What is Bias in AI and How to Avoid It?
When we are weighing things, events, or people using different ways for various goals, the algorithms cannot be neutral. Thus, to develop solutions for the creation of impartial systems of artificial intelligence, we need to understand these biased algorithms. The goal of this article is to reveal the AI Bias sense, its types, bias in ai examples, and how to mitigate risks associated with them.
First, let us define what AI Bias is.
What is Bias Algorithms and Why They are Important?
Bias Algorithms are the types of algorithms describing computer system repeating and systematic errors, which lead to unfair results, such as the preference of one random user group over other groups.
Two types of Bias in AI exist. One is the AI algorithm Bias is trained with a Biased system of data. Another type of AI Biases is bias AI in society. Here our social norms and assumptions make us have blanks or some definite expectations in our minds.
For instance, a fair algorithm of a credit ranking can refuse you in giving a loan, if it constantly weighs appropriate financial indicators.
Why bias algorithms are so significant?
The explanation is simple – people write algorithms, select data, that these algorithms use, and decide about the application of these algorithms’ outcomes. People may accept such subtle and unconscious AI biases without various commands and careful thorough training, which can lead IA to automatize and immortalize them.
Application Bias in Machine Learning
Machine learning bias sometimes the name Bias in AI is a kind of event when algorithms create outcomes that always have a form of biases systematically as machine learning has wrong assumptions.
There are follow wing common Bias AI known:
Algorithmic types of biases
This event takes place when an algorithm has such a problem with computations with support calculations for machine learning.
Bias of samples
It occurs if a certain problem with data intended for training a model for machine learning appears. The data of this kind of machine learning bias are not too much big or suitable enough for teaching the system. For instance, when we use the data for teaching which foreseen only women teachers making tuition of the system, the conclusion arises that all tutors have only female gender.
Preconceived artificial intelligence bias
Here, we use records for tuition in the system accounting for actual preconceptions, stereotypes, or wrong social assumptions that can introduce these true biases in computing learning. For instance, we use the medical specialists’ data that include only women nurses and men doctors, thus, creating a timeless stereotype of medical employees in machine systems.
Measurement AI Bias
As its title says, this bias in AI is caused by the fact that data are not enough precise and the measurement and evaluation of data. If a system intended for an assessment of the workplace area is touted with the help of the photos of happy employees, it can be a biased system, if these employees already knew that the purpose of their training was the achievement of luck. When the system is trained to evaluate the share, it will have a bias type, if the shares in the data for such tuition were successively surrounded.
Bias of exception
It takes place when an important data period stays beyond the data which are applied, which means, something occurs, when the developers refuse to acknowledge the data period as indirect.
The Most Common Bias in AI Examples
Bias in AI is a belief, which is not based on famous facts about a person or a certain group of persons. Thus, there is a well-known belief that females are weak, however many women worldwide are known for their strength. Another one belief – all black people are not honest, but in fact, most of them are honest.
The meaning bias algorithms describe repeatable systematic mistakes, which lead to unfair results. For instance, loan ranking algorithms can refuse to issue a credit, even it is fair if is constantly weighing appropriate financial indicators. If this algorithm provides credits for one customer group but refuses to give them to another group of customers, which are almost the same, based on unlinked criteria, and this kind of behavior repeats several times, we can call it AI algorithm bias in this case. It can be intended or not intended bias, it, for instance, can come from the biased records received from an employee, who performed a job, which will be made by an algorithm from this moment).
Let us consider an example of an algorithm for recognizing faces, which can more easy thought to detect a white person, than a person with black skin, because this type of data is more often used in tuition. The minors can suffer from it as equal opportunities are not possible in discrimination and oppressing can be endless. These biases are not intended and can be hardly revealed until they are programmed with appropriate soft, and this is the problem.
Here are some common Bias in AI examples we can face in real life:
Racism in the medical system of the USA
Technology must facilitate the reduction of health inequality, but not make it worth it when the population fights with continuous preconceptions. Artificial intelligence systems learned on the basis of health data, which is not representative, usually work badly with not enough represented population groups.
A scientist in the USA discovered in 2019 that the algorithm used in American hospitals for the prediction of which patients need medical care gave a privilege to white patients more than to black ones by a great margin. As medical care expenses indicate the needs of a human in medical care, this algorithm takes into account the health expenses of patients in the past.
This figure was associated with race in a significant grade. Black people with the same diseases pay less for medical care, than white ones with the same problems. The scientists and the medical service provider Optum cooperated to make the Biased system less by 80%. Although, if there were no doubts about artificial intelligence, the AI preconditions would have discriminated against black people.
Imagination that CEOs can exclusively men
27% of Chief directors are women. Although, according to the reports of 2015, 11% of people emerging in Google picture search by the key „CEO“ were female representatives. Later, Carnegie Mellon University made is independent study and concluded that the online advertising Google showed more high-income positions for males, than females.
Google reacted indicating that advertisers can point to the persons and web portals to which the search engine must show this advertising. One of the features set by the companies is gender.
Nevertheless, it has been an assumption, that the algorithm of Google could define itself that men are more suitable for leading positions at companies. Researchers think Google could make it based on the behavior of the users. If, for example, men are the only people who see and click on the ads for high-income vacancies, the algorithm will be able to learn to give these ads only to males.
AI Bias algorithm common in personnel hiring by Amazon
Automation played a key role in Amazon’s domination over other companies in e-commerce. Some people, who worked with the company, stated that it uses artificial intelligence in hiring staff to assign 1 to 5-star rankings to job seekers, which was similar to the customer’s estimate products on the Amazon platform. When the company noticed that is new Biased system cannot assess the job seekers who are looking for software developers positions and other leading positions in a gender-neutral way, mostly because it was biased concerning women, the company made necessary adjustments to create a new non-biased ranking system.
After analyzing the summary of the computer model of Amazon, the similarities are in the applications of candidates. Most applications were drawn by males, which certifies that there are more men in this area. The algorithm in Amazon concluded, that male candidates are preferable. Thus, it punished CVs containing that a job seeker was a woman. It also reduced the number of applications from those people who visited one of two women’s educational establishments.
After that Amazon made software changes to make them neutral in relation to these keys. However, it does not prevent emerging of other AI Biases during its work. HRs used the proposals of the tool for searching for new staff, but never fully depended on these ratings. After the Amazon leadership lost their belief in this initiative, the project was closed in 2017.
AI Bias algorithm common in personnel hiring by Amazon
Automation played a key role in Amazon’s domination over other companies in e-commerce. Some people, who worked with the company, stated that it uses artificial intelligence in hiring staff to assign 1 to 5-star rankings to job seekers, which was similar to the customer’s estimate products on the Amazon platform. When the company noticed that is new Biased system cannot assess the job seekers who are looking for software developers positions and other leading positions in a gender-neutral way, mostly because it was biased concerning women, the company made necessary adjustments to create a new non-biased ranking system.
After analyzing the summary of the computer model of Amazon, the similarities are in the applications of candidates. Most applications were drawn by males, which certifies that there are more men in this area. The algorithm in Amazon concluded, that male candidates are preferable. Thus, it punished CVs containing that a job seeker was a woman. It also reduced the number of applications from those people who visited one of two women’s educational establishments.
After that Amazon made software changes to make them neutral in relation to these keys. However, it does not prevent emerging of other AI Biases during its work. HRs used the proposals of the tool for searching for new staff, but never fully depended on these ratings. After the Amazon leadership lost their belief in this initiative, the project was closed in 2017.
How AI Bias Can be Prevented?
Based on the above-mentioned issues, we would like to propose some ideas to overcome occurring of Bias algorithms in our life and work.
Trying machine teaching Bias algorithms in life
For example, candidates for a job. The AI-based decision you made may not be trustworthy if the information of your computer tuition system is given by a certain group of candidates. Although it cannot be a problem, if you apply artificial intelligence to the same seekers, the issue occurs when you apply it to another group of candidates, which your data set did not include before. In this case, it looks like you ask the algorithm to apply the preconditions, which it found out about the previous seekers, to the group of people with the wrong assumption.
To prevent this artificial intelligence bias and find a solution, you need to perform testing for the algorithm in such a way as you could use it in your practical life.
Accounting for justness in Bias in AI prevention
Moreover, we should understand that the term “justness” as well as the way it is calculated must be discussed. It can change under an influence of external factors, which means the AI should consider such changes as well.
Scientists already created many methods to make artificial intelligence systems meet them, such as the preliminary treatment of data, changing the choice of postpartum system, or integrating a certain justness into a tuition program. The contrafactual justness is its method warrantying that the choice of the model would be equal in the contrafactual environment where susceptible features, such as gender belonging, race type, or a sexual focus.
Considering the “Man in a cycle” system
The purpose of the “Man in a cycle” system is to make what a man or a computer cannot do themselves. In case a PC is not able to address an issue, people must help and find a solution instead of a machine. This procedure causes an unbroken feedback cycle.
This unbroken feedback teaches the system and increases its productivity at every further launch. Thus, the participation of a human in this cycle leads to more precise seldom data sets and increased safety and accuracy.
Creating a non-biased system by making changes in technical education
Craig Smith in his article published in the New York Times, while suggesting fighting with Bias in technology, expressed his opinion, that we need to make serious changes in the ways people obtain knowledge in the field of technological science. He states we need to create reforms in technical education. Nowadays, education is based on an objective point of view. We need to make it on a more inter-disciplinary level and educational revision.
He declares we need to consider and agree with some important issues globally, while other problems should be discussed on the local level. We must create regulations and rules, manage authorities and specialists, supporting control of such algorithms and events. More various collecting of information is only a single criterion, but it will not address the artificial Bias problem.
Conclusion
Biases in all fields of our social, private, and professional life are very important issues. It is very hard to overcome them only by trusting the ordinary computation methods based on AI and standard assumptions. Bias phenomena can cause errors associated with the wrong interpretation of collected data by algorithms. This problem can lead to wrong results and bad productivity in science, production, medicine, education, and other spheres. It is necessary to fight biases using testing methods, creating fair systems, allowing the right human to interfere in the automated computation processing, and changing methods of education.