What Makes An Expert? Reviewing How ML Researchers Define "Expert"

Expertise is rarely explicitly defined in ML research: The paper reviewed 112 academic publications that engaged human experts but found that most did not clearly define what they meant by "expertise" or who qualified as an expert, raising concerns about consistency and reproducibility in how expert knowledge is incorporated into ML systems.

Formal credentials are overvalued while other forms of knowledge are overlooked: The research found that expertise is predominantly understood through formal education and professional certification, while forms of knowledge outside these channels—such as lived experience, community knowledge, or practical skills—are "rarely sought," potentially excluding valuable perspectives from ML development.

Expert knowledge is often underutilized: Rather than leveraging experts' full range of knowledge and judgment, ML developers tend to use experts primarily for narrow tasks like data annotation, which the authors describe as "mining textbook knowledge" rather than engaging experts' deeper understanding and contextual insights.

Current practices may contribute to "deskilling" in ML development: By reducing expert involvement to specific, limited tasks rather than meaningful consultation, the field risks losing opportunities to build deeper expertise and understanding within ML teams themselves.

Better documentation and reflection on expert engagement is needed: The paper calls for ML researchers to be more explicit and intentional about why they're engaging experts, how they're using their knowledge, and what kinds of expertise they're recognizing, which would improve both reproducibility and the quality of responsible AI development.