It is a well-known fact that Artificial intelligence technologies take businesses to a different level. Businesses have also shown immense interest in adoption AI as part of their process and vision. However, it is important to understand the risks that come with the benefits. Companies can go smart with the use of AI, at the same time, they can be smarter if they understand the risks that come along with it and prepare for the mitigations and workarounds so that their solutions are much more adaptable and acceptable. Understanding what can go wrong with technology is a crucial part of adopting, designing and implementing new technologies. Overcoming the risks and negatives is an important part of business growth. Let us analyse broadly what can possibly go wrong with a technology that can mimic human intelligence.
What is a black box, what does it do or convey? As the name suggests, it is just a simple black coloured box. We are unaware of what is inside and how it works. Sometimes AI technology and processes come across like a black box. Just as we are not fully aware of what actually goes into human thinking and behaviour, something such as an AI that mimics human behaviour also carries the same black box attribute. AI modelling sometimes contains complex data analysis and algorithms that we fail to understand and explain the reason behind its predictions and behaviour. We may have designed it using complex data sets and training and it may not be explainable or understandable. In businesses that are regulated and require an explanation on the outcome, such complex AI black boxes will prove to be deterrent. For example, a loan application reject can warrant an explanation from the system side and failing to provide one does not reflect well of the business model.
What is a helpful direction, workaround or a solution to this problem is to use simple appropriate explainable algorithms so that the model is not a black box. The use of a correct fit algorithm can render the AI system explainable and appropriate reasoning can be brought out to the benefit of the user. For example, the usage of a decision tree is more explainable. However, attention must be paid to the fact that oversimplifying the data set or algorithm can render the model inaccurate. Therefore set of algorithms and data that is a perfect match to the problem statement becomes imperative during the AI design process.
Data bias is a huge factor in influencing an AI model to predict the outcomes. Once a bias is introduced into the system, it becomes almost impossible to remove the bias and product trustable outcomes. To explain this better, let us take the example of facial recognition systems and processes. We know that data set containing different faces and different profiles have to be fed into a system and it needs to be trained to recognize the same. Extensive rounds of training and testing will result in the AI facial recognition system to function in the correct manner. When we take a data set with faces that are mostly pertaining to a particular type of skin colour or pertaining to a particular gender, then the predictions and decisions will be in favour of the majority of data that is being fed. Assuming that the faces that the data set contain has mostly males, then the AI mode that is designed will fail to work accurately for a female. This is the bias that makes the AI model falter in its predictions and credibility.
The solution to prevent data bias is a process called Data Preparation. Before data preparation, data has to be collected. Data collection must be from a reliable source and should have comparable amounts of different parameters. For example, in a face recognition process, care must be taken to take comparable amounts of male and female faces for training. Once the data is collected, then the data needs to be prepared for the balance. Missing values or bias must be analyzed and appropriate measures to reduce the same must be done. Such extensive data analysis and preparation can reduce the bias and produce a model that is accurate and efficient
Another interesting risk of AI technology is its naivety. You must be wondering how an Artificially intelligent system mimicking a human intelligence system can be naive and intelligent. Let us explain the same with an example. Assume that an AI model is trained and modelled using a data set containing multiple pictures of the dog. The algorithm recognizes the dog using pixel intensities and contours and gets training with numerous data pictures that it has been fed. Now when we produce a picture of a lion to this system, it is highly possible that it recognizes the picture as a dog! Here is the twist to this, we agree that the system is intelligent however it is a mathematical and mechanical system which is modelled and designed based on the data set it has been provided. Therefore it uses its deduction skills and compared the pixel intensities and contours and finds it very similar to the dog pictures you have provided and concludes the same as a dog. It does not understand the difference between the two animals and where it lives and how it relates to the human world.
This naivety of the AI system can also be overcome by using large and extensive data sets. The system has to take in and analyze different varieties of dog pictures, different positions or different profiles of the dogs. It not only has to learn to recognize a dog, it also has to learn to recognize when it is not a dog as well. Once again, feeding the right type of data and preparing the data becomes a crucial factor to mitigate this risk.
Ultimately we recognize and understand that the AI system largely behaves on what is being given to it as input and what training it has undergone to make its prediction. Diligent work needs to done before we even think of using AI into our processes because we finally need to produce outcomes that are more efficient and quicker than the current processes.
We can help you build cloud and data products faster. Our Microsoft certified professionals can help you in driving business value with exceptional quality.Contact us