It is quite evident that AI has a great potential in the global market and with its new innovations, AI is all set to expand its global boundaries in the coming years. It is predicted by PwC, AI will be a potential contributor to the global market by the year 2030 with a total contribution of USD 15.7 trillion (approx.) . This massive growth of Artificial Intelligence has also given rise to many problems. One of the major issues faced by the investors are the biases surrounding artificial intelligence. This concept of bias has been prevalent for quite a long time now.
Let us have a look at some of the biases revolving around Artificial Intelligence:
- Prejudicial bias: In a report on Mitigating bias in Artificial Intelligence presented by Haas School Of Business, it is understood that the bias in Artificial Intelligence systems is present as the systems are developed by humans. The softwares is developed while mirroring the society. In an AI system, these biases can be formed in its developing phase.
- Sample Bias: These biases can also occur due to the inaccuracy in data entered. In a report presented by McKinsey titled ‘ Tackling bias in artificial Intelligence’, it was found that the main cause of bias in the majority of cases is the underlying data. The bias is directly linked to the demographics which is not portrayed by the data presented in the model. A perfect example of this situation is facial recognition AI system. In a case where the system is developed with the dataset containing information on white people, then it will result in an AI model being more biased towards that particular category of people.
- Historical Bias: Biases present in the previous AI systems can get carried on to the new systems in many cases. A very apt example of this bias is when in 2018, Amazon had to dissolve its former AI recruiting tool as the system was extremely biased towards women candidates. On the basis of resumes submitted from over the past 10 years, the AI program was trained on that basis, which contained majorly applications from men , as a matter of fact this became a major reason for the bias in the system.
- Aggregation bias: Bad representation and bias is also caused due to the classification of people belonging to different classes and races. To understand this situation, a wonderful example is of the income growth of a lawyer and athlete. A lawyer enjoys a heavy growth in his earnings as he gains more experience and ages with time. Whereas, when we talk about an athlete, his earnings are at the peak only during his young age. When both are combined, they give rise to bias.
- Evaluation Bias: Bias can occur several times even though the data provided is accurate. This bias takes place during the process of model evaluation. For example, the models have to be screened and necessarily have to undergo testing on the basis of benchmarks. What can be seen is that in the majority of scenarios, these benchmarks are not even related to the model and its criterias. Hence, these situations result in biases. These biases also occur during times when there is new data. Although the AI models are able to function precisely in predicting values out of the training dataset, they fail miserably with the new data.
Treating these Biases
One of the initial steps in solving the biases is to provide accurate data for the programming of the algorithms. This step is highly crucial as it forms the basis of the development and further working of AI models. It is the responsibility of the developers to ensure that the samples and dataset does not lead to bias. In order to deal with bias, McKinsey has also suggested an analysis on subpopulation by the developers of AI models. This analysis will help the developers to identify if the model is matching the subpopulations in the dataset. The company in its report, also emphasized on a routine check of the AI models in order to keep a track of the changes in algorithms and the possible change in the training data. In a research published last year by PwC, it is found that AI ethics framework is present in only 20% of companies. On top of that only 35% intend to enhance the working of the AI systems and its related processes. Indeed it is highly essential to work on spreading awareness, and ensuring higher clarity in the AI systems.
Author- Toshank Bhardwaj, AI Content Creator
