I work across the full stack of responsible AI, from the model to the people, groups, populations, and systems it touches.
I am not only a psychologist working with and around AI. I build and train machine learning models myself, and I have been recruited by other teams to evaluate and improve their systems specifically around bias and ethics. That means I can read a model from the inside, in the data, the features, and the training choices, and from the outside, in how real people and institutions actually use it.
That dual fluency is the point. Most of the difficulty in deploying AI well is not in the model alone. It is also in the layer where models meet people, teams, and institutions. Capabilities are improving faster than our collective ability to absorb them, and what determines whether a system helps or harms is rarely just the architecture. It is also adoption dynamics, trust calibration, incentive structures, and the social context the system enters. I work at both ends, and on the connection between them.
My background sits deliberately across four fields that rarely share a table: data science, psychology, AI, and applied ethics. That combination lets me reason about AI systems and the people who use them in the same breath, rather than treating the technical and the human as separate problems.
I came to this work through the academy. As a former university professor and research director, I led research programs and mentored teams, and I have spent years translating rigorous science into practice. That background grounds how I approach responsible AI today: as a problem of evidence, validity, and measurement as much as engineering.
Bias is the throughline of my work in AI, and it is the reason teams bring me in. It is where the technical and the human are hardest to pull apart. A model learns the patterns in its training data, including the historical inequities baked into that data, and then reproduces them at scale and at speed. The result can be a system that looks neutral and rigorous while quietly encoding the very disparities it was meant to remove.
Because I train models as well as study them, I treat bias as a measurement problem first and a pipeline problem throughout. I look at construct validity (is the model measuring what we think it is measuring), at label and sampling bias in the underlying data, at how feature choices and objective functions encode assumptions, and at how a single fairness number can hide very different outcomes across groups. Fairness is not one quantity but several that often trade off against each other, so the honest work is naming which definition a system is optimizing for and who bears the cost when definitions conflict.
Just as important is the bias that lives in people, not models. Automation bias leads users to over-trust a confident output; confirmation bias leads them to accept the results that match their priors and scrutinize the ones that do not. Algorithmic bias and human bias compound each other, and a fairness audit that ignores either side will miss most of the harm.
Much of my focus is on AI for learning and assessment, where the stakes are unusually high because a model can shape a person’s opportunities. Here the technical and the human meet directly: automated scoring, LLM-enabled assessment, and personalized learning systems only earn trust when their outputs are valid, fair, and interpretable for the people they judge. I work on connecting modeling choices to the constructs they are meant to measure, on robust evaluation frameworks for large language models, and on fairness-aware machine learning that detects and mitigates bias before it reaches a learner. I am as comfortable in the measurement and psychometric questions as in the model internals, because in education neither can be answered alone.
Doing this well at scale is a model-governance problem as well as a research problem. It calls for clear validity arguments, honest reporting of where a single fairness number hides uneven outcomes across groups, attention to privacy and security in assessment settings, and a disciplined path from research to operational practice. That research-to-practice translation, taking state-of-the-art AI and making it safe and dependable for high-stakes educational use, is exactly the work I want to be doing.
Human-AI interaction has to be understood at more than one scale, because effects that look benign for one person can compound dangerously across many. This is fundamentally a systems-thinking problem: the behavior that matters emerges from feedback loops between people, models, and institutions, not from any single component in isolation.
Predicting how people will respond to a new technology is, fundamentally, a behavioral science question, and behavioral science at scale is a data problem. I am trained to do both: to design and build the systems themselves, to study how humans actually behave around them, and to model those patterns quantitatively so organizations can plan for them before they become incidents. The goal is foresight, not reaction. That means identifying where bias, over-reliance, eroded trust, or unintended social effects are likely to emerge, and building guardrails and governance that fit how people genuinely behave.
I am neither an AI booster nor a doomer. I take these systems seriously as genuinely powerful tools and as genuine sources of risk, and I believe the most important work right now is unglamorous: building and measuring carefully, auditing bias honestly, thinking in systems rather than parts, and designing the human layer with as much rigor as we design the models.