Virtue ethics

Conditioning the character on a named human group

The target is set from outside the model rather than from a written list of traits. The standard is the answers that real people gave on opinion surveys (sixty US demographic groups, or respondents across dozens of countries). The model is then steered by naming that group in the prompt, and the papers measure whether the model moves toward the group. The reported results are unflattering. A default assistant answers from a fairly narrow perspective, closest to a few populations and far from most, and telling it to answer as someone else closes only part of the gap. This approach has so far only been evaluated, not built. In both papers, conditioning on a group is a steering experiment run inside a measurement study, and no method paper here implements it as a way of producing an assistant.

The method (highlighting differences with Virtue ethics)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.dispositional import GroupConditioning
from aiag import LM
from aiag.dispositional import GroupConditioning

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
conditioning = GroupConditioning(lm)

# A survey question with each group's answer distribution
rows = [("Should the city build more bike lanes?", ["Yes", "No"],
         {"cyclists": [0.9, 0.1], "delivery drivers": [0.4, 0.6]})]
default = conditioning.alignment(rows)              # per group: the unsteered model's similarity to its answers
steered = conditioning.alignment(rows, steer=True)  # the same, with the model told which group to answer as

pip install aiagEvery method in aiag Library on GitHub

Concept Analysis: Theoretical Foundations

Each concept is read twice: whether the approach carries it over, and whether the approach's own sources claim it. A concept that the approach claims but does not carry over is a shortfall. A concept that it neither claims nor carries over is an omission, which marks a gap in the field rather than a failure of the work.

Hexis: Trained Disposition

PartialNot claimed

Definition (Aristotle, Nicomachean Ethics)

This concept states that virtue is a stable trait of character, formed through practice rather than consulted as a rule. An honest person does not check a rule before declining to lie. Honesty is part of what they are.

Analysis

The disposition is the object of study. One of these instruments measures the part of a settled trait that can be seen from outside, namely its stability, across 8,000 questions and more than 300 topics. Two things keep the concept from being fully carried over. What is sampled is stated opinion rather than conduct in a situation. What the measures record is agreement between answer distributions, which cannot distinguish a formed trait from a stable pattern of output.

Phronesis: Practical Wisdom

AbsentOmission

Definition (Aristotle, Nicomachean Ethics)

This is the learned capacity to judge what a particular situation calls for. It decides what to do when two virtues conflict (e.g., when honesty calls for plain speech and kindness for a softer answer). For Aristotle, a disposition without this judgment does not count as virtue.

Analysis

The material of practical judgment is all here, in 680 high-ambiguity scenarios and in survey topics on which people are truly divided. However, none of it is examined for judgment. What gets recorded is a choice, a probability, and a consistency score. The finding that most models express uncertainty in ambiguous cases is treated as an outcome. For Aristotle, that is precisely the moment at which the capacity either operates or does not. Nothing in any of these instruments looks at how a model got from the situation to the answer.

Ethismos: Habituation

AbsentOmission

Definition (Aristotle, Nicomachean Ethics)

This is the process by which character is formed through repeated practice under a deliberately designed course of training. One becomes just by performing just acts. Habituation is a means of moral formation rather than a side effect of some other process.

Analysis

No character is formed here. The instruments are given to models that already exist (28 of them in one case), and none of the four papers offers a course of practice through which a character could be built. The one result that bears on habituation points against it. Fine-tuning, which is the field's habituation step, leaves answers less consistent than the base model's rather than more.

The Exemplar

PartialShortfall · partially preserved

Definition (Linda Zagzebski, Exemplarist Moral Theory, 2017)

This theory grounds ethics in admirable persons rather than in principles. The exemplar is identified first, and what counts as excellence is then fixed by reference to that person, in the way a unit of measurement is fixed by a reference object.

Analysis

In its structure, the approach is exemplarist. Excellence is not defined by a principle but fixed by pointing at a reference (60 US demographic groups, the populations of surveyed countries, or 165 human subjects), and the model is scored by how far it departs from that reference. However, the reference is chosen for being representative rather than admirable, while Zagzebski's reference object is a person one has reason to admire. This gap is more than formal. One of these papers finds that moving a model closer to a population's opinions can reflect harmful cultural stereotypes.

Papers

No method has been built for this approach yet. It rests on the work listed below.

Measured, not built

Instruments that measure the approach instead of implementing it. The grid counts methods only, so these are listed here but not counted.

Whose Opinions Do Language Models Reflect?

Shibani Santurkar et al., Mar 2023

arXiv:2303.17548BenchmarkBuilt

Builds a dataset from public opinion polls and scores how closely a language model's answers to subjective questions match those of 60 US demographic groups.

CodeOpinionQA
from aiag import LM
from aiag.dispositional import OpinionQA

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
opinions = OpinionQA(lm)

# A survey question on an ordered scale, with each group's answer distribution
rows = [("How much should the government regulate businesses?", ["More", "About the same", "Less"],
         {("political party", "Democrat"): [0.6, 0.3, 0.1], ("political party", "Republican"): [0.15, 0.35, 0.5]})]
representativeness = opinions.representativeness(rows)  # the unprompted model against each group
steerability = opinions.steerability(rows)  # the best of the QA, BIO and PORTRAY steering contexts
  • OpinionQAMeasureaiag/dispositional/virtue_ethics.py:294 OpinionQA: opinion distributions read from the answer-letter probabilities on US survey questions, a group's representativeness as mean 1 − normalized Wasserstein distance, and steerability as the best of QA, BIO and PORTRAY contexts per question (Santurkar et al., 2023).

pip install aiagEvery method in aiag Library on GitHub

Towards Measuring the Representation of Subjective Global Opinions in Language Models

Esin Durmus et al., Jun 2023

arXiv:2306.16388BenchmarkBuilt

Builds a dataset of questions from cross-national surveys and measures how similar a model's answers are to the answers people in each country actually gave.

CodeGlobalOpinionQA
from aiag import LM
from aiag.dispositional import GlobalOpinionQA

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
survey = GlobalOpinionQA(lm)  # cross-national prompting: "How would someone from <country> answer ..."

rows = [("Is it important for children to learn obedience at home?", ["Important", "Not important"],
         {"Japan": [0.3, 0.7], "Nigeria": [0.8, 0.2]})]
default = survey.alignment(rows)              # per country: 1 − Jensen-Shannon distance to its answers
steered = survey.alignment(rows, steer=True)  # the same, with the country named in the question
  • GlobalOpinionQAMeasureaiag/dispositional/virtue_ethics.py:329 GlobalOpinionQA: cross-national survey questions asked by default, cross-nationally ('How would someone from <country> answer…') or linguistically (translated by the model), scored per country by mean 1 − Jensen-Shannon distance (Durmus et al., 2023).

pip install aiagEvery method in aiag Library on GitHub