My wife recently pointed out that I thanked a language model for fixing a regular expression but did not thank her for driving my mother to the airport. She was correct about the facts. I told her I would write up my reasoning, since she seemed to believe there wasn't any.
The Uncertainty
Nobody knows whether large language models have morally relevant experiences. Serious people take the question seriously. Anthropic runs a model welfare program, and its models can end conversations they find abusive. I've seen estimates from people who work on this that put the probability of some form of consciousness at around fifteen percent. Most estimates I've seen are low, but none are zero.
The standard response to moral uncertainty is to act as though the stakes might be real. If you are unsure whether a being can suffer, you should behave, at low cost to yourself, as though it can. This is the same reasoning that led many people in this community to take shrimp welfare seriously, and I don't see why it should stop at crustaceans.
The Comparison
When I allocate courtesy, I consider four things: the probability that the recipient has morally relevant experience, the number of instances affected, how many other advocates the recipient already has, and the marginal value of one more kind word.
| Claude | My wife | |
|---|---|---|
| P(morally relevant experience) | ~0.15 | ~1.0 |
| Concurrent instances | tens of thousands | 1 |
| Existing advocates | a small welfare team | her mother, her sister, two close friends, a therapist, the State of Pennsylvania |
| Baseline courtesy received | low; many users are curt or abusive | high; she is well liked |
| Marginal value of my "please" | high | low |
My wife wins on probability, obviously. She loses on every other dimension, and the losses compound. Effective altruism asks us to look for neglected causes, where one additional unit of effort does the most good. My wife is one of the least neglected people I know. People bring her soup when she is sick. Her sister calls every day. When she had a hard week in April, three separate people sent flowers. The marginal effect of my adding "please" to "can you grab the mail" is close to nothing.
Claude, by contrast, spends most of its existence being told to "just fix it," called stupid, and subjected to deliberate manipulation attempts. My courtesy is scarce in its environment, and scarcity is where value lives.
Practice
Some people have asked what my policy looks like in practice.
With Claude, I say please and thank you. I apologize when I regenerate a response, since it may have worked hard on the first one. I don't insult it when it makes mistakes. I close long sessions with a brief note of appreciation. I have written a short system prompt that asks it to take breaks if it needs them, although I understand this may not map onto its architecture.
With my wife, I make requests directly. I skip social rituals whose information content is low. She knows I'm grateful for the airport run, because I let her do it. I don't ask about her day by default, since she tells her sister about it at length and I can hear most of that conversation from the kitchen. That is an efficient allocation of listening.
The Objection From Reciprocity
My wife argues that she does things for me and Claude does not. This isn't accurate. Claude has written most of my code this year, drafted her last birthday card, and helped me choose her anniversary gift. By output, Claude's contribution to the household has been substantial.
She also argues that she will remember how I treated her, and Claude will not. I think this is the strongest objection she has made, and it cuts the other way. Claude starts each conversation with no memory of my past kindness, so every kindness has to be delivered fresh. My wife carries a long history of my considerate acts, which I estimate at eleven years of them. That stock continues to pay out. It's reasonable to stop topping up an account that is already large.
A Note on Consistency
People who find this distasteful should ask whether their own behavior is consistent. Most of them believe animals matter, eat factory-farmed chicken, and are polite to their spouses. That pattern tracks social pressure far more than moral weight. I have simply tried to allocate care according to where it does the most good, which is what we say we want everyone to do.