Coded Bias Reflection (HW 1)

#cs390#ethics#hw

Tips:

1. Clear Purpose
2. Good Structure
3. Analytical (draw conclusions)
4. Honest and Reflective
5. Clear and Concise
6. Free of error

Main Points:

1. Bias exists in AI/ML algorithms
2. Bias is harmful to users and can have real detriment
   1. Facial recognition
   2. Invading privacy
   3. Amazon Resume bias against women
   4. Tenure & Firing for teachers
   5. Credit score
3. AI exists everywhere and can be incorporated more and more behind the scenes

Goal: Discuss potential causes and implications of biases Causes:

1. Biased data
2. Biased testing
3. Lack of foresight in harm and implications
4. Quick/Hasty implementation (think amazon face rec)
5. No accountability (hidden systems like credit score, mortgage, etc.)

Implications:

1. Harm specific groups of individuals not represented in data or testing
2. Untested systems or systems without much thought against bias released, causing unintended side effects (think Microsoft Tay)
3. Invasion of privacy (Amazon face, ads, etc.)
4. Lack of regulation or ability to perpetuate bias without repercussion (due to no accountability)

I have been aware of a variety of ethical issues arising from bias in artificial intelligence and machine learning systems after watching the Coded Bias documentary. With my familiarity of AI systems and from examples provided by the documentary; I observed that the potential causes of bias include lack of representation in data and testing, little foresight into the harm and implication of systems, hasty implementation and commercialization, and less observable systems (credit score, teacher performance, etc.). I believe algorithm bias is a pertinent issue for our society given it’s ability to harm specific groups, cause unintended harm, invade individual privacy, and operate with little regulation.

One of the initial topics in my ’Data Mining’ was representation in data. With a skewed dataset, the outcome of various intelligent algorithms and models are bound to cause bias. A recurring example in the documentary was bias in the accuracy of facial detection which favored white males. This was due to the disproportionately high amount of them within the training data; an outcome that I believe to be obviously expected. The lack of representation of other skin tones and genders may be attributed to lack of foresight or the lack of sufficient data collection. A situation like this one may be avoided by proper testing of all groups the model will be used with, and then sufficient data collection to ascertain model accuracy.

A personal thought I have towards the ‘data collection’ suggestion, is that for ’low stakes’ computing (Face-ID and convenience face detection), collecting representative data may not be possible. Companies may not have sufficient resources to collect all representative data. In the case of societally unimportant algorithms, I think it is okay to have disproportionate accuracies.

However, for a more important face detection implementation — Amazon’s facial recognition, it is important to be accurate to all groups. This AI model was created with a lack of foresight and quickly released to investors / buyers, and as such, turned out to be biased. In this situation, given it’s commercial use and especially for it’s marketing towards government entities, it has real capability to harm individuals. Similarly, police face detection of potential criminals which lead to arrests or stops are also an important implementation where bias must be minimized. I deduce the implications of bias in these models harm both the commercial entities involved and the end users / targets.

Systems without direct accountability are a dangerous point mentioned in the documentary. I too, am worried about these systems, and I believe we must be very cautious when implementing them. Bias in these systems may propagate systematic harm to groups without their knowing. An example of this which I personally would be affected by is school admissions. For example, if an AI algorithm is to take in input of admissions essays, academic history, and to make conclusions of my performance, I would believe it is worse than having an erroneous human make the judgement. Humans in a sense, are also biased, but not to a systematic extent. An example of this which resonated with me in the documentary is the example of teacher’s performance measures by an AI algorithm. This algorithm marked a seemingly good teacher as underperforming, a decision I believe that no human would make under normal circumstances. Even if this outcome is an outlier, bias can present itself to only outlier groups. For systems not readily available to their users (admissions decisions, credit score, performance measures, etc.), they should be strenuously tested for bias, and used as sparsely as possible. Alongside this, data given to these algorithms should be limited — if one was to predict my performance; my race, gender, family, etc. should have nothing to do with it.

Overall, the documentary has taught me a lot about where bias can present itself within algorithms, and how it can harm groups of people. In my journey with AI, I will take extra care to make sure that I avoid the causes of bias, and consider the implications that my model will have towards society. Alongside this, I will also be more critical of existing and upcoming models which present themselves as the technology evolves.