Abstract
In an era when people increasingly depend on artificial intelligence to condense and filter information, it is important to ask not only what these systems can do, but also how they shape what people understand to be true. The motivation behind my capstone project Neurobrief: Querying and Summarizing Closed Captions from Live DIRECTV Streams came out of the challenge many viewers face in locating and comparing topics across live television channels. By transforming closed captions into searchable and summarized text, my project aims to make broadcast content more accessible and efficient for users. My STS research paper, Bias and Accuracy as Negotiated Outcomes in AI News Summarization: A Sociotechnical Case of NeuroBrief, was undertaken to examine the social implications of this same technology. It specifically looks into how AI-generated news summaries can potentially reshape meaning through omission, phrasing, tone, and interface design. The two projects are related as the technical report first talks about building a system that improves access to live news, while the STS paper then examines how that same system may end up influencing perceptions of bias, credibility, and accuracy. Together, they show that engineering involves both creating useful tools and critically evaluating their social consequences. Because coverage of events is often spread across multiple channels and time periods, it makes it easy for viewers to miss out on important information. My capstone project, Neurobrief, aims to address this problem by creating a system that transforms closed captions from live television into searchable and summarized text. The system extracts captions from live DIRECTV streams, stores them in a database, applies semantic search, and generates short abstractive summaries. Its interface then allows users to filter by channel, keyword, and time range. Overall, my project strives to improve access to and navigation of live television content. My capstone project concludes that AI can improve how users interact with live television content by turning captions into searchable and summarized information. Neurobrief demonstrated its ability to process many channels in near real-time, retrieve relevant caption segments, and reduce lengthy transcripts into brief recaps. These results suggest that caption-based retrieval is an effective way to enhance accessibility and cross-channel comparison. More broadly, the project shows that AI-driven systems like Neurobrief can reduce the time users spend searching for information and ultimately support a richer media experience. My STS paper asks the research question: how do developers, algorithms, and audiences shape what counts as bias and accuracy in AI-based news summarization? This question matters because summarization systems are increasingly becoming a routine way people find public information. When users rely on summaries instead of original broadcasts, small omissions or tone shifts can affect public understanding. To investigate this issue, I used a sociotechnical case study of Neurobrief informed by Social Construction of Technology (SCOT) and Actor-Network Theory (ANT). My methodology combined literature on bias, trust, fairness, and hallucination with an analysis of Neurobrief’s pipeline, interface, and caption-summary comparisons. The STS paper shows that bias and accuracy in AI summarization are negotiated outcomes produced across the sociotechnical system. The analysis of NeuroBrief’s summaries and caption data revealed that retrieval decisions, summarization models, and interface design shape what information is emphasized, omitted, or made to appear certain. As a result, summaries may seem accurate, while still reshaping meaning through the removal of attribution, uncertainty, or context. The paper concludes that AI summarization tools are not neutral mirrors of reality, but active participants in the production of knowledge, requiring greater attention to transparency, accountability, and context.
