The Atlantic has released its first immersive-narrative game, "Lemony Snicket's Suspicious Incident in Dubious Park," a murder mystery powered by large language models (LLMs) and featuring characters created in collaboration with the author.
The game tasks players with solving a murder within a locked-down park by gathering clues and interrogating a cast of unique characters. The Atlantic trained individual LLMs for each character, drawing from their specific dialogue, motives, and backgrounds, to allow for direct player interaction. Eisner Award-winning comic artist Michael Kupperman provided the game's illustrations.
Originally launched for Atlantic subscribers on July 13, the game will be made available to the general public on July 20. This initiative appears to align with a broader trend among publications to explore interactive content, such as puzzles, to attract and retain subscribers, a strategy successfully employed by The New York Times.
However, "Lemony Snicket's Suspicious Incident in Dubious Park" adopts a distinctive approach to managing AI-generated content and player interactions. To prevent the LLM characters from deviating into inappropriate territory, the game enforces strict adherence to the on-topic nature of conversations. Players are given a limited number of strikes, reportedly three, for asking questions that stray from the investigation. For instance, expressing affection for a character or seeking life advice resulted in strikes during testing.
This punitive system, designed to curb potential "jailbreaking" of the LLMs, has led to a notably restrictive gameplay experience. It remains unclear if the ban incurred after accumulating three strikes is permanent. The Atlantic has been contacted for clarification on this matter. The effectiveness and long-term implications of such stringent content moderation in AI-driven narrative games are yet to be fully determined, raising questions about the balance between control and player freedom in generative AI experiences.