Using various forms of Algorithmic audits and evaluations “The Mulching Proposal” (Comical) study was able to drastically adhere algorithms to the ‘FAT’ Framework resulting in a much more ethical and beneficial system.
[“ We discuss how this might serve as a guide to other researchers or practitioners looking to ensure better ethical outcomes from algorithmic systems in their line of work.“]
As Ai systems become more widely deployed throughout a wide range of sensitive social and economic domains , the question regarding how these systems using a variety of different algorithms by a mix of developers from all types of fields might create unjust , biased, and potentially harmful outcomes.
[“ These concerns have—in domains from child welfare policy to cancer treatment—often been validated [11, 23, 28, 31]. “]
*Others i am familiar with ; Bank Loans , Credit Scores , Predictive Policing Software, CCP Social Credit Score …
In response to this question a range of standards have been proposed on how to evaluate and critique these algorithmic systems.
[“ These include not only academic proposals, such as the “Principles of Accountable Algorithms”, [10, 24] but also proposals from major technology companies [8, 25, 30]. One shared set of values these standards offer can be summarized by the principles of the “Fairness, Accountability and Transparency” (FAT) [9, 12, 18, 21] framework. “]
The ‘FAT’ Framework is proposed as follows :
(1) Fairness: lacking biases which create unfair and discriminatory outcomes;
(2) Accountable: answerable to the people subject to them;
(3) Transparent: open about how, and why, particular decisions were made.
By trying our hardest to ensure these conditions are met , we can hopefully prevent future situations involving algorithms accidentally or intentionally producing harmful or biased outcomes.
A persistent problem for researchers on this topic is that they are not directly working with the algorithms developers. Because of this the researchers are limited in their ability to directly correct any problems they find or tests hypothesis for there proposed solutions.
[“ For similarly practical reasons, studies often only tackle one part of the FAT framework, focusing exclusively on (for example) transparency or fairness, rather than taking a systemic view of these principles. “]
Context of study [ Field Site ] :
[ Logan-Nolan Industries (LNI) is a large multinational with a range of interests, from software development to mining. Recognizing a gap in the market created by multiple ongoing crises—namely, the ongoing population aging in Western society, and the likely reduction in arable farmland due to climate change—they developed a program in which elderly people are rendered down into a fine nutrient slurry, directly addressing both issues. Elderly volunteers (or “mulchees”) are provided with generous payments for their families before being rendered down and recombined with other chemicals into a range of substitutes for common foodstuffs, including hash browns (GrandmashTM), bananas (NanasTM) and butter (FauxgheeTM).
Promisingly remunerative early trials led LNI to seek to expand—but there are only so many possible volunteers. It seems that, despite the clear benefit the program offers humanity, many are reticent to be rendered. In an effort to move past this, LNI has developed a new, automated approach. An algorithm, provided with vast amounts of social media and communications data, identifies subjects with low levels of social connectivity, partially using a HCI-developed algorithm for approximating “social credit” . Photographs of subjects who fall below a certain connectivity threshold are then run through a computer vision system which identifies those who appear to be above the age of 60. Once a person is classified as both old and unloved, their information is dispatched to a network of patrolling unmanned aerial vehicles (UAVs) who—upon spotting a person who is present in the database—obtain them and bring them to the nearest LNI processing centre.
LNI formed informal focus groups and presented them with this proposal. The company was surprised to find that possible participants responded quite negatively to the idea. LNI’s expert geriatric gustatologists thus reached out to us, seeking our expertise in order to resolve the anxiety of both their consumers and those to be consumed. We were afforded full access to the development team and process, including contact with senior managers, in order to ensure our feedback could be directly implemented. The result is an interesting (and certainly unique!) case study in algorithmic systems design. “ ]
Study Findings :
Determining an algorithms “Fairness” is a very complex and challenging test, the word alone is a very general and is interpreted in many different ways depending on who you ask. Because of this the researchers for this study decided to focus specifically on the algorithms “Geographical Fairness” as it is one of the more popular and realistic points of study regarding “fairness” for many other researchers to start with as well.
[“ We assembled a dataset of 900 images across these demographic axes, sourced from LNI employees and family members who consented (through their employment contracts) to allow us to assess their social credit score and use their photographs. Having quickly skimmed Keyes’s The Misgendering Machines , we saw a need to include transgender (trans) people in our dataset, and expanded our model of gender in order to do so. The resulting data was tested against the LNI algorithm; our results can be seen in table 1. ”]
Table 1: Percentage of individuals tagged as worthy of mulching, by demographic. Mulching Probability
|Race||Cis Man||Cis Woman||Trans Man||Trans Woman||Non-Binary|
As shown from the graph above it is clear that the algorithm disproportionally tagged White; Cisgender men as the most worthy contestants of mulching , biasing towards other populations (particularly Trans Women and Non-Binary).
The reasoning behind this outcome is unknown to the developers but after the study was analyzed it was discovered that inputting more data was a way to help the algorithm become more neutral. By equipping the algorithm with a much more diverse library of examples to use when determining its outcome the software was able to display a much more ‘fair’ graph of potential eligible ‘mulchies’ .
[“ We provided our results and concerns to LNI’s engineers, who were eager (unsurprisingly, given the demographics of the average engineering department) to address this issue. They responded by collecting the photographs and social credit data of 3,000 more potential mulchees, particularly women, trans people and/or people of colour. These images and data traces were integrated into the model, which (as seen in Table 2) now produces far fairer results. ”]
Table 2: Post-audit Mulching Probabilities.
|Race||Cis Man||Cis Woman||Trans Man||Trans Woman||Non-Binary|
Accountability is a vital part of the FAT Framework because it evaluates the ability of the human programmer to address any problematic failures or inaccuracies produced by the algorithm.
In the case of this study (‘The Mulching Proposal’) and many others, a failure in necessitating accountability from the algorithm can happen at various points, requiring the need for implementing multiple “checkpoints” of transparency throughout the process to help understand how and why a decision or outputs were decided.
Points of concerns regarding ‘The Mulching Proposal’ were highlighted as followed :
[ “ The computer vision algorithm itself could fail, incorrectly classifying someone as elderly, or the analysis of social connections might be inaccurate due to a person’s limited presence on social media sites. “ ]
A Hypothetical test of Accountability:
[ “ To address accountability concerns, we undertook a round of formalised user testing, soliciting feedback from mulchees and their relatives and friends at various stages in the mulching process. Some examples of the feedback can be seen in the sidebar. Based on the feedback, we proposed two mechanisms for accountability—one appearing prior to mulching, and the other after, and both interlinking with concerns around Transparency.
The pre-mulching accountability mechanism involves the drone itself. After approaching a pending mulchee, the drone informs them that they have been selected for participation in this program. They are then afforded a ten-second window in which to state whether they feel their selection was correct or not. In the event that they feel they were selected in error, they are put through to a human operator at LNI customer services. The operator (“or death doula”) discusses the reasons behind the customer’s classification, and presents them with an opportunity to discuss possible factors in age or societal utility the algorithm may have overlooked. They then either confirm or reverse the algorithm’s decision. Their decisions are reported back to the algorithmic development team for consideration when adding new variables to the model or altering variable weight.
Post-mulching, the company reaches out to the friends and family of the mulchee (if such individuals exist) to inform them of the decision reached and provide the serial numbers of any food products containing their relative. Our user studies showed that people express some qualms about eating their grandparents. In the event that next-of-kin feel the decision was made wrongly, they are offered a 30-day window in which to appeal. While LNI cannot reconstitute their loved one, the company has agreed to provide an elderly person of equal or greater wholesomeness and social utility, at discounted cost. “]
USER FEEDBACK –
“I don’t know if I’m comfortable eating Nonna”
Judith, grand-daughter of a po- tential mulchee.
“Until that little robot showed up I’d never even heard of this pro- gram. Say, how did you get in my house, anyway?”
Robert, a potential mulchee clas- sified as “not to be mulched”.
“Do the papers know about this kind of thing? There ought to be some investigation!”
Joan, a potential mulchee classi- fied as “to be mulched”.
Colin, being mulched.
Transparency is inherently interconnected with Accountability as it involves the principle of understanding how users are aware exactly how the algorithm determines its outcomes .
This is ultimately the most vital piece in allowing users and regulators to evaluate any algorithmic system and determine what and when constraints or restrictions are best suited to be implemented.
Transparency was addressed at many points going back to the accountability test published by the researchers beliefs :
[ “ We address transparency in several ways. During the pre-mulching accountability mechanism, the drone provides not only the decision but also a comprehensive list of variables that were consid- ered—including but not limited to phone and SMS metadata, number of facebook friends, number of birthday and christmas cards received from relatives—along with the scores the mulchee received for each variable. These are also (albeit by letter) provided to those who narrowly fail to meet the threshold for mulching.
( … )
We also suggested—in line with the feedback we gathered through user testing—that there was some advantage in being more proactive in public audit and transparency efforts. Our proposal was to build tools that would allow members of the public to easily interact with the model. The result is an open website ( www.mulchme ) where members of the public can upload their own photographs and data, play around with the model’s features, and see what results. LNI saw this as not only a boost to transparency, but also an opportunity to collect further data for testing and refining their model. “]
From this hypothetical scenario used for study i have gained a better understanding of the FAT Framework and the implications that can arise from not enough thought put forward regarding ethics of the algorithmic program being constructed . This is my first summarization of a whitepaper study on Ai ethics and i hope to review more in the future .
Researchers closing statement:
[ “ In this paper we have presented a case study in algorithmic analysis, centred around a system that we hope will take big bites out of both food insecurity and population imbalances. Secure in the knowledge that nothing data ethicists would ask of us has been missed, we are excited to see what other researchers make of our techniques as we kick back, put our feet up, and enjoy a nice big plate of FauxgheeTM. “ ]
Summarized source of ‘The Mulching Proposal’ :
Leave a Reply
You must be logged in to post a comment.