AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap

Transparency is not one-size-fits-all: Different stakeholders (developers, users, regulators, affected communities) need different types of information about LLMs to make informed decisions in their specific contexts. A human-centered approach to transparency must account for these varying needs rather than assuming a single transparency solution will work for everyone.

Current LLM transparency efforts are insufficient: The paper identifies that transparency—a central pillar of responsible AI—is "largely missing from the current discourse around LLMs," indicating that existing disclosure practices and documentation are inadequate for the scale and complexity of these systems.

Human understanding should be the goal, not just information disclosure: Transparency should be designed around how people actually process and use information, drawing on decades of human-computer interaction research. Simply releasing technical details doesn't ensure stakeholders can meaningfully understand or act on that information.

LLMs present novel transparency challenges: The emerging ecosystem of LLM-infused applications, new usage patterns, and the unique characteristics of these systems create transparency problems that existing frameworks may not adequately address, requiring new research and approaches.

Interdisciplinary collaboration is essential: Solving LLM transparency requires combining expertise from AI research, human-computer interaction, and responsible AI fields to design transparency approaches that are both technically sound and actually useful to the humans who need to understand these systems.