Reference no: EM133147953

Dynamic Sound Identification ethics case study

Also known as "query-by-example," dynamic sound recognition recently found commercial success as a means to identify music through short audio snippets, captured through a microphone. First-generation algorithms recognized unique signatures in a particular sound, which they could then match with a most likely source or an equivalent sound stored in a large database of previously identified auditory signatures. Early mobile apps employing these algorithms were amusing and effectively enabled music listeners to identify a song's title and the performing artist. One Los Angeles-based research and Development Company determined that the underlying technologies might have further, public-minded implications as well, and began exploring new uses for sound recognition algorithms. The most promising output of this research was a mobile app, dubbed Epimetheus. Epimetheus was particularly proficient at recognizing music, advertisements and human voices. Unlike previous apps using dynamic sound identification, Epimetheus was also adept at picking up subtle auditory signals and sorting through environmental noise in order to accurately identify natural phenomena, such as the changing tides. This functionality was meant to benefit scientific researchers who could employ Epimetheus as a tool to track ecological change in remote locations. It also proved popular among students and casual hobbyists who enjoyed the app's educative and informative capabilities. In addition to identifying sounds with a high degree of accuracy, Epimetheus incorporated a machine learning algorithm that adapts to new inputs and provides users with useful information about the sounds being processed. For example, the app might identify personal information about those speaking, links to websites selling a product being advertised on television, encyclopedic entries about bird calls in the wild and other relevant resources. It wasn't long before the titans of Silicon Valley recognized Epimetheus' commercial and scientific potential and started bidding to acquire the underlying software. At that point, the research team behind Epimetheus began preparing demos that leveraged the strengths of its sound classification engine. For example, engineers developed an entertaining demo that was able to identify with high accuracy the voice actors/ actresses for cartoon characters. It even worked in cases where the cartoon characters were voiced by actors/actresses of the opposite sex (e.g. Bart Simpson is voiced by female voice actress Nancy Cartwright). One company, Cronus Corp., was especially impressed by these demos, and was eager to acquire Epimetheus and incorporate its sensing technology, databases and information provisions into its own products. However, before negotiations could proceed, Cronus Corp.'s lawyers asked the research team behind Epimetheus to prove that they had minimized the risk of unexpected harmful results. Programming an algorithm that is sensitive to societal norms and cultural flux is notoriously difficult, and Cronus Corp. did not want to unwittingly produce a bad outcome or acquire a public relations scandal. A problem arose when one of the adversarial testers, Sybel, tested her voice on the system. Sybel, who had been born as a biological male, had recently begun sexual reassignment and now identified both psychologically and publicly as a woman. Based on her voice sample, however, Epimetheus identified Sybel as male and displayed further information about her known history, including a link to several online videos that showed Sybel prior to her transition. The researchers only then realized the potential for this error to cause substantial dignitary and material harms for transgender individuals. When transgender users of Epimetheus are misidentified, they may feel like they are not being respected for who they are. For those who are "passing" as the gender with which they identify, being publicly identified by their biological sex might even make them targets for abuse. And while this one isolated error may have seemed minor now, researchers notes the potential for a larger, systematic problem. Transgender individuals comprise only a small percentage of the world population, but mass adoption of the Epimetheus app through an enormous technology company, like Cronus Corp., would mean that the algorithm might categorize individuals in ways that did not match their gender identity multiple times per day. As members of a historically marginalized group, this is would be no small thing. The research team revealed this issue to the acquisition team at Cronus Corp., apologized and promised Sybel a swift resolution of this rather embarrassing issue. However, time was running out for the Epimetheus team to devise a workable solution, and all the usual solutions proved inadequate. Regardless of the amount or type of new data the researchers fed into Epimetheus's training sets, the engineers could only marginally reduce the error rate of categorizing the sex of transgender persons. Even efforts to create focused and auxiliary training data using a significantly diverse set of transgender persons did not yield the necessary results in subsequent tests. The team had to concede that this may not be a problem that can be solved with more data or improved calculations but would require a different strategy entirely. The researchers at Epimetheus organized several workshops and focus groups with experts from a variety of fields. Participants signed non-disclosure agreements before being invited to critique the approaches and help think through possible solutions. Experts were also asked to help identify any additional red flags or areas for concern. These review sessions produced several findings. Regarding the Sybel problem, some reviewers suggested that the team might want to rethink whether the benefits of using Epimetheus' algorithm on any particular sound would always be worthwhile. Epimetheus' low error rates-calculated at around 0.016% of identified issues-were well within the acceptable range for each interaction. However, given the scale of operations at Cronus Corp., even this small rate of error would likely be amplified beyond what the researchers and the interested companies may consider to be negligible levels. In instances where such an error might harm members of already marginalized groups, several reviewers argued that the only acceptable rate of error should be zero.

Drawing on one or more moral reasoning theories examined in class, including Utilitatianism, Kantianism, Virtue Ethics, and a Human Rights' approach, discuss:

a) Do you agree with the view that that 'the only acceptable rate of error should be zero' in the case of the Epimetheus mobile application?

b) Should Cronus Corp. invest in the Epimetheus application despite the fact that Epimetheus engineers could only marginally reduce the error rate of categorizing the sex of transgender persons, from a moral and/or business point of view?

c) To what extent is/are the moral reasoning theory/theories employed helpful in addressing the the two questions above?