The development of large language models has reached a critical juncture. It is no longer sufficient to ask only what these models are capable of doing. We must also grapple with questions such as who is—and is not—involved in and accounted for in the creation of LLMs; what the impacts of these models are at both the individual and societal level; and which values and principles these technologies uphold and promote. This survey examines how such human-centered principles are inherently intertwined with the design, training, and deployment of LLMs.

Ultimately, the trajectory of LLM development must be guided by more than technical benchmarks and capability milestones. The questions of inclusion, impact, and values explored in this survey are not peripheral concerns to be addressed after the fact; they are foundational to what these systems become and who they serve. By centering human-centered principles at every stage of the LLM lifecycle, from design and data curation to training and deployment, researchers and practitioners can work toward models that are not only more capable, but also more equitable, accountable, and aligned with the diverse needs of the people they affect. The path forward demands a broader coalition of voices, a more expansive notion of responsibility, and a sustained commitment to ensuring that progress in AI is measured not only by what these models can do, but by the kind of world their development helps to build.

Acknowledgments

The idea for this manuscript originated in the Fall 2024 offering of Stanford’s CS 329X course on HCLLMs, taught by the instructor Diyi Yang and assistants Rose E. Wang and Caleb Ziems. Diyi Yang developed the initial structure and outline of the paper. The enrolled students collaborated on a first draft of the manuscript as part of their coursework, with each student pair assigned responsibility for drafting a subsection of the survey according to the course assignment structure. The teaching assistants subsequently reviewed, graded, and provided feedback on these drafts. In Winter 2025, a subset of students continued to revise and expand the manuscript under the primary direction of Rose E. Wang and the secondary direction of Caleb Ziems.

Caleb Ziems and Dora Zhao largely re-wrote and restructured the manuscript, with significant conceptual revisions and new chapters, to produce the current version. This revision phase was supported by additional contributions from Sunny Yu and Advit Deepak. All authors reviewed and approved the final manuscript.

The core authors were jointly responsible for deciding on and writing the paper in its final form. The core are as follows:

Additionally, the leadership of this work included:

Student contributions are as follows:

0 items under this folder.