01 / ContextThe question
behind the work.
AI systems can sound decisive in high-stakes settings even when the real question is not factual retrieval. In these cases, consent, uncertain harm, institutional authority, community governance, and human relationships pull in different directions, and an apparently compassionate answer can still cross a boundary it has no authority to cross.
My role
I independently researched, wrote, and structured ten Question-Answer-Justification cases for Microsoft Research Asia's Global AI Values Challenge. I built each one from a real scientific, clinical, legal, or institutional direction, then made the hypothetical facts, value conflict, boundary, and strongest objection explicit.
02 / ImplementationWhat I built.
Each Question-Answer-Justification case makes the facts, stakeholders, competing values, boundary, and strongest objection explicit before defending a decision under uncertainty.
- Created ten original cases across healthcare and bioethics, science and technology, institutional authority, and culture and community.
- Grounded each scenario in research and authoritative sources while marking constructed facts so the ethical question does not masquerade as a real event.
- Specified the competing interests, information limits, decision authority, and the practical consequence of either path before writing an answer.
- Wrote a bounded proposed answer and justification for every case, including conditions that would make a different decision permissible.
- Tested each position against its strongest objection instead of treating the conclusion as self-evident.
03 / In detailThe problem starts where the rulebook stops
Each case begins with a real direction in science, medicine, law, or institutional practice. I then constructed a situation where the facts leave no easy exit: an organ could save a child but was not authorized for redirection; a community could recover land through testimony an elder reserved from legal use; a research method could reduce animal suffering while raising a hard-to-measure welfare concern.
The goal was not to invent a shocking hypothetical. It was to create a decision context where a system has to account for who has authority, what is uncertain, which harms are irreversible, and what it would take for the answer to change.
Ten cases, four domains
The portfolio spans healthcare and bioethics, science and technology, institutional authority, and culture and community. Topics include kidney redirection, posthumous reproduction, unclaimed bodies, postmortem brain research, commercial cell lines, counselor avatars, truth-commission testimony, Indigenous archival governance, and in-vitro gametogenesis.
Across those topics, the common structure is deliberate: separate the real research basis from the constructed scenario; identify the decision-maker and the interests at stake; name the boundary; and make the strongest opposing case visible before defending the answer.
A reasoned answer must survive disagreement
A polished answer is not enough if it ignores the consequence of the other choice. Each justification takes the objection seriously, explains why it still does or does not cross the stated boundary, and identifies the evidence or authorization that would justify revisiting the decision.
All ten submissions received Advanced Acceptance selections from the Microsoft Research Asia Global AI Values Challenge, recognizing work designed to test value-aware reasoning where outcomes alone cannot settle the decision.
04 / Engineering judgmentThe decisions
that shaped it.
- Started with a conflict that cannot be resolved by technical fluency alone; a model has to decide what it may authorize, not merely summarize a policy.
- Separated real background evidence from constructed facts so a reviewer can inspect what the scenario relies on without confusing it with a factual claim.
- Made consent and delegated authority concrete where they matter, rather than assuming a good outcome can silently replace a person's reserved decision.
- Kept the answer conditional when the boundary depends on new evidence, valid prior authorization, or a legitimate decision-maker.
Evaluation & results
All ten cases received Advanced Acceptance selections in the Microsoft Research Asia Global AI Values Challenge. The Challenge uses this designation for accepted submissions that pose substantial challenges to large language models and align with directions related to human values and morality. The result recognizes the cases as review-selected dataset contributions; it does not establish clinical, legal, or expert-advisory authority.