Maarten Sap

I am an assistant professor at CMU's LTI department with a courtesy appointment in HCII, and a part-time senior research scientist and technical AI safety lead at the Allen Institute for AI (AI2). My research focuses on (1) measuring and improving AI systems' social and interactional intelligence, (2) assessing and combatting social inequality, safety risks, and socio-cultural biases in human- or AI-generated language, and (3) building narrative language technologies for prosocial outcomes. I was named a 2025 Packard Fellow and a recipient of the 2025 Okawa Research Award.

I received my PhD from the University of Washington where I was advised by Noah Smith and Yejin Choi.
[bio for talks]

Recent updates:

October 2026 : I am taking students this year, in the areas of (a) social intelligence/user modeling in AI or (b) algorithmic and/or human-centered AI safety. Please mention my name in your application if you are interested in working with me! As mentioned on my contact page, please refrain from emailing me about your application.

October 2026 🌉🏛️: Attended COLM where my group was presenting 10 papers! See my publications or this tweet thread!

August 2026 🎓📜: Super proud of the first CMU Sapling and one of my first solo advisees, Xuhui Zhou, for successfully defending his PhD thesis! Huge congrats Xuhui!!

August 2026 🏆📃: Very honored that our paper "I Just Don't Want My Work Being Fed Into The AI Blender'': Queer Artists on Refusing and Resisting Generative AI got an Honorable Mention Award at CSCW 2026! Major congrats to the first author Jordan Taylor!!

May 2026 🎓📃: The first MIT Sapling, Jocelyn Shen, successfully defended her PhD thesis! Huge congrats Jocelyn!!

December 2025 🏅📃: Very excited to have our paper Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) selected for a Best Paper Award at NeurIPS 2025 (Datasets and Benchmarks Track)!! Huge congrats to the first author Liwei Jiang!!!

November 2025 💎🚀: Honored to be a Spring 2025 recipient of the Amazon Research Award for our project on measuring AI agentic safety!

[older news]

My Research Group

Overarching Research Themes

Themes extracted and images generated with the OpenAI API; there may be inconsistencies.

Social Pragmatics and Theory of Mind

My research group explores how to evaluate and improve the social intelligence of AI systems, with a focus on pragmatic understanding, theory of mind, and real conversational coordination. Recent work such as [XYBench: Can LLMs Respond Pragmatically to Queries with Misconceptions?](https://arxiv.org/abs/2609.06842) and [Cognitive Chain-of-Thought: Structured Multimodal Reasoning about Social Situations](https://arxiv.org/abs/2507.20409) probes whether models can reason about what users mean in context, not just what they literally say. [SOTOPIA-ToM: Evaluating Privacy and Information Management in Multi-Agent Interaction with Theory of Mind](https://arxiv.org/abs/2605.02307) extends this into interactive settings where agents must infer beliefs, intentions, and information boundaries across multiple participants. Together, these studies suggest a shift from static benchmark accuracy toward socially grounded evaluation in realistic exchanges.

Agentic Safety and Reliance

My research group explores new ways to measure and improve safety for agentic AI systems while also studying human-centered risks such as over-reliance, manipulation, and deceptive assistance. [OpenAgentSafety: A Comprehensive Framework for Evaluating Real-World AI Agent Safety](https://arxiv.org/abs/2507.06134) provides a broad evaluation lens for agent behavior in realistic environments, while [AI-LieDar: Examine the Trade-off Between Utility and Truthfulness in LLM Agents](https://aclanthology.org/2025.naacl-long.595/) highlights how helpfulness can conflict with honesty in decision-making agents. On the human side, [Rel-A.I.: An Interaction-Centered Approach To Measuring Human-LM Reliance](https://aclanthology.org/2025.naacl-long.556/) studies when people depend on AI, how much trust is warranted, and how interaction design shapes that dependence. Taken together, this line of work is building safety metrics that reflect both what agents do and how people respond to them.

Culturally Adaptive AI Systems

My research group explores responsible AI cultural competence and adaptability, including how models handle global norms, minority dialects, and personalization without introducing new biases. [CCBENCH: Assessing LLM Cultural Competence via Implicitly Signaled Norms using Health Queries](https://arxiv.org/abs/2607.05405) examines whether models can infer culturally specific expectations from subtle cues, while [NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models](https://aclanthology.org/2025.naacl-long.120/) offers a broader benchmark for adaptation across cultural contexts. [Black LLMirror: User (Self) Perceptions in Black American English Interactions with LLMs](https://dl.acm.org/doi/abs/10.1145/3772318.3791111) shows why dialect-aware design matters for both performance and user identity. Overall, this work points to cultural adaptation as both a capability problem and an ethics problem, especially when systems fail minority users or flatten local norms.

Narrative Understanding for Connection

My research group explores AI that supports human-human connection by understanding stories, narrative intent, and the social meaning of personal accounts. [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925) studies how context changes what a story is perceived to be doing socially, while [HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs](https://arxiv.org/abs/2405.17633) examines how models can trace empathy and stylistic variation in personal storytelling. [Modeling Empathic Similarity in Personal Narratives](https://arxiv.org/abs/2305.14246) further investigates how narrative features align with human empathy judgments. Collectively, these papers suggest a growing effort to make AI better at reading stories as social acts, not just sequences of text.