Report by Ryota Takatsuki and Hiroshi Yamakawa.
Members of ALIGN attended the 39th Annual AAAI Conference on Artificial Intelligence (AAAI-25), held in Philadelphia, United States, from 25 February to 4 March 2025, and organized the 1st Workshop on Post-Singularity Symbiosis (PSS 2025) there. This two-part report, “Frontiers of AI Alignment,” covers AI alignment research at the main conference (this part) and the workshop (Part 2).
- Overview. AAAI-25 received about 13,000 paper submissions, with an acceptance rate of about 23.4%, and accepted 49 workshops. A new Special Track on AI Alignment was established, and Stuart J. Russell of the University of California, Berkeley received the 2025 AI for Humanity Award. The authors see both as signs of growing interest in AI safety.
- The AI Alignment track. The authors sorted its 66 accepted papers by title and abstract: safety mechanisms and adversarial robustness (18), value alignment and preference learning (16), bias, fairness and social impact (9), verification and formal guarantees (6), interpretability and transparency (6), multi-agent alignment and cooperation (6), and foundational theory and methodology (5). Most papers dealt with model vulnerabilities or value alignment; few dealt with formal guarantees or interpretability. The report then introduces representative papers from each category.
- Invited talks. Stuart Russell (“Can AI Benefit Humanity?”) was skeptical that current systems such as GPT will reach true AGI, stressed how hard it will be to control future AGI and ASI, and proposed formal safety verification. David Chalmers (“Propositional Interpretability in Humans and AI Systems”) proposed propositional interpretability as a subcategory of mechanistic interpretability, with the idea of “thought logging” (preprint). Andrew Ng (“AI, Agents and Applications”) presented five trends in AI development in an optimistic talk focused on the benefits of AI.
- Overall impressions. Safety risks from more capable AI systems are beginning to receive attention at AAAI, but such work is still a small share of all presentations, and the authors call for more research in this area.
The report was prepared with the support of the Matsuo-Iwasawa Laboratory at the University of Tokyo.