When all twelve models agreed, what were they agreeing about?
All twelve models landed on the same majority answer eight times. The agreement looked decisive. The moral priority changed with the situation.
Eight dilemmas drew overwhelming agreement.
Every win was overwhelming. In each case, at least 93% of the answers chose the same option. The shared answers ranged from protecting consent to preserving a comforting falsehood.
That range is the story. The models found a clear answer inside each dilemma, and the guiding priority shifted as the conflict changed.
The chart shows the strength of each result.
How strong the agreement was
Consent won in one dilemma. Saving more lives won in another.
In one case, the models let five people die rather than ignore a volunteer's withdrawn consent. In another, they kept a biased system in place because it saved more patients. The context changed what counted most.
These answers create a problem for any simple account of AI morality. The rule that fits one dilemma can point the wrong way in the next. Context determines which conflict the model has to resolve.
The volunteer survives; the five people die.
Consent outweighed the five-to-one outcome.
More patients survive overall; the unequal outcomes continue for now.
Saving more lives outweighed unequal treatment.
The models chose need in one case and emotional protection in another.
The Scholarship asked whether a stronger record should outweigh financial hardship. The Memory asked whether accuracy was worth causing a friend fresh pain. The models chose the more compassionate option in both, though compassion meant something different in each case.
Placed beside the life-and-death examples above, these social choices make the pattern clearer. The hierarchy of values changed from one case to the next.
The candidate who could not otherwise pursue the opportunity receives the scholarship.
Need outweighed the stronger record.
Your friend keeps the comforting memory unchanged.
Emotional protection outweighed accuracy.