AI Development for the Public Interest: From Abstraction Traps to Sociotechnical Risks

Different AI subfields approach social responsibility through distinct historical lenses: The paper identifies three major AI research areas—AI Safety, Fair Machine Learning, and Human-in-the-Loop Autonomy—each with different understandings of how technical systems interact with society, based on past failures and integration problems specific to their domains.

Graduate-level AI education largely neglects social context despite undergraduate efforts: While universities have begun incorporating ethics and social contexts into undergraduate AI curricula for "Public Interest Technology" goals, graduate-level interventions remain "largely unexplored," potentially leaving advanced AI researchers without adequate training in sociotechnical considerations.

Siloed AI research communities miss opportunities for integrated learning: The paper argues that AI Safety, Fair ML, and HIL Autonomy operate as separate fields with limited cross-pollination, suggesting that a "unified approach to sociotechnical graduate pedagogy" could better prepare AI researchers to address complex real-world problems that span multiple technical domains.

Historical context shapes how different AI communities conceptualize responsibility: Each subfield's approach to responsible AI development is rooted in particular past failures—meaning that AI Safety, Fair ML, and HIL researchers may literally be solving for different problems based on their field's history, potentially creating blind spots when addressing novel risks.