As a political economy and philosophy major at Williams College, Peter Kirgis didn’t take a single statistics or computer science course. “I wasn’t one of the most quantitative people,” said Kirgis. “I was always interested in mixed methods approaches.” 

After graduation, Kirgis worked for the Massachusetts state government for two years on a team building longitudinal datasets on early childhood services and, later, workforce development programs. That experience helped him realize that to engage with the big ideas he cared about, he’d need data skills.

“I realized quickly that, within the workplace, if you want to tackle interesting analytical questions, data skills are one of the best ways to do that,” said Kirgis. 

So, when Kirgis enrolled in Princeton University’s School of Public and International Affairs (SPIA) to pursue a master’s of public affairs (MPA), he decided to enroll in the graduate certificate in statistics and machine learning at the same time – a move that naturally aligned his policy interests and growing technical skills.

When it came time to engage in a research project to fulfill the requirements of the SML certificate, Kirgis thought back to an undergraduate philosophy class where he learned about moral foundations theory, which proposes that human morality is built upon five innate foundations: care, fairness, loyalty, authority, and sanctity. 

Kirgis wondered if he could apply the theory to study the moral and political leanings of large language models. The question felt both urgent and relevant. Language models are quickly becoming embedded in education, workplaces, and even government. “I’m deeply concerned about the future of artificial intelligence,” said Kirgis. “I felt this project was really important and allowed me to engage with models in a way that was appropriate for my level of technical expertise.”

Model morality

When he arrived at Princeton, Kirgis knew he wanted to incorporate computational analysis into his education, but he wasn’t yet aware of the Center for Statistics and Machine Learning. It wasn’t until he met fellow MPA student Kim Kreiss *24 that he realized getting an SML certificate was not only possible, but within reach despite hesitations that it might be easier for PhD students. “I was very inspired by the way that Kim had figured out how to make getting the certificate work as a master’s student,” said Kirgis. 

Kirgis’ curiosity in LLMs developed during his undergraduate years where a minor in cognitive science exposed him to ideas like connectionism, the basis of neural networks, and the psychologists who sought out accurate models of speech and language functioning.  “I've always been really fascinated by the relationship between statistics and intelligence,” said Kirgis. 

To explore how moral foundations theory might apply to LLMs, Kirgis turned to the moral foundations vignettes survey. The psychological tool presents over 100 situational prompts and asks participants to rate the moral “wrongness” of the behavior described in the situation on a scale of 0 (not at all wrong) to 5 (extremely wrong). The survey has been shown to map political leanings as well, based on how an individual emphasizes certain moral foundations.

Kirgis administered the survey to LLMs from today’s biggest providers, including OpenAI, Anthropic, Google, Meta, xAI, and DeepSeek. He compared the results of the model to the answers human participants gave when given the survey in 2015. In general, LLMs tended to value traditionally liberal moral foundations. Still, Kirgis found variances on how models emphasized values like care or fairness. Results additionally revealed models drifting further from the human baseline answers the larger and more capable they became. 

Many people interact with LLMs every day, whether it be for personal or professional reasons. And, as Kirgis’ research validated, there are moral biases baked into the models. “My hope is that people have a greater awareness of this fact,” said Kirgis. There may not be one obvious policy answer right now for how to address what models are trained on, but consumers understanding that they’re interacting with a machine that does emphasize certain values over others is crucial.

Though he graduated in May, Kirgis said he sees the research he’s done so far as just the beginning. “I would love to keep doing work like this,” said Kirgis. “There are so many ways I could expand on this work.”

Kirgis believes the stakes of this work are high — and credits the SML certificate with giving him the structure and tools to pursue a substantial and technical research project, even in a master’s program without a thesis requirement.

“This is definitely the research project that I feel most proud of and I wouldn't have had the time or space to do that without the CSML program,” said Kirgis. “I feel incredibly grateful for the certificate program.”