Reflections and New Directions for Human-Centered Large Language Models

Large Language Models (LLMs) are increasingly shaping the private and professional lives of users, with numerous applications in business, education, finance, healthcare, law, and science. With this rise in global influence comes greater urgency to build, evaluate, and deploy these systems in a manner that prioritizes not only technical capabilities but also human priorities. This work presents a framework for developing Human-Centered Large Language Models (HCLLMs), which integrates perspectives from Natural Language Processing (NLP), Human-Computer Interaction (HCI), and responsible AI. Considering the ethics, economics, and technical objectives of language modeling, we argue that model developers need to address human concerns, preferences, values, and goals, not only during a cursory post-training stage, but rather with rigor and care at every stage of the pipeline. This paper offers human-centered insights and recommendations for developers at each stage, from system design to data sourcing, model training, evaluation, and responsible deployment. Then we conclude with a case study, applying these insights to understand the future of work with HCLLMs.

Large Language Models (LLMs) have transitioned from research artifacts to production infrastructure. They now power developer tools, enterprise copilots, search and recommendation systems, content moderation pipelines, and domain-specific assistants across healthcare, finance, education, science and law. As LLMs are integrated into individual and collective processes, they can no longer be understood as isolated tools bounded by static performance metrics or leaderboard positions. LLMs are sociotechnical systems with global influence, and should be developed and evaluated in starkly more human terms. Are these models helpful, steerable, and safe under adversarial pressure, aligned across global markets, robust to distribution shift, and adaptable to evolving user goals and expectations? Do models comply with data governance regimes, privacy regulations, and ethical concerns around intellectual property? How can we build models that not only avoid harm but also actively contribute to human flourishing? Can LLMs do more than just passively assist humans; can they also actively collaborate with us as equal partners?

This survey advances the framework of Human-Centered Large Language Modeling (HCLLMs) as a unifying lens for understanding and answering these questions. Rather than treating human-centered objectives as simple patches or alignment problems downstream of capability scaling, we argue that human-centered methods must be embedded across the entire LLM development pipeline, from data sourcing and filtering, to post-training and alignment, evaluation, deployment, and long-term maintenance. This survey paper elaborates on and endorses the alternative, Human-Centered Design (HCD), where users and other stakeholders are centrally involved with ideating, building, evaluating, and deploying Large Language Models. Their centrality at every stage of the design process is what distinguishes HCD from other instantiations of human factors design that account for general user needs in only a small slice of the design or deployment process.

Importantly, we will demonstrate how human-centered objectives tend to resist universal solutions. The optimal path will depend both on who you ask and how you operationalize concepts like harm and benefit. Broad themes like transparency, privacy, safety, and justice frequently emerge, but there will be significant variation in perspective on how these ideals should be implemented. Governments and non-profit organizations may codify the most dominant perspectives into laws and policies, but high-level guidelines may fail to account for the nuances of real-world use, and lag behind the rapid evolution of language models themselves. In the face of these challenges, stakeholders often remain passive, which only endorses the status-quo.