Part of a recurring series that highlights some piece of particularly pertinent news regarding AI
Above: an AI-imagined image of Johansson from the film Her
The news about Scarlett Johansson’s dispute with OpenAI over the use of her voice garnered much attention in the past week, and rightfully so. For a brief refresher, OpenAI debuted a voice assistant named “Sky” that for many recalled Johansson’s voice (and, for good measure, her character in the 2015 film Her in which she plays…an AI voice assistant). Johansson herself has demanded answers from the company on how it was built, though OpenAI has denied using Johansson’s voice itself as the training model.
Nonetheless, the episode has raised questions about the ethics of creating synthetic media. Should there be restrictions on the uses of AI technology to mimic others’ likenesses — whether for profit, for malicious purposes, or even simply for fun? Should technologists be either compelled or at least encouraged to make synthetic media readily identifiable (such as via watermark technology)? Should they be forced to disclose the data on which such products have been trained, and could a system be developed to compensate the ostensible owners of such data? Finally, how will regulations of these activities be written in such a way that they are not vague and overbroad enough to accidentally stifle less controversial applications of the technology?
While the Johansson case might have raised these questions in a highest-profile context to date, they have actually been in the sights of academic researchers who consider the ethical and socio-cultural implications of technology for a while.
One recent research study by a team of researchers at Google examines the use of voice data in a context far from the red carpet, in fact: the collection of voice data from medical patients to train generative AI models that could be used by clinicians. The researchers note that such applications are currently in development, as “Electronic Health Records (EHR) system vendors and clinical AI labs have recently developed natural language models to automate medical chart reviews by evaluating EHR data and identifying opportunities for improvement and validation of documentation for patient encounters.” Ultimately, this could help to “alleviate the burden of data entry by clinicians” — an increasingly burdensome task that some say detracts from the care they provide.
The study interviewed doctors using an approach that invited them to consider the various problems that could arise with the collection of voice data to train these models and its implementation in the clinical setting. The interviews suggested a number of potential issues that could undermine trust between patients and physicians, cause patient self-censorship, and lead to inaccurate or biased assessments of patients. As the researchers note, such reminders are critical, as “not paying enough attention to the ethics of data collection and data quality can have outsize impact on vulnerable communities.”
This kind of work on emerging AI technologies involves a number of different traditions in communication research. First, the medical study in particular underscores how the communication that occurs in health contexts can have a profound impact on the people involved and on the medical outcomes themselves. At SUNY Plattsburgh, the course CMM 303: Health Communication takes an in-depth look at this kind of research over the course of a semester. It’s also important to remember that the methodology of this study utilized interviews, which is an important method of data gathering that is useful across many different professional contexts. Interview skills are cultivated in different parts of the Comm Studies curriculum, such as classes in the Broadcast Journalism concentration as well as CMM 304: Business & Professional Communication.
######
An organization that I find indispensable in tracking and unpacking developments in AI is called the Montreal AI Ethics Institute. I highly recommend signing up for their newsletter, where I learned about the study covered in this post.
Benjamin A. Medeiros 