I came to AI research from psychology, after noticing that the methods I had built to read what a person is doing underneath what they are saying transfer almost directly to language models. A large language model holds information in a high-dimensional vector space and emits language as a translated representation of its internal state; that is close to how people work, where speech is the surface and cognition is the thing underneath it, each one a reflection of the other. Once that parallel is in view, a model starts to look like a subject you can assess.
The computer half I come by honestly. Both of my parents were software engineers, our house had the first desktop computer on the block by a wide margin, and my father built our first television from a CRT Heathkit; I will leave the size of that margin to your imagination. The psychology half arrived just as early. Doctor Sbaitso left a mark on me, though not as a therapist; what I wanted was to do assessment at scale and with implicit measures alone, and that combination needs two things, a computer and a good source of psychometrically meaningful behavior. Language is the most available source there is, where content, form, deviation, and something as small as which pronouns a person reaches for all carry empirically validated signal.
My dissertation is where the computer, the language, and the psychology meet most cleanly. Perpetrators of mass murder cannot be interviewed or tested after the fact, so I worked with the record they left behind, their pre-attack writing, and asked one question, whether the quantitative structure of that language, scored independently of what it says, could predict how its author would act. Two distinct behavioral typologies separated out of the clustering, and the real attacks bore them out, with the groups tracking victim type and number, how the event ended, and where it was carried out. That is a different lens on mass violence, and on the psychological states underneath it, than the field has usually brought to the question.
Before any of the research, my first grownup job was as a 911 operator and police dispatcher for the second largest sheriff’s office in the country, covering more than nineteen municipalities. It set the habits I still work by. Declining to decide is itself a decision, every decision carries consequences, and the work is choosing which consequences you are willing to live with while learning to see around the corner before you reach it; to the limit of what is reasonable, I trace a choice all the way down its likely downstream effects before I commit to it. My sergeant said it more plainly. Pretend everything you do will be tomorrow’s front page, and decide what you want the headline to say.
The robopsychologist title is half a joke, and I keep it because the other half looks more plausible every year. We are getting close to a point where these systems will need people who understand their cognitive processes, their behavioral drivers, and their “personalities” the way a clinician understands a patient; if we mean to treat this as technology we can steer, we first have to understand what makes it behave as it does, find the predictors, and act on them before the behavior nobody wanted has a chance to surface.
If the range here reads as a lot, that is the ADHD, and it is also how I get to spend a morning on forensic psychology, an afternoon on AI alignment, and an evening building a retro-futuristic, LLM-powered neo-noir detective game in the spirit of late-eighties LucasArts, while keeping the threads tied together (and collecting facts I never needed along the way; ask me sometime why men started wearing pants, the answer is better than it has any right to be). I paid for part of my doctorate by DJing clubs, and when I am not profiling language or building measurement systems I am cutting paper art with a very large laser, doing cloisonné badly, building furniture whenever I can find where my husband has hidden the tools, and becoming a temporary expert in whatever I turned up on YouTube that week.