Introduction
As artificial intelligence (AI) transitions from academic curiosity to a foundational societal technology, the discourse surrounding its ethics and governance has intensified. However, a critical perspective is often missing: that of the researchers actually building these systems. This study surveys 524 top-tier AI/ML researchers to bridge the "consensus gap" between technical experts, policymakers, and the general public. By understanding the attitudes of those at the forefront of the field, we can better predict the feasibility and support for future AI regulations.
Methodology
The researchers conducted a comprehensive survey of 524 technical experts who published in two of the most prestigious AI conferences: NeurIPS and ICML (2018). To provide a longitudinal perspective, the study included a subset of researchers from a 2016 survey. Furthermore, the results were benchmarked against a 2018 survey of the US public to highlight the divide between expert and layperson sentiments. Issues were measured using 4-point scales for importance and trust, alongside Likert scales for specific policy support.
Key Insights and Results
Divergent Priorities: Experts vs. The Public
One of the most significant findings is the disparity in how researchers and the general public prioritize AI governance challenges. While the US public views AI through a lens of national security and competition, researchers tend to view it as a global scientific and safety challenge.
As shown in the figure above, the "US-China competition" represents the most striking gap; it is a top-tier concern for the public but is rated significantly lower by researchers. Conversely, researchers place higher relative importance on technical challenges like "Critical AI systems failure" and "Criminal justice bias" compared to the lay public.
The Rise of Safety Culture
The data reveals a burgeoning "safety culture" within the machine learning community. Longitudinal analysis shows a marked shift: in 2016, 49% of researchers believed AI safety should be prioritized more; by 2019, this figure rose to 68%. This suggests that as AI capabilities advance, the technical community is becoming increasingly cautious about the risks of unintended consequences.
Trust and Globalism
Researchers exhibit a "globalist" outlook, placing their highest trust in international bodies (such as the UN and EU) and non-governmental scientific organizations. In contrast to the US public, which views national militaries as trustworthy actors for AI development, researchers rank national militaries and private corporations (like Facebook) among the least trusted entities.
Military Applications and Selective Transparency
Attitudes toward military AI are nuanced. While there is overwhelming opposition to lethal autonomous weapons (LAWs)—with 58% of respondents strongly opposing their development and 31% willing to resign if their organization pursued them—there is general support for non-lethal military AI, such as logistics algorithms.
Furthermore, researchers are moving away from "open-source absolutism." While 84% believe high-level methodology should remain open, only 22% feel the same about sharing fully trained models. This reflects a sophisticated understanding of "dual-use" risks, where the dissemination of powerful models could lead to misuse. Notably, 59% of researchers support institutional pre-publication reviews to assess potential harms.
Conclusion and Future Directions
The survey highlights a technical community that is increasingly safety-conscious, skeptical of nationalistic AI narratives, and supportive of self-regulation. However, the study is limited by a lack of demographic diversity (91% male) and a focus on academic publication as the primary metric for expertise. Future research should incorporate the views of industry engineers and social scientists to create a more holistic map of AI governance attitudes in the era of Large Language Models.
Summary updated