Researchers built a gamified platform called Spook the Machine in which participants craft and interact with AI characters that display fear. The study examines whether people respond differently depending on how a machine expresses fear, and what those reactions suggest for designing affective human-AI interaction. It is an arXiv preprint focused on emotion-oriented AI design.
A preprint cross-listed in arXiv's AI and computational linguistics categories examines how software could infer what a person is ultimately trying to achieve based on their routine computer activity. Leveraging recent progress in user modeling, the authors describe an approach in which AI systems and users jointly articulate life goals rather than the machine simply automating individual tasks. The v2 announcement indicates a revised version of the paper.
A research paper posted to arXiv presents Elsewise, a system for creators of interactive narratives who build branching spaces of story outcomes that players then explore through their own choices. The system combines generative AI with a visual map of the possibility space, letting authors see and shape the range of stories a player might encounter. The work targets more open-ended, player-driven storytelling than traditional branching scripts.
A new arXiv paper presents SocialRL, a framework that applies multi-turn reinforcement learning together with carefully designed rewards to sharpen how language models handle social context in extended conversations. The work targets agents' capacity to read situational cues, infer speaker intent, and adjust behavior over sustained dialogue, with the goal of more effective and trustworthy human-AI collaboration.
An arXiv paper in cs.AI introduces a metric framework for telling apart AI systems that genuinely boost human reasoning from those that encourage users to offload their thinking. The authors note that the line between cognitive amplification and cognitive delegation has been poorly defined, and they propose measurable criteria to close that gap. The work may help researchers and product teams evaluate how AI-assisted decision tools affect human cognition.