Image is courtesy of PhonlamaiPhoto via Getty Images.
What is Machine Learning in Artificial Intelligence?
Machine learning is a branch of Artificial Intelligence (AI) that uses algorithms to train a machine to learn and improve based on sample data. Without the need to explicitly program every step, the machine is able to autonomously classify or make predictions, simulating human intelligence. Currently, AI is used to assist humans in different aspects, such as sorting primary and spam mail and predicting weather patterns. As a growing industry, AI holds the large potential to solve pressing world issues by mimicking human intelligence at a global scale, tremendously increasing efficiency and accessibility. In order to protect and empower the most vulnerable groups, it is essential that programmers consider the tradeoff between fairness and accuracy. With a loose definition of “fairness” in AI and the complications of increasing equity or accuracy, it is difficult for programmers to minimize the potential biases within their model.
A Brief Process of Machine Learning
Before training the data on the model, the machine has to gather and prepare the data. Although this step may often be overlooked, it is essential that the quality and quantity of the training data are maintained to maximize accuracy. The model then goes through a training loop, learning, and backpropagating from each iteration. Next, the model is evaluated and tested to measure the performance and accuracy. Lastly, the model is deployed into an external context and is further monitored for changes.
While a high accuracy score is optimal, programmers are careful not to overfit the data, which is the case when the accuracy is too high, and the model fits too closely with the training data. Thus the model is unable to learn new information from the training data and loses predictive power.
Data Bias
A majority of reported biases within AI models are caused by biases and inconsistencies within the training data. When learning from an incomplete or preferential dataset, the machine also draws biased or inaccurate conclusions based on the data given. Examples of data bias include:
Racial bias in facial recognition systems
Facial recognition systems employ AI algorithms by training machines on facial feature datasets to autonomously recognize patterns and classify whether the face is the owner of the device. Facial recognition claims to be 90% accurate, but studies show that there are diverse error rates between different demographic groups. The Gender Shades project evaluated face recognition tech companies such as Microsoft, Face++, IBM, Amazon, and Kairos and found that individuals who were black, female, and between 18-30 years old consistently received the lowest accuracy compared with other demographics.
The evaluation of face recognition technology shows accurate discrepancies between different demographic groups (Gender Shades project).
Bias in Amazon employment system
In 2014, the Amazon company started a project to train a model to autonomously review job application resumes. According to Reuters, the AI system could give applicants scores ranging from 1-5 in order to automate the repetitive process; however, in 2015, the company realized that their new AI recruiting system was not fairly rating the job candidates. The system showed shocking results that were unfair against female applicants, which were due to the use of biased data. The company used ten years of historical data to train its model, but there was evident male dominance in the tech industry in the past, with 60% of Amazon employees identifying as male. The machine incorrectly learned to consider gender as a factor, penalizing resumes that included keywords such as “women” within applicant resumes.
Bias in the criminal justice system
Correction Offender Management Profiling for Alternative Sanctions (COMPAS) is a machine learning algorithm used by the U.S. to predict the risk of recidivism - the likelihood of convicted criminals reoffending. Trained on the data of 34,000 federal offenders, COMPAS considered 137 factors such as age, race, and criminal history to autonomously rate the risk of recidivism ranging from 1-10, with 1 being the lowest risk and 10 being the highest risk. Yet, after an evaluation of over 10,000 criminal defendants and comparing COMPAS’ rating to their actual recidivism rates, it was clear that there was racial bias between white and black defendants. Black defendants were predicted to have a higher risk of recidivism and were twice as likely than white defendants to be misclassified by the algorithm. COMPAS is used as additional information for judges to consider when declaring the verdict; therefore, biases within the data may cause the judge to deliver inaccurate justice.
What are the impacts of biases in AI?
If left unaddressed, biases in AI will continue to represent human prejudices and reinforce dangerous stereotypes within the general public.
There is no doubt that AI has the potential to increase efficiency in many sectors ranging from healthcare to criminal justice. Even so, biases in AI not only prevent machines from making accurate decisions, but issues of accountability lead many to question to ethics of employing AI. In fact, MIT Technology Review claims that programmers today are still working to find a way for AI to explain its reasoning of why it came to a certain conclusion. Without truly creating a machine to understand concepts, it is difficult to determine who is accountable, especially since the results of an AI model is representative of its respective creators.
Article Author: Rachel Weng
Article Editors: Sherilyn Wen, Victoria Huang