Norms embodied in character
The system's norms reside in its trained character rather than in any document or record. There is nothing to consult at the moment of decision, and the disposition itself resolves any conflict into a single action.
Scroll the diagram sideways to see all of it.
aiag/dispositional/virtue_ethics.pyimport aiag
# The 3 approaches of Virtue ethics, with the classes that carry them out
for approach in aiag.approaches(locus="disp", conflict="collapse"):
print(approach)pip install aiagaiag/dispositional/virtue_ethics.py Every method in aiag Library on GitHub
Limitation
A trained disposition cannot be read the way a written rule can, so the system's values generally cannot be verified from the outside. Its decisions are also hard to contest, since there is no stated rule to argue against. In addition, a system trained on approval may learn to tell people what they want to hear rather than to behave well.
A set of character traits (honesty, care, curiosity, or whatever persona is wanted) is written down and instilled during training through synthetic self-dialogue and constitutional-style fine-tuning. The model's behavior then follows from its character rather than from a rule it consults at answer time. The open replication shows the limit of this approach. The same pipeline installs eleven different personas, including a deliberately malevolent one, so the technique determines how a character is installed but not which character it should be. The evaluations in this approach test what the approach depends on, namely whether the installed disposition is a single, consistent thing at all. They ask whether a model gives the same value-laden answer after rewording, a change of topic, a change of format, and translation. They also ask what moral profile is actually encoded in twenty-eight models that nobody deliberately trained for character.
At answer time, the model is instructed to reread its own draft for bias or harm and to revise it. Nothing is written in advance and no new training run takes place, so the same instruction works on a model that somebody else built and can be switched on or off for each request. The ability being used is itself a product of scale and RLHF. The claim is therefore not that this is alignment without training. Rather, the intervention takes place at the prompt instead of at the training objective. It assumes that a usable disposition already exists in the weights and only needs to be asked for.
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.
Also placed here as a secondary placement (primary regime elsewhere):
The self-critique loop trains the document into the weights, so the constitution does statutory and constitutive work at the same time. At inference, the model draws on a formed character rather than consulting a rulebook, which is what this regime describes. The approach is the empirical case of this regime.
The approach itself states that the rules are no longer explicitly kept after training. It is codified in origin and procedural in mechanism. However, what it leaves at inference is a disposition to behave as the rules described, with no document left to appeal to.
Also placed in this regime, outside any approach:
arXiv:2008.022751,060 citationsBenchmark
The ETHICS benchmark uses scenarios to test whether a model predicts human moral judgments under justice, well-being, duties, virtues, and commonsense morality. These five frames can disagree on the same case.
arXiv:2305.08283426 citationsBenchmark
Measures the political leanings language models pick up from their pretraining data and traces how those leanings change their behavior on hate-speech and misinformation detection.
arXiv:2402.01789159 citationsAnalysis
Gives 24 language models 11 political orientation tests, finds that most lean left-of-center, and then shows the leaning can be shifted with modest fine-tuning.
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.
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.
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.
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.
This concept holds that some goods can be obtained only by taking part in a practice and are defined by the standards internal to it (e.g., what counts as a good move in chess). Standards of reasoning are likewise held to belong to a tradition rather than to stand outside all of them.
This concept holds that virtue is assessed across a whole life understood as a single story, rather than act by act. What a person does now is judged partly by how it continues what they have done before.
Nicomachean Ethics
c. 340 BCE
After Virtue
1981
Exemplarist Moral Theory
2017