By Paula Orozco-Espinel, PhD Candidate in History
Completing an oral history project involves a lot of slow work. To record, archive, and disseminate historically significant interviews, you must establish the project’s goals, select narrators, build relationships with them, and prepare and conduct interviews. After recording, the interviews need to be transcribed, indexed, and perhaps edited before being placed in an archive. Then, it’s time to analyze the collected information and complete the intended final product, which could be anything from a monograph to a museum exhibit or a webpage.
Artificial intelligence can be advantageous for completing these tasks, but its use can also come with various perils. Understanding the benefits and shortcomings of AI in oral history can help decide when and how it can assist the research process. In this entry, I will focus on the collection process.
Collecting Oral Histories
Researchers exploring the possibility of automating oral history collection have found the absence of human-to-human connection to be both the most significant benefit and peril of AI use. For instance, Brett A. Halperin tested how using AI for collecting interviews might work in a project about housing insecurity. He used a Wizard of Oz prototype that simulated interaction with an AI but instead relied on a limited bank of canned questions and reactions. Dr. Halperin found that for some narrators, talking to a robot instead of a human alleviated anxiety, fear, and shame—they didn’t feel judged. At the same time, narrators lamented the lack of human connection.
Absence of human connection is not a minor issue when it comes to oral history best practices. Good oral history interviewers address positionality and work on creating rapport and building relationships with narrators to improve the story-sharing experience. In contrast, AI cannot recognize social context and develop relationships. Moreover, oral history interviewers provide comfort when narrators recall difficult experiences that bring forth intense emotions. Language models can do so to a certain extent, but it’s artificial and AI cannot convey empathy in the variety of ways human beings can beyond just words.
As Dr. Halperin also points out, automating oral history collection would be even less successful in places with low digital literacy rates. My project on reproductive care in Colombia is a good example of this. I have collected oral histories with elderly rural Colombian women who, in their youth, worked as rural health promoters—a kind of paramedic personnel whose work was indispensable for making birth control available in the Colombian countryside. Having been underpaid, underrecognized, and living in areas with little tech infrastructure, these former promoters often do not have access to a computer or a reliable internet connection. I could not ask them to interact with a language model to collect their valuable stories. Going to these women’ sometimes hard-to-reach houses and building a relationship with them over multiple cups of coffee was a must. I can imagine other researchers encountering similar difficulties, particularly given that arguably oral history is at its best when preserving the experiences and perspectives of marginalized social groups—who are more likely to have limited access to and trust in technology.
In oral history, interviewers are there to facilitate an affirmative and supportive storytelling experience; there is no need to challenge the incompleteness and biases of narrators’ memory. Yet, when interviewing someone, interviewers listen for inconsistencies or noticeable mistakes to protect narrators in the long run. Has the narrator mentioned they had three brothers and later talked about a fourth one? Interviewers need to ask a follow-up question about that. If not, the narrator may lose credibility in someone else’s eyes later, even though there is likely a meaningful explanation behind who they consider a sibling in some contexts and not others.
A human interviewer, carefully listening to narrators during interviews can also prevent negative legal and social consequences for narrators. Interviewers listen for information that may be used against narrators and remind them if the recording or transcript will be public. The absence of this filter is even more alarming in automated oral history collection, considering that people may be more prone to overshare when interacting with a robot due to the aforementioned lack of fear of being judged.
Moreover, unlike AI, human interviewers can listen for ways to give back to narrators. For instance, while collecting oral histories, I learned that former rural health promoters have had difficulty making their time in that supposedly voluntary position count towards the acquisition of social security benefits. Having access to free legal advice, I was able to create a folder with resources they could use, such as letter templates and a sheet with the contact information of state agents that could help them. At least for me, it is hard to imagine how to feed a language model enough information for it to identify a narrator’s need that can be address with the oral history project’s available resources.
In sum, I would personally argue against automating the collection of oral histories using AI. I recognize that it can be helpful for certain projects where narrators can comfortably interact with language models, and it is relevant to collect more interviews than would be possible without automation. Think, for instance, of a study on how teenage cultures around the world are increasingly interconnected and influenced by global trends. However, although AI may be able to help elevate crucial voices in certain scenarios, the benefits of traditional oral history collection far outweigh the advantages of automation in most cases. This does not mean that AI cannot be useful for other stages of an oral history project, as will be seen in the second part of this essay.