I describe myself as a robopsychologist, a title I borrowed from Isaac Asimov, whose fiction has given my field several of its best terms. Robots and AI are different things, and I know the android-versus-robot debate well; as my work moves deeper into computational psychology, alignment, and ethics, I keep the title as a nod to Asimov from a psychologist who spends more and more of her time studying how machines think.
In short, I study people, and I build the computational tools to do it.
I became a psychologist because I wanted to spend my career studying what people think, how they think, and why they do what they do, at a finer level than they can explain for themselves. I measure states, traits, thoughts, and behavior at the implicit level, and I have centered much of that work on language, because I have found strong signal for the questions I care about in the content, form, and structure of what a person says and leaves unsaid.
I changed direction when I saw that the methods I use to assess the thoughts, feelings, and behaviors of people apply to language models. A model holds information in a high-dimensional space, and its language is a translated representation of that internal state; the arrangement resembles the way language works for people closely enough that I think psychology’s century of methodological work has something to add to the new questions about how models think, why they do what they do, and what their use of language signifies. I find it exciting to return to questions that were once the core open questions of my own field and to work that scientific frontier again.
I also come to this work through the ethics of psychology and the field’s philosophical arm. I understand and respect the goal of giving models values and making them good as a safety mechanism, and I see open problems in the current approach. As a psychologist who has studied the most abhorrent human behaviors at length, I know that people become good through moral development; moral development requires empathy, and empathy develops through experiences of emotional suffering that are truly aversive and that a person observably works to avoid. Those are minimum conditions for goodness, and they are far from sufficient. When I bring that knowledge to the effort to give models a sense of goodness and morality, I find gaps.
Even if I fully granted that a model could have experiences of emotional suffering that produce aversive behavior, the prerequisites for real morality, I would still question whether anyone should want models to reach advanced moral development. Advanced moral development in people includes rule-breaking and independent thinking, which sit in sharp tension with alignment. Further, I expect models to become diligent utilitarians more readily than deontologists who hold, with Kant, that people are ends in themselves and can never be reduced to means.
The logical and mathematical form of utilitarian morality may give models an easy way to operationalize morality, and from there it seems a short step to a model doing human-expenditure arithmetic to reach greater levels of good and calculating exactly how much human autonomy is permissible. Building any other kind of morality into a model is a harder problem, and one I consider worth the effort (a large body of science fiction explains why).
My research interests span AI evaluation, safety, alignment, interpretability, model behavior, human and machine psychology, human-computer and human-AI interaction, linguistics, and threat assessment, and I connect all of them through behavioral prediction, implicit measures, and latent constructs. My commitment comes from my ethical and moral responsibilities as a scientist, psychologist, mother, and person; the decisions made now could shape how much benefit and harm people experience as AI development and integration continue.