"So There's a Catch-22 Here": How Early Adopters Who Build Multi-Agent LLM Systems Conceptualize Transparency

Transparency in AI systems is more complex than previously understood: The study found that early adopters of multi-agent LLM systems conceptualize transparency in at least five distinct ways—reproducibility, debugging, boundary-setting, visualization, and auditing—rather than as a single, unified concept, suggesting that "one-size-fits-all" transparency approaches may be insufficient.

There's a fundamental tension between building and governing these systems: Participants described a "Catch-22" where the technical complexity of multi-agent systems makes transparency difficult to implement, yet transparency is essential for responsible deployment. This gap highlights an urgent need for better tools and frameworks before these systems become more widespread.

Transparency must be tailored to different audiences: The research identifies that developers, users, and governance stakeholders have different transparency needs and expectations. Effective responsible AI practices require designing transparency solutions that address these distinct perspectives rather than treating all stakeholders the same way.

Transparency is a practical, social challenge, not just a technical one: The study frames transparency as a "situated socio-technical practice," meaning it depends on the specific context, relationships, and capabilities of the people involved. This suggests that responsible AI governance requires attention to organizational culture and communication, not just technical documentation.

Current practices may not scale: Since this research focuses on early adopters within a large technology organization, the findings suggest that transparency practices being developed now may need significant revision as multi-agent LLM systems become more common and are deployed across different organizations and contexts.