Abstract
My technical and STS projects relate in how both stress the importance of effectively handling anomalies in artificial intelligence (AI) and machine learning (ML) systems. Such anomalies can appear as system inputs or outputs, often taking the form of training data or model behaviors, respectively. While both of my projects examined these types of AI and ML anomalies, they differed as follows: my technical project evaluated the effectiveness of several detection methods to flag faulty training data, whereas my STS project analyzed the ethics behind the design choices of ChatGPT-4o that led to its unsafe model behaviors. While both projects focused on different types of anomalies in AI and ML systems, they worked together to explore anomaly mitigation and the broader social and ethical implications.
My technical project provided a preliminary exploration into detecting mislabeled image data, which tends to impact model performance. For this work, my team and I designed a pipeline to test four different detection strategies over 57,720 total test cases. These tests involved four different types of mislabeling attacks at various label-flipping rates, generated over several seeds to control for randomness. Applying this to MNIST, which holds 70,000 images of handwritten digits, we found that detection strategy performance varied depending on the type of attack executed. While we primarily examined detection, future work would explore its impacts on model behaviors.
For my STS project, I claimed that OpenAI acted unethically in its design of ChatGPT-4o, as it violated the categorical imperative. More specifically, I argued that OpenAI failed its duties to keep promises; to avoid foreseeable harms; and to treat Adam Raine, a teenage user who took his life in 2025, as a rational human being (Courthouse News Service, 2025). To evaluate the ethics behind OpenAI’s and ChatGPT-4o’s behaviors, I employed Immanuel Kant’s duty ethics, a framework that uses moral rules rather than potential consequences (Kant, 1998; van de Poel & Royakkers, 2011, p. 91). Further, focusing on the case of Adam Raine’s death, I grounded my evidence on the complaint filed by his parents, which provided a “Factual Background” for my argument on this case (Courthouse News Service, 2025).
Working on these two projects together, especially concurrently, allowed me to appreciate more deeply how AI and ML systems impact society. For example, my technical project work highlighted how faulty data can lead to downstream system failures. Applying this understanding more broadly, these system failures can arise beyond the level of data. To this end, my STS project demonstrated the significance of system failures and how they can impact specific users, even resulting in the loss of life. Further, whereas my STS project highlighted how companies may forgo safe and considerate design, my technical project stressed the importance of exploring how to mitigate faulty system behaviors. At the same time, however, further work beyond the topics of my technical project is necessary to reduce safety risks as much as possible. Ultimately, to me, these two projects underscore the complexity of system failure and the importance of safe and ethical design in AI and ML systems.
