A Taxonomy of Foundation Model based Systems through the Lens of Software Architecture

Foundation models require careful architectural choices with real trade-offs: The paper identifies that designing systems around foundation models (like ChatGPT) involves multiple architectural decisions, and choosing one approach often means sacrificing benefits of another—meaning there is no one-size-fits-all solution for responsible AI implementation.

Responsible AI must be built into design from the start, not added later: The taxonomy includes "responsible-AI-by-design" as a core architectural category, suggesting that safety, fairness, and accountability considerations should be central to how foundation model systems are structured from inception, rather than treated as an afterthought.

How foundation models are adapted matters significantly: The paper examines "pretraining and adaptation" as a key category, indicating that the way organizations customize and fine-tune foundation models for specific uses has important implications—suggesting that responsible deployment requires thoughtful decisions about model modification and training approaches.

Foundation model systems are still in early stages with limited established best practices: The authors note that "design of foundation model based systems has not yet been systematically explored," highlighting that organizations deploying these systems today are navigating largely uncharted territory without established architectural guidelines for responsible implementation.

Software architecture decisions directly impact AI responsibility: By framing responsible AI as an architectural concern alongside technical design choices, the paper emphasizes that responsible AI outcomes depend on fundamental structural decisions made by engineering teams, not just policy or ethics reviews.