Explaining Any ML Model? -- On Goals and Capabilities of XAI

XAI algorithms cannot deliver all promised benefits: The paper reveals that despite widespread claims, current explainability tools can only reliably answer one core user question about ML models, not the diverse range of questions stakeholders actually need answered. This means XAI cannot single-handedly solve problems like building trust or ensuring fairness as often claimed.

Vague terminology obscures what XAI actually does: Terms like "interpretability" and "explainability" lack precise definitions in XAI research, creating confusion about what these tools can realistically accomplish. This definitional gap means policymakers and organizations may have unrealistic expectations about XAI's capabilities for responsible AI deployment.

XAI effectiveness depends heavily on context: The paper emphasizes that whether an explanation is even possible—let alone useful—depends entirely on the specific ML application. This means there is no one-size-fits-all XAI solution, and organizations need to carefully assess whether explanations are actually achievable for their particular use case.

User needs should drive XAI development: Rather than assuming XAI tools will create desired properties like trust, the paper argues researchers should start by understanding what specific questions different users actually need answered. This user-centered approach could lead to more practical and genuinely useful explanation methods.

Realistic expectations are essential for responsible AI governance: Policymakers should understand that XAI algorithms have real limitations and cannot guarantee accountability or fairness on their own. Responsible AI requires acknowledging these constraints rather than relying on explanations as a catch-all solution.