Which three of them seem most fundamental to achieving ethical data use for individuals?
Respond to the following:
1. Of the four principles of data ethics listed in Box 10.4 in Chapter 10 (
fairness, accountability, transparency, and explain ability),
which three of them seem most fundamental to achieving ethical data use for individuals?
2. What’s the difference between lawfully obtaining personal data and having accountability for obtaining it?
3. Provide specific examples of behaviors that you could follow that apply each of the four principles as they relate to data that you would need to collect for the development of the project scenario app.
BOX 10.4
Data Ethics Principles (FATE)
· Fairness refers to impartial and just treatment or behavior without favoritism or discrimination. While this is something to which organizations adhere in areas such as promotion and hiring, in the context of data analysis, it refers to how fair a dataset is and what the impact will be from using the results. For example, sometimes, the dataset is biased toward a particular demographic that results in unfair decisions being made. If these could be identified and revealed by the system, it would make it possible to rectify them while also developing new algorithms that can make the system fairer.
· Accountability refers to whether an intelligent or automated system that uses AI algorithms can explain its decisions in ways that enable people to believe they are accurate and correct. This involves making clear how decisions are made from the datasets that are used. A question that arises is who is accountable for doing this? Is it the person providing the data, the company coding the algorithms, or the organization that is deploying the algorithms for its own purposes?
· Transparency refers to the extent to which a system makes its decisions visible and how they were derived (see Maurya, 2018). There has been much debate about whether AI systems, which typically depend on large datasets when making a decision, should be designed to be more transparent (see Brkan, 2017). Examples include medical decision-making systems that can diagnose types of cancer and media service providers (for instance, Netflix) that suggest new content for you to watch based on their machine learning algorithms. Currently, many are black-box in nature; that is, they come up with solutions and decisions without any explanation as to how they were derived. Many people think this practice is unacceptable, especially as AI systems are given more responsibility to act on behalf of society, for example, deciding who goes to prison, who gets a loan, who gets the latest medical treatment, and so on. Some of the rules of the GDPR on automated decision-making are also concerned with how to ensure the transparency of decisions made by machine learning algorithms (Brkan, 2017).
· Explain ability refers to a growing expectation in HCI and AI that systems, especially those that collect data and make decisions about people, provide explanations that laypeople can understand. Research into what is a good explanation to provide has been the subject of much research since expert systems came into being in the 1980s. Following this early work, there was research into what context-aware systems should provide. For example, Brian Lim et al. (2009) conducted a study that provided different kinds of explanations for a system that made automated decisions. They found that explanations describing why a system behaved in a certain way resulted in a better understanding and stronger feelings of trust. In contrast, explanations describing why the system did not behave a certain way resulted in lower understanding. More recently, research has investigated the kinds of explanations that are appropriate and helpful for users of automated systems (see Binnes et al., 2018).
The FATE framework suggests that the design of future systems, which use AI algorithms in combination with personal or societal data, should ensure that they are fair, transparent, accountable, and explainable. Achieving this goal is complex, and it involves being aware of the potential for bias, discrimination in big data and algorithms, ethics in big data, legal and policy implications, data privacy, and transparency (Abdul et al., 2018).
Achieving this objective is inevitably difficult. For example, as pointed out by Cynthia Dwork at a panel on big data and transparency (transcribed by Maurya, 2018), it is difficult to know what a good explanation of a decision might be for human beings. She uses the example of what should a system say when a user asks, “Why was I turned down for the loan?” to illustrate this. The system might be able to reply, “There is a classifier, we feed your data into it, and the outcome was that you were turned down.” However, that is of little help to a user, and it is likely to be more annoying than not having any explanation.
Reuben Binnes et al. (2018) conducted an experiment to determine what kinds of explanations users found to be fair, accountable, and transparent for an automated system. In particular, they compared four different styles of explanation, ranging from being largely numerical scores to more comprehensive ones that provided a breakdown of the statistics used for certain demographic categories, including age, gender, income level, or occupation. The different styles were presented for scenarios in which a decision had been made about individuals automatically, such as applying for a personal financial loan and where passengers on over-booked airline flights were selected for rerouting. The results of their experiment showed that some of the participants found that they engaged with the explanations to assess the fairness of the decisions being made, but at times they found them impersonal, and even dehumanizing. What constitutes a fair explanation may need to be more than providing an account of the processes used by the algorithms. From an interaction design perspective, it might help if the explanations were interactive, enabling the user to interrogate and negotiate with the system, especially if a decision that has been made is contrary to what they expected or had hoped.
Jure Leskowec (2018) comments on how the consequences of a system making a decision on behalf of a human can vary. This will determine whether an explanation is needed to support a decision made by a system and what it should include. For example, if a decision is made to pop up an ad for slippers in a user’s browser, based on an analysis of their tracked online app usage (a common practice used in targeted advertising), it might be mildly annoying, but it is unlikely to upset them. However, if it means a person is going to be denied a loan or a visa based on the outcome of an automated algorithm, it may have more dire consequences for someone’s life, and they would want to know why the particular decision was made. Jure suggests that humans and algorithms need to work together for system decisions that implicate more important societal concerns.
Another reason why ethics and data have become a big concern is that automated systems that rely on existing datasets can sometimes make decisions that are erroneous or biased toward a particular set of criteria. In doing so, they end up being unfair. An example that caused a public outcry was the misidentification of people with dark skin. Traditional AI systems have been found to have much higher error rates for this demographic. In particular, 35 percent of darker-skinned females were misidentified compared with 7 percent of lighter-skinned females (Buolamwini and Gebru, 2018). This difference was exacerbated by the error rate found for lighter-skinned males, which was less than one percent. One of the main reasons for this large discrepancy in misidentification is thought to be due to the make-up of the images in the datasets used. One widely used collection of images was estimated to have more than 80 percent white images of which most were male.
This bias is clearly unacceptable. The challenge facing companies that want to use or provide this data corpora is to develop fairer, more transparent, and accountable facial analysis algorithms that can classify people more accurately regardless of demographics, such as skin color or gender. A number of AI researchers have begun addressing this problem. Some have started developing 3D facial algorithms that continuously learn multiracial characteristics from 2D pictures. Others have introduced new face datasets that are more balanced (see Buolamwini and Gebru, 2018).
Order an Essay Now & Get These Features For Free:
Turnitin Report
Formatting
Title Page
Citation
Outline
