Can AI Agents perform clinical tasks?
Ashkan Khilwatgar ’28, Suning Hong ’27, former Hamilton student Simon Li, and Wenbo Wang, visiting assistant professor of computer science
Genetic variant classification requires scientists to manually collect and evaluate multiple forms of biological and clinical evidence, which can be time-consuming. This project, AI Agents for Scientific Discovery, tested whether AI agents could perform complex clinical workflows without the risk of hallucinations and unstable behavior. Specifically, the Hamilton researchers designed an AI agent workflow using LangGraph to retrieve evidence, apply individual criteria, check intermediate decisions, and combine the results into a final classification.
Using the disease Hereditary Hemorrhagic Telangiectasia (HHT) along with a large portion of its variants, the team coded tools that the AI agent used to classify each variant as pathogenic, likely pathogenic, benign, likely benign, or variant of uncertain significance. The result shows improved accuracy compared with a zero-tool LLM baseline, demonstrating the feasibility of using an AI agent to automate variant classification.
The result also revealed that non-automatable criteria (criteria that cannot be automated because they require input of patient-specific data), inaccessible internal laboratory data, and a performance cap in the functional assay retrieval process contribute to most errors in the final classification label. So the results of their research suggest that while LLM-based systems can automate substantial parts of the variant-interpretation workflow, reliable final classification still requires better handling of difficult criteria and ambiguous evidence.
Can AI make ethical decisions?
William Ford Collins ’27 with Justin Clark, associate professor of philosophy
Almost all corporate industries have begun to employ machine learning and AI in some capacity in their day-to-day operations. Often, this application lends itself well to the goal of making the individual employee more efficient and capable. In other areas, especially those requiring a philosophical consideration of particulars and abstract generalizations, the application of AI remains contentious.
To investigate the possible deficiencies of AI, this research revolved around developing a DeBERTa-based binary classification model capable of assigning an “acceptable” or “unacceptable” verdict to ethical scenarios. To train the model, Collins developed a dataset of some 3,800 philosophically labeled samples that worked to emulate the principles of Immanuel Kant’s deontology and John Mill’s utilitarianism. From the model’s output, they found it failed to correctly apply universal maxims across situations and also substituted the severity of an act for the aggregative comparison of harm and benefit that utilitarian calculations require.
Based on these findings, Collins contends that AI remains unable to reason and judge a human agent’s actions with the same care and consideration that would be applied by a philosopher. Further, he extends this claim by analogy to demonstrate why machine learning ought not to be applied to legal decision-making.
Can AI improve alternative communication tools?
Sarah Morrison-Smith, assistant professor of computer science, with Madeline Brogen ’28, Royce Carol ’28, Gwen Sawicki ’28, and Han Dong, assistant professor of computer science
Millions of people who cannot rely on speech use augmentative and alternative communication (AAC) tools to express themselves, from tablet-based symbol boards to text-to-speech systems. As generative AI works its way into these tools, it raises a pressing question: Do AAC users have more or less control over their own voice? Morrison-Smith has been collaborating with Blade Frisch at Bowling Green State University to lead a study that puts AAC users at the center of that question. The project AI, Voice and Choice: Centering AAC Users in GenAI Adoption examines how people who communicate through AAC understand and evaluate the move from predictive text toward generative AI.
The ongoing study pairs a five-week online focus group and interviews with AAC users with a broader survey distributed through AAC communities. Early analysis has drawn out themes that matter deeply to participants, including autonomy over their own words, concerns about ableism, and the desire for technology that reflects individual personality rather than flattening it.
Does sustainability influence AI tool choice?
In a second project, Carbon Footprint Disclosure and LLM Model Choice, Morrison-Smith and Han Dong are studying how people make decisions about the AI tools they use and what role sustainability plays in those choices. Their work brings participants into guided writing sessions where they work with a range of AI writing tools and reflect on their experiences. Using a mixed-methods approach that combines survey measures with in-depth interviews, the study looks at how people weigh factors like output quality, ease of use, and broader values when they decide which tools fit their needs.
When sustainability is not perceived as something the research team is interested in, participants prioritize AI performance over the environment, despite noticing their carbon footprint. However, when asked about the environment, participants attest to being eco-conscious, in opposition to their earlier behavior.
Can better policy fight AI misinformation?
Maxwell Gold ’28 with Ashley Gorham, assistant professor of government
Gold focused on the dangers of AI-generated political misinformation and how to manage its influence. By examining case studies on current efforts to regulate online misinformation — such as the rollout of Community Notes on X — he built a case for not censoring misinformation, but instead allowing viewers access to more information should they choose to pursue it.
Gold cites the philosopher John Mill, who argued that no opinion should be censored or ignored as it often has some aspect of truth, or, at the very least, people should be allowed to rebut it. This carried over to AI regulation. After examining policies already enacted to combat AI misinformation, Gold drafted a proposed federal policy that would require AI providers to include clear disclosure on any political content created using their program. This would mean that any GenAI system must be programmed to insert a digital marking, as well as visual indication, that any piece of political media it creates is artificial. This disclosure would be presented in a way that cannot be cropped, edited out, or otherwise hidden when the person generating the content shares it on another platform.
Although this would mean that artistic creations may also be subject to a visual marker, the risks of leaving an open-ended interpretation of what constitutes an artistic work would allow many pieces of media, created with the same intentions or look as deliberately misleading content, to have the same effect of that content. The disclosure policy wouldn’t ban such media — whether for political messaging or creative expression — it would simply tell viewers that the material is artificial. The goal of this regulation is to give people the necessary information they need to form informed opinions.
Posted October 13, 2026
