Community feedback has highlighted several areas for potential improvement and consideration:

  • Writing Quality: Some users suggested that the game's marketing materials and dialogue samples need more polish to ensure the narrative feels engaging rather than repetitive.
  • Inference Economics: A common technical question raised by the community concerns the sustainability of the business model. Given that LLM inference carries a non-zero cost per session, observers have questioned how the developer plans to manage these expenses while keeping the game accessible.
  • User Sentiment: The concept of interrogating AI has drawn mixed reactions. While some users find the prospect of "detective work" using LLMs to be a compelling evolution of gaming, others noted that because they spend their professional lives identifying hallucinations and lies in LLM outputs, they may not find the experience relaxing for leisure time.

Sources