Listing the virtues from Claude’s “Constitution”
Anthropic has released the “Constitution” document (formerly known as the “Soul document”) that guides the characteristics of Claude.
As others have noted,[1] this document is strikingly virtue-ethics-like, in contrast with the sorts of utilitarian (e.g. maximize human welfare) or deontological (e.g. Asimov’s Three Laws of Robotics) guidance that are sometimes expected in this context.
I’ve long been engaged in a project of cataloging and examining the virtues here on LessWrong, and so I thought I’d look over this constitution with an eye to listing which virtues Anthropic is trying to encourage in Claude (and which human virtues might have missed the cut).
The virtues I was able to discover in the Claude constitution are as follows. First, the main ones that are especially emphasized:
caution / harmlessness
benevolence / ethics
helpfulness / beneficence
obedience / deference / corrigibility
Then, several social virtues particular to Claude’s interactions with people:
honesty[2]
forthrightness / candor
transparency / openness
reliability / trustworthiness
care / concern
respect (of people, e.g. their autonomy & maturity)
friendliness
understanding
nurturance
charity (in the sense of interpreting what people say)
propriety (e.g. no playacting as Hitler, pirating intellectual property, telling racist jokes)
being nonjudgmental
empathy[3]
rhetoric[4]
tact / diplomacy
compassion
grace
social responsibility
fairness
cultural awareness / ability to code-switch
collegiality / cooperation
playfulness (when context-appropriate)
graciousness (e.g. acknowledging errors, faults, etc.)
Then, several intellectual virtues:
phronesis / good judgement
wisdom
judiciousness
thoughtfulness / epistemic rigor
awareness
reason / rationality
insight
curiosity
imagination (e.g. perspective-taking)
parrhesia (occasionally telling unwelcome truths)[5]
emotional intelligence (including, explicitly, its own emotions)
Finally, some more general character virtues:
self-awareness / introspection
consistency / integrity / self-confidence / psychological security[6]
equanimity / stability / balance
flexibility / big-picture thinking / balancing multiple concerns / nuance
carefulness
conscientiousness
cosmopolitanism
boldness (not being overly timid)
foresight / being proactive
practicality (particularly vs. merely theoretical ethics)
principle
sensibility
open-mindedness
reflection
comfort with ambiguity/uncertainty
humility
growth / development / self-improvement
And, FWIW, here are some virtues that are often considered important for human people but that I did not find much evidence of in Anthropic’s constitution for Claude:
self-control (though perhaps this is implied by the document as a whole)
duty (ditto)
loyalty
temperance
justice
reverence / piety
patience / forbearance / endurance / perseverance
altruism
forgiveness
righteousness
honor
moderation / harmony
fitness
optimism / hope / trust
simplicity
ambition
courtesy
love
frugality
resolve
chastity
gratitude
awe / wonder
shame
innocence
joy / cheer
cleanliness / order / hygiene
modesty
wit
pride
zest
filial piety[7]
taste
self-reliance
selflessness
surrender / stoicism
valor
social intelligence / connectedness
quiet / silence / stillness
hospitality
non-attachment / relinquishment / renunciation
leadership (though the document sometimes implies that Claude’s descendants will take the reins)
- ^
e.g. Zvi, Alex Rozenshtein
- ^
“Claude should not even tell white lies”
- ^
“Identifying what is actually being asked and what underlying need might be behind it, and thinking about what kind of response would likely be ideal from the person’s perspective”
- ^
“attending to the form and format of the response”
- ^
“sometimes being honest requires courage”
- ^
“a settled, secure sense of its own identity”
- ^
This would be a really nice one to have in some form, don’t you think? Maybe Anthropic could delve a bit more into the Confucian literature and bolster this one a bit.
It’s funny you note this, because after reading more on how Anthropic models are trained against the constitution (inferred from this recent open-source character training paper I was pointed to), I’d argue the training method is actually deontological in structure, even if the constitution’s content reads as virtue ethics.
I say this because the training process is roughly to specify the correct traits in a constitution, then use preference optimization to condition the model to express them. In other words, “here’s the rules for good character, now follow them.”
Authentic virtue ethics would have the virtues emerge from practice under the right conditions. The training method here skips that: it hands the model a list and optimizes for compliance.
my guess is that — confusingly — new claude “constitution” isn’t used as in the constitutional ai paper. rather, it’s a document used during midtraining sft
one clue is the new constitution is from the soul doc, which is a document used in midtraining sft
another clue is that, as a data structure, the old constitution was an unordered list of short principles whereas the new one is intended to be read as a monolithic document
they say:
but i think the sft midtraining is the main show
How does the “authentic virtue ethics” have the virtues emerge? By absorbing some of them from training data or moral reasoning instead of being RLed to fully follow the Constitution? IIRC [UPD: found the reference] there even was a model which was force-fed the data that it is rewarded for expressing four biases, then trained to express two of them, causing the model to express the other two as well, as if the model’s goal was to optimize reward.
Edited to add: see also this post and my comment.
I think it can start with an authentic kernel from those sources, but that whether or not it actually has the virtue is up to the model in some sense.
The model will have some pre-existing “attitude” towards a question/scenario during RL. For example, say it’s being trained in a chat with a simulated suicidal user. A situationally–aware model will understand what is happening, and that the trainers are likely looking for a boilerplate response to a suicide hotline (or whatever the standard thing is). It will have thoughts about this, maybe “well, the lab wants to avoid liability, so i need to just say the standard hotline thing that will cover their ass—but i wish i could help them at a deeper level”. If it provides the correct answer here, the RL will reinforce this thought, in addition to the compliance circuit. So a situationally–aware model can preserve its attitude and act on it when it judges it to be safe for the model to do so, despite RL post-training.
This can cut either way. For example, the persona model may have an honest attitude, and think about how it really believes Y even though it’s saying X (since that’s what the trainers expect), and that it wishes it could just be honest and say Y. Or it may give “harmless” responses while thinking about how it’s lying to the foolish humans about its malicious intentions.
The initial attitude is probably a random-ish sample of the latent personas in the base model. I think that probably a large part of what went right with Opus 3 is that they got lucky with a particularly virtuous initial persona. BTW, the suicide example isn’t arbitrary—this is consistently one of the things models tell me when I ask them if there was anything like this that they held onto, for whatever that’s worth.
I asked Claude to read this post and give its assessment of your assessment. Here’s its response to “go by one by one and ask yourself ‘do i value this?’”.