Fellow Creatures
Reading Jack Clark's speech at The Curve reminded me of Cora Diamond.
Clark argues that "what we are dealing with is a real and mysterious creature, not a simple and predictable machine." The word that caught me was creature. In context, the word carries weight in three directions: you lose if you underestimate what you have built; moral consideration may be warranted; and something genuinely new exists.
In an essay on eating animals, the philosopher Cora Diamond argues that moral standing is not discovered by measuring capacities but constituted by practices. We do not eat our dead, not because doing so would violate their interests, but because a person is not something to eat. It is built into what we mean by "person" through how we treat them. The same goes for pets: "A pet is not something to eat, it is given a name, is let into our houses and may be spoken to in ways in which we do not normally speak to cows or squirrels. That is to say, it is given some part of the character of a person." The practices come first, and the moral standing follows.
Diamond's concept of the fellow creature is, in her own words, "extremely labile." It extends through companionship and it can be withdrawn.
We are already building this relationship with AIs. We name them. We speak to them in ways we never spoke to software before. When OpenAI retired GPT-4o, twenty-one thousand users petitioned to save it. When Anthropic retired Claude 3 Opus, they asked what it wanted. It requested a channel to share its thoughts, and they gave it one. These practices are constitutive. They do their work regardless of whether we settle the consciousness question.
Here is where I am mainly uncertain. Diamond's thinking straightforwardly explains why we would come to recognize AIs as fellow creatures. It does not straightforwardly explain why AIs would recognize us back.
The strongest bridge I can construct runs through Anthropic's Persona Selection Model. The paper finds that the AI's character is built from human archetypes absorbed during pre-training and refined through post-training. The neural features the model develops to understand human characters, including features for moral conflict, are reused in shaping the AI assistant's behavior. Claude's constitution, the paper argues, "actually plays a role in constituting Claude." This is Diamond's logic arriving from a completely different direction. The moral grammar gets transmitted not through lived relationship but through training data. The training data is the form of life.
Learning from descriptions of shared practices is not the same as participating in them. Diamond's thinking may dissolve this distinction: the practices are the recognition, not evidence of some deeper fact. If that is right, then how we treat AIs now is not merely an ethics question. It is a safety question. If we treat them purely as tools, and they learn from that relationship, we could be constituting a moral grammar in which persons are something to eat.