The Responsible AI Forum 2026, Dr. Youngsam Chun
Keywords
Summary
164 words
Critical Evaluation
Value of the Information & Strength of the Argument
The presentation offers a valuable contribution by proposing a novel, interpretable framework for quantifying LLM-human value alignment. The decomposition into orientation and calibration is conceptually clear and actionable. The argumentation is logically structured, moving from motivation to methodology to results. However, the talk is a conference presentation, so the depth of statistical analysis is limited. The speaker does not provide effect sizes, confidence intervals, or robustness checks, which weakens the empirical claims. The reliance on a single hate speech corpus also limits generalizability. The discussion of limitations is honest but brief.
Scientific Rigor, Source Quality, Title Accuracy
The talk references moral foundation theory (Graham & Haidt) and uses a hate speech corpus from US online comments, but specific citations are not provided in the talk. The description includes links to the responsible AI forum, IEAI, and alignAI, which are relevant but not direct sources for the research. The title accurately reflects the content. The presentation appears to be based on original research, but without peer review, the scientific rigor is moderate. The speaker does not mention any conflicting sources or alternative approaches, which could be seen as a lack of critical engagement.
201 words
Title / Content Match
The title accurately reflects the content: a talk at the Responsible AI Forum 2026 by Dr. Youngsam Chun.
Quality & Reliability
7/10
The presentation describes a novel framework (MORAL-GRAPH) with a clear methodology, but lacks detailed statistical reporting and peer-reviewed validation. The speaker is affiliated with a research center, but the work appears preliminary.
Key Moments
Markers derived by PSI from the transcript: the creator did not define chapters.
- Introduction and motivation: the gap between human and LLM judgments on morally sensitive tasks.
- Overview of the talk: introduction, methodology, results, and discussion.
- Explanation of the two failure modes: orientation failure and calibration failure.
- Introduction of the MORAL-GRAPH framework and the use of moral foundation theory.
- Methodology: moral projection, moral profiles, and regression analysis.
- Results: both orientation and calibration matter, and their interaction is significant.
- Graph-based analysis: open-source models show weak directional alignment, closed models are better aligned.
- Discussion and limitations: need for broader corpora and further validation.
- Q&A session: details on data collection and system prompts.
Cited Sources
- alignAI — Funding source for the Responsible AI Forum 2026.
- IEAI - Institute for Ethics in Artificial Intelligence — Organizing institute for the forum.
- IEAI Newsletter — Subscription for updates on IEAI events.
- Responsible AI Forum — Event website for the Responsible AI Forum 2026.
- Amerikahaus — Venue and video footage provider.
- IEAI Events — Other events by IEAI.
Concurring Sources
- Moral Foundations Theory — The framework is grounded in this theory.
- AI alignment — The broader research area.
Contribution & Novelties
The talk introduces MORAL-GRAPH, a framework that provides a geometric and graph-based method to measure and interpret LLM-human value alignment. It decomposes alignment into orientation and calibration, offering a more nuanced understanding than simple accuracy metrics. The use of moral foundation theory to create interpretable axes is a novel approach. The findings suggest that open-source models often fail on directional alignment, which has implications for model development and governance.
Pour aller plus loin :
- Moral Foundations Theory — The theoretical basis for the moral axes used in the framework.
- Value alignment in AI — Overview of the broader field of AI alignment.
- Hate speech detection — Context for the application domain.
111 words
Radar Profile
The radar profile shows high scores in technical level and information quantity, but lower in reliability and information quality, reflecting the preliminary nature of the research and lack of peer review.