The development of artfcial intelligence (AI) systems and their deployment in
society gives rise to ethical dilemmas and hard questons. This is one of a set
of fctonal case studies that are designed to elucidate and prompt discussion
about issues in the intersecton of AI and Ethics. As educatonal materials,
the case studies were developed out of an interdisciplinary workshop series
at Princeton University that began in 2017-18. They are the product of a
research collaboraton between the University Center for Human Values
(UCHV) and the Center for Informaton Technology Policy (CITP) at Princeton.
For more informaton, see htp://www.aiethics.princeton.edu
2 | AI Ethics Case – Dynamic Sound Identfcaton
A lso 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
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.
Discussion Question #1:
Regardless of how much testng is done in the development stage, it is impossible to predict all
the potental harms that may occur when an AI system goes live. This is why Cronus Corp. only
asked Epimetheus to show they have done as much as reasonably possible to “minimize” the risk
of unantcipated harms. But while developers cannot predict everything, they should be able to
antcipate common discriminatory harms. What are examples of such harms? What might companies
do to minimize these risks?
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
This work is licensed under a Creatve Commons Atributon 4.0 Internatonal License. AI Ethics Case – Dynamic Sound Identfcaton | 3
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.
Discussion Question #2:
Even when AI companies aim to be inclusive, limitatons on data and human imaginaton ofen mean
that minority populatons lose out. As an engineer at Epimetheus, what might you do to make your
products accessible to all? What can you do?
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, lest the negotiators begin looking elsewhere for advanced
dynamic sound recognition technologies. Unfortunately, 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
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 a tiny 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.
Discussion Question #3:
Just because a technological capacity exists, does that mean it should be pursued? What factors
should companies take into account when determining whether or not to launch a product?
Discussion Question #4:
Epimetheus began as a small company without much market penetraton. Being acquired by the
much larger Cronus Corp. presents many opportunites, but also raised certain challenges. For one
thing, low margins of error that may have seemed fne initally may no longer be acceptable when
Epimetheus is scaled up and out. In that case, what would consttute an acceptable margin of error?
4 | AI Ethics Case – Dynamic Sound Identfcaton
While they were at it, the reviewers also alerted the Epimetheus team to several ethical dilemmas they
thought the company ought to consider prior to any sort of major expansion or buy-out. While not an
exhaustive list, they identified three major areas for concern. These centered around questions of cultural
insensitivities, concerns about the act of categorization, itself, and the lack of control over Epimetheus’s uses
once it had been made publicly available.
Ethical Objection #1: Cultural Insensitivity
Some experts pointed out that Epimetheus’ identfcaton of sounds and subsequent labeling had been
well trained on American cultural norms and the subtletes of the English language. However, this
training would not necessarily translate to non-Anglophone societes and different cultures. While
privacy laws had been mapped around the world and were taken into account when processing the
sounds, the researchers had not foreseen every culturally specifc ethical issue at play. Even within
the United States, experts warned that the sensitvity of certain labels and categories are disputed.
For example, while some may consider the term “American Indian” to be a valid descripton of one
whose ancestors lived in North America before European setlement, others may consider it offensive.
It would be a struggle—perhaps impossible—to develop a categorizaton schema that did not offend
Ethical Objection #2: Categorization as Harm
Beyond causing offense when categorizaton goes wrong, the mere act of sortng people and ideas into
groups struck some reviewers as wrong. To categorize a person, idea or thing is to assert a judgment
about what they are. Each instance of labeling may be minor, but taken together, programming
decisions about how to draw distnctons can influence society’s values and cultural understandings of
what is good, right and feasible. Where these categorizaton schemas are used to inform decisions or
actons within a larger, more complex informaton system, the results can be real material harm.
Ethical Objection #3: Unforeseen Uses
Epimetheus’ engineers had designed the app to do good. Even if it also produced some harmful side
effects, the intenton was to create something that would beneft society by increasing knowledge. But
what about bad actors who might want to use the technology for more nefarious purposes? The expert
panel argued that Epimetheus would need to think more about its moral and legal responsibilites in
the event that the sound recogniton capabilites were coopted to knowingly inflict harm.
The research team at Epimetheus was glad for the input and advice, but they struggled with how to
implement it. They argued that it was technologically impossible to reduce error rates to zero in the case
of misidentification of transgender voices—they’d tried!—and so they proposed a quick, though inelegant,
compromise. They decided that the least harmful approach would be to delete certain labeling categories
that had yielded insulting results for marginalized groups and would continue to do so as a service whenever
Epimetheus’ technology is included in new applications. This meant that, for example, Epimetheus would no
longer differentiate between genders when identifying voices. Such an ad-hoc approach would essentially
function as a band-aid solution, though one that might, in fact, do the trick.
Discussion Question #5:
Given that many outcomes cannot be predicted prior to a product’s release, how should companies
address individual or group harm afer it has occurred? Are ad-hoc solutons ever acceptable? What are
some alternatves?
This work is licensed under a Creatve Commons Atributon 4.0 Internatonal License. AI Ethics Case – Dynamic Sound Identfcaton | 5
Some observers were disappointed with this approach. They argued that deleting that one category
didn’t address the harm of labeling people in the first place. Furthermore, the act was one of erasure for
a community that has fought hard to make themselves seen and heard. They would have preferred for
Epimetheus to have committed more efforts into eliminating error rates for transgender voices.
The Epimetheus team defended their ad-hoc approach, while acknowledging that it is far from ideal. In the
extreme unlikely situation where an error is also insulting, the engineers decided that it is best to remove
the problem, rather than continuing funneling resources into efforts to marginally decrease the likelihood of
it occurring further. In the current nascent stages of the development of machine learning approaches, they
argued that it is not worth discarding the technology due to growing pains. Rather, ad-hoc solutions should be
embraced to allow the technology to mature further.
Indeed, the team went one step further, arguing that the development of new technologies is always going to
require a learning curve. Technology companies need room to experiment, and it is impossible to predict with
perfect accuracy the challenges a new product will present once it is unleashed in the wild. This is especially
true in the case of products that are likely to be used millions or billions of times a day, in which the question
may not be whether one can avoid inaccuracies as such inaccuracies are inevitable, but how best to deal with
them when problems arise.
In an attempt to address emerging concerns and dilemmas, the Epimetheus team committed to organizing
further stakeholder meetings, conducting interdisciplinary research and ensuring diversity of races, genders,
ages and socioeconomic backgrounds in development teams. They hoped that these measures would
contribute to the ongoing development of products and services that would not only push development
forward but would do so in a way to best serve social welfare.
6 | AI Ethics Case – Dynamic Sound Identfcaton
Reflection & Discussion Questions
Rights: The complexity of machine learning technologies makes it difficult to mitigate error rates entirely.
Furthermore, it is nearly impossible to judge the true impact of error rates during the testing phase. What
appears to be a negligible error rate in advance of a product launch may turn out to have significant
consequences once the technology is up-and-running and used millions or billions of times per day by people
across the globe. In many cases, the overall utility of these AI systems may outweigh the harms associated
with the large-scale effects of low error rates, which can range from the benign to the highly detrimental.
However, especially in cases where the consequences of a materialized errors constitute significant dignitary
and/or material harms, the rights of those on the losing end of a technology may need to be balanced against
the overall utility of the system.
○ If technology is imperfect and some error is inevitable, how should companies, product managers,
engineers, etc. balance the overall utility of a product with its potential harms to individual and
group rights? What are the factors that should be considered in making that determination? Are
these decisions that should be made in boardrooms and labs, or should they involve societal
○ Due to technological limitations, the Epimetheus research team chose to accept low error rates in
certain instances and not in others. In particularly sensitive cases, researchers decided it would be
best to remove a search category altogether rather than accept some inevitable degree of error,
which may cause harm. Which other solutions might have been available during the testing and
negotiation phases? Does a “better” solution exist?
Representational harms: Technologies that assess and sort the messy physical and social worlds
into predefined or emerging categories contribute towards human understanding and enable us to explore
connections that might otherwise have remained invisible given the limits of human intelligence. However, the
act of categorization may also inflict harm on the social standing, peer perception and/or self-understanding
of some people. To name a person and place her in a particular box is to detract from her individuality
and undermine her complexity. When that label contains negative social connotations or harmful political
associations, or when it is one with which the individual does not identify, the experienced harm can be
especially grievous.
○ What responsibilities, both to its users and society in general, does Epimetheus have to protect
against labeling that is not only improper but harmful? What role, if any, might diversity in
development have to play in reducing such instances?
○ Epimetheus is based in the United States, embedded with some American set of values. Virtually
any technical decision it makes could be considered a form of moral imperialism, imposing their
understanding of the problem onto a population through their technology. How, if at all, can
Epimetheus engage meaningfully with this accusation? If you were a manager at Epimetheus, how
would you balance conflicting definitional claims and standards between communities?
This work is licensed under a Creatve Commons Atributon 4.0 Internatonal License. AI Ethics Case – Dynamic Sound Identfcaton | 7
Neutrality: The process of drawing distinctions is never entirely neutral. By choosing certain categories
rather than others, and by defining those categories in particular ways, Epimetheus is implicitly making
value judgments about what is good, right and possible. For example, to categorize one sound as “music”
and another as “noise” indicates something about what Epimetheus believes both those ideal types
represent. These kinds of value judgments may then go on to influence the values of those humans who use
Epimetheus, creating a self-reinforcing pattern.
○ How should companies like Epimetheus decide which values to promote through its use (or nonuse) of particular categorizations?
○ Given that a technology can never be perfectly value-neutral, what if Epimetheus were to decide
to take a proactive approach towards promoting social values by, for example, designing its
categorization function to over-represent the prevalence of female scientists? Is it right or even
desirable for private companies to engage in social engineering? What might a “free speech”
absolutist have to say about such practices?
Downstream Responsibility: Technologies can be used in a variety of ways, or they may influence
others to create similar technologies for other ends. Once a system has left the hands of the original
engineers, they may not have much say in how their technologies are used. Sometimes, this means systems
that were designed to produce positive social ends get coopted to negative purposes, such as facial
recognition software being used by authoritarian regimes to identify and persecute political dissidents.
○ Could the researchers at Epimetheus be held responsible when lives are endangered on the basis
of erroneous identifications or labeling? Should they be responsible? Think of this in terms of not
only legal liability, a well-defined jurisdictional term, but moral culpability as well.
○ Can we expect engineers to foresee how the data they create through machine learning inferences
may be used in further systems that make decisions about people? Up to which hypothetical data
reuse moment should an engineer think ahead? What if the inferred data would be directly useful
for authoritarian governments who could justify crackdown on minorities or special interest groups
based on erroneously inferred or collected data?
AI Ethics Themes:
Representational harms
Downstream responsibility
This work is licensed under a Creatve Commons Atributon 4.0 Internatonal License.

YOU MAY ALSO READ ...  Qss – (1) No, because everything changes in the world. There are environmental changes, mental health changes, and people get sick, die, etc. There are always new studies about the same subject that we ca

Assignment status: Already Solved By Our Experts

(USA, AUS, UK & CA Ph. D. Writers)


Order from Australian Expert Writers
Best Australian Academic Writers