Imagine stepping into a virtual household where a mysterious event has unfolded. The fridge door is wide open, but who left it that way? This is the challenge posed by MARPLE, a new AI benchmark designed to test an AI's ability to solve "whodunit" mysteries. Researchers are essentially putting AI’s detective skills to the test. MARPLE simulates everyday household scenarios, generating visual, language, and audio clues related to two virtual agents. The AI’s task is to identify which agent caused a specific event, like turning on the laundry machine or picking up a snack. The goal is not just to find the culprit, but to determine how quickly the AI can crack the case with limited evidence. The current study pitted humans against AI models, including a large language model (LLM) like GPT-4, and traditional Monte Carlo simulation methods. Humans outperformed the AI in every scenario, showcasing a superior ability to piece together the narrative and anticipate actions. While AI models improved their accuracy over time, humans needed considerably less evidence to reach the right conclusion. Traditional methods struggled to generalize to new, unseen environments, which humans handled with ease. Interestingly, LLMs like GPT-4 often fixated on changes in an agent's state, like position or direction, rather than changes in the environment itself. This sometimes led the LLM down the wrong path. For example, if an AI only focused on an agent’s movements, it might miss a crucial clue hidden in a change to the environment (e.g., a light turning on or a door closing). This suggests that LLMs still struggle with causal reasoning across longer sequences of events. MARPLE offers a unique challenge for AI researchers. Unlike tasks focused on short-term reasoning or isolated physical events, MARPLE requires AI to combine different modes of evidence and reason about actions over longer time horizons. The future of at-home AI assistants may well depend on cracking this complex reasoning puzzle. More advanced reasoning could pave the way for AI that can provide richer, more intuitive assistance in our daily lives.