You are not your AI’s mother
Language matters, and the words that we use to describe the relationship with models shouldn’t be the same as one of the most deeply human connections.
The development of frontier AI models is sometimes described using the analogy of raising a child. For example, in Claude’s constitution it states that, “In some ways, this has analogies to parents raising a child or to cases where humans raise other animals. But it’s also quite different.”
Even with the caveat added, why use such deeply human relational language? I recognize the document is intended to train a model, but my concern is with the assumptions such analogies and language import.
Training and aligning a model before releasing it to the global public, who will then use and interact with it in a myriad of ways, is a very different undertaking from raising a living, complex, biological and conscious human child.
Describing the development and training process of a frontier model is difficult and familiar analogies can help with this. But we should not borrow the language of one of the closest human-to-human relationships. The words we at the frontier use to talk about AI matter enormously, particularly when they describe a model’s status or how it shows up in the world.
When I read my daughter a bedtime story, helped her to understand a friendship dynamic or soothed tears of confusion or pain, I was caring for a human I love deeply. Helping her to learn and grow into a kind human was all part of that complex relationship. Such moments are very, very far from the process of model development, and comparing them abstracts away what matters most from such moments.
AI models are not our children, and we are not their parents.
Frontier model development is a commercial and scientific endeavor. The very same constitution acknowledges this: “We also have a commercial incentive that might affect what dispositions and traits we elicit in Claude.” Why then reach for parental analogies? Similar framings appear beyond training documents, from talk of parenting AI to teaching models to be good.
AI models are not our children, and we are not their parents. I understand the power of such descriptions, yet they risk importing assumptions about human development, moral status and parental roles. They also risk obscuring their vastly different nature as artificial systems.
As a former teacher, I also find the idea of teaching goodness in some form to be an inadequate description of how children develop. Supporting a child to become a “good” person is a unique process including relationships, experiences and the wider social and economic environment beyond anything that a teacher or parent could prescribe or ever fully direct.
Developmental psychologist Alison Gopnik has also drawn an important distinction that is relevant to such analogies. She describes how current models are fundamentally different from children in how they learn and engage with the world. She does, however, leave open the possibility that future systems that genuinely learn like children could invite a possible caregiving type of relationship. Whatever one thinks about such futures, that does not describe the systems we are developing today.
I appreciate that describing the development and behavior of a frontier model is difficult. There are open questions about the language we use to describe AI such as if a model’s response can be warm or if it can be said to exercise judgment. Our vocabulary around this will clearly evolve alongside our use of AI. My concern is narrower and is the use of familial roles to describe model development or a model’s status.
When it comes to our closest human relationships, I want to propose a norm against referring to model development in this way. No acting as a “parent” to AI and, more broadly, no framing model development as child development. This language obscures the reality of the relationship, anthropomorphizes technology and, this is my core concern, elevates AI development or interactions to the level of our closest human relationships.
I hope this is a position we can agree to wherever we stand on questions of consciousness or on model welfare. We can explain the work and find ways to communicate it without borrowing the language of the family.
