What Areas of AI Safety and Alignment Research are Largely Ignored?

With the wide variety of research directions pursued in AI safety and alignment, are there common properties or categories of research that are routinely overlooked, thereby creating systemic risks? Ignored areas of AI safety and alignment research can be categorized according to the following properties and themes:

1. Conceptual Over Empirical:
   • Research that focuses on theoretical innovations, like semantic backpropagation (the concept that backpropagation might exchange graphs representing meaning rather than just exchanging numbers or other symbols), is often neglected because the AI safety community places a high value on empirical evidence. The insistence on testing in real-world settings limits the exploration of ideas that could address non-linear risks before they manifest​​.
2. Decentralized Collective Intelligence (DCI):
   • Approaches that emphasize decentralization and collective problem-solving are sidelined in favor of centralized methods that fit within existing institutional frameworks. This leads to underestimation of strategies aimed at reducing systemic risks posed by centralized AI control​​.
3. Semantic Representation and Interpretation:
   • The absence of detailed exploration into semantic models—which integrate meaning-rich interactions—has left interpretability and collaboration at the multi-agent level underexplored. Such models could improve transparency and facilitate better alignment strategies​​.
4. Exponential Problem-Space Growth:
   • Research into areas where the solution space grows exponentially, such as multi-agent systems and emergent synergies, lacks attention. This is because current safety frameworks prioritize linear models of oversight and are ill-equipped to address combinatorial threats or opportunities​​.
5. Non-Traditional Collaboration and Outreach:
   • Innovative methods for engaging the broader community and integrating interdisciplinary approaches, like involving underrepresented voices or using creative mediums (e.g., documentaries, games), are overlooked despite their potential to increase awareness and participation in alignment efforts​​.
6. High-Risk, High-Reward Ideas:
   • Speculative or high-concept proposals, such as implementing systems for semantic backpropagation to enable both better safety and greater collective intelligence, are deprioritized due to their lack of immediate, measurable results. This contributes to a cycle of underfunding and low visibility​.
7. Addressing Bias in Safety Governance:
   • The reliance on consensus-based validation mechanisms and institutional inertia stifles innovation. This creates a bias toward maintaining the status quo, even when alternative frameworks could address systemic risks more effectively​​.
8. Scaling Human Problem-Solving Abilities:
   • Research that aims to exponentially scale human group problem-solving, especially through graph-based knowledge systems and AI-enhanced collective intelligence, is often misunderstood or dismissed as impractical without demonstration​​.

This categorization highlights a systemic undervaluing of theoretical, decentralized, and speculative ideas in favor of empirically grounded but often incrementally effective approaches. Addressing this gap requires fostering interdisciplinary dialogue and creating funding mechanisms for early-stage, high-impact concepts.