The Series of Unfortunate Events series of books by Lemony Snicket (who in actuality is also Magnetic Fields’ accordionist Daniel Handler) are a treasure trove of macabre writing for children, that went on to be a bad Jim Carrey movie and a quite good Netflix series. Now the series lives on in—oh lord—an LLM-driven game to be published by The Atlantic.
Lemony Snicket’s Suspicious Incident in Dubious Park is described by centrist mag The Atlantic as “its first immersive-narrative game,” and in it players attempt to solve a murder mystery via a “combination of a traditional whodunit and a visual scavenger hunt.” Players will take on the role of a detective in a locked-down park, hunting for clues and interviewing “a cast of peculiar witnesses and eccentric parkgoers” to catch the killer. Which, so far, sounds great! And then (emphasis ours):
To create the game, The Atlantic worked with Lemony Snicket to create the game’s cast of suspicious and dubious characters, dialogue, and plot. The Atlantic then trained an LLM for each character based on their unique dialogue, motives, and background, allowing players to interrogate and interact with the characters directly to uncover the perpetrator of the crime. The game is illustrated by the Eisner Award–winning comic artist Michael Kupperman.
The game, which launched for Atlantic subscribers today, July 13, will open up to the wider public on July 20, as part of the magazine’s attempt—alongside so many other publications—to pivot toward the lucrative world of puzzles that has so boosted the New York Times. However, most have focused on crosswords and Wordle-likes to draw in a subscribing crowd, rather than genAI-powered murder mysteries based on the work of a fictional author of children’s books. (It’s not clear from the press release if the “Lemony Snicket” writing the game is actually Daniel Handler, although there’s been no word that the author has ever let anyone else use the name.)
However, the game has taken a unique approach to preventing inappropriate conversations from being generated: it requires you to stay entirely on-topic in your interactions with the game’s characters, punitively locking you out from playing if you ask three innocuous questions that stray from the matter at hand. I told a character I loved them, and was given a strike. I asked another what I should do next in my life, and received another.
The Atlantic‘s concern about the possibility of people making the game’s colorful cast of suspects say things they really shouldn’t is understandable. LLM chatbots are sometimes incredibly vulnerable to being “jailbroken,” such that they’ll gladly go far off script. But the solution that’s been implemented here results in the game feeling like the most weirdly restrictive affair. It’s currently unclear if the ban you receive after three strikes is permanent. We’ve reached out to The Atlantic to ask.
It’ll be interesting to see if sneaky people figure out a way to get past these most draconian of restrictions, and what mayhem could then ensue. Or if the current pearl-clutching reaction to the most banal of off-topic questions might be loosened. But this seems to be the damning issue of genAI slop: you either restrict it so much it barely functions, or you allow it enough rope to hang itself.