`It is currently hodgepodge'': Examining AI/ML Practitioners' Challenges during Co-production of Responsible AI Values

Responsible AI remains fragmented in practice: The researchers found that despite numerous RAI guidelines from organizations, implementation across companies is "hodgepodge"—lacking coordination and consistency. Practitioners across 10 organizations reported significant gaps between published RAI principles and what actually happens when shipping AI/ML products.

Organizational structures create barriers for different roles: Both top-down mandates from leadership and bottom-up pressure from individual contributors prevent practitioners from upholding RAI values. The study found that different roles face distinct institutional burdens, suggesting that one-size-fits-all RAI approaches don't work across organizations.

Conflicting values are a major execution challenge: When RAI values conflict with each other (such as fairness versus accuracy, or privacy versus utility), practitioners struggle to navigate these tradeoffs without clear guidance. The paper emphasizes this as a particularly difficult problem that current frameworks don't adequately address.

Practitioners use informal "value levers" to resolve conflicts: Rather than following formal processes, the 23 interviewed practitioners developed their own strategies and workarounds to balance competing RAI values. This suggests organizations need to formalize and support these practical problem-solving approaches rather than leaving them ad-hoc.

Supportive organizational structures are essential: The researchers recommend that companies create inclusive structures and opportunities that actively support practitioners in upholding RAI values, rather than treating responsible AI as an additional burden on top of existing work.