Skip to content Skip to footer

The Limitations of Artificial Intelligence in the Real World

Artificial intelligence services might look great on paper, where imagination is key, but it faces limitations when it comes to integrating it in the real world. In the recently published Forbes article “Moving AI from the Safe and Secure Lab into the Chaotic Real World”, I discuss the unpredictable challenges that Artificial Intelligence faces when we set it loose in a new setting. 

Usually, algorithms can effectively predict behaviors in established settings, but when the layout set up changes, the behavior patterns turn unpredictable. This happens because the market conditions or the parameters of the business case change too fast, making the data unreliable or affecting the algorithm created for it. 

It is important to highlight how the COVID-19 pandemic affected industries everywhere. No one could have imagined the significance of it or the consequences it brought. We witnessed this first-hand when the accuracy of a model we developed to predict patient churn at a major healthcare insurer decreased drastically in the real world when compared to our 90% accuracy in the lab. The algorithm predicted the likelihood of a decrease of customers during open enrollment periods so the insurer could take actions to reduce churn. During the height of COVID, people avoided doctor appointments, and because of this, the results of the algorithm were off. On the other hand, a human doing the evaluation would have noticed that the doctor visits waned, not because customers were not interested, but because no one was scheduling appointments at all. The algorithm wasn’t able to differentiate between patients avoiding doctors because of COVID and those who were planning to leave the health plan. 

Other factors can change the effectiveness of an algorithm, such as natural disasters and changes in laws or regulations. For example, an algorithm that can predict home prices in California might lose effectiveness if a new tax law makes it less financially advantageous to purchase a home. 

Algorithms depend on at least 10,000 data points to be reliable. If the behavioral patterns change by the time a solution is released, data scientists would need to establish and add new data points. 

Artificial Intelligence in the Real 2022 World

Useful Tactics for Successful Artificial Intelligence

There are six ways in which companies can move their algorithms from the lab to the real world successfully. These are:

  • Gathering as much diverse data as possible. By accumulating as much diversity in data as possible, you improve the accuracy of your algorithm because it will represent the gamut of people that will come into contact with or be affected by your algorithm. Take time to consider the biases that might accidentally be integrated into a program, specifically those apps that help humans make decisions that could affect others, such as who gets approved for a loan. Because of this, all AI solutions should be continually tested for biases or reviewed for harmful outcomes.
  • Making AI practical. Aim for AI solutions that are relevant to today’s business problems instead of long-term projects that focus on pure AI to create more human-like machines.  As Melvin Greer, Intel’s chief data scientist for the Americas, noted, “I approach AI implementation as moving investigation and adoption outside the laboratory into real-world, reliable, and practical applications. We have a major gap between AI ambition and reality.
  • Getting Data Scientists out of the lab and into the world. It is important for data scientists to possess different skills, apart from their technical skills, because, by having soft and business skills, they might have a better understanding of user needs as well as the challenges present in the business world. They need to get out into the field and learn about the people impacted by their product and how their applications will be used to solve real-world problems.
  • Using synthetic data to see what could happen. Data scientists can fast track the development of their app by using synthetic data to initially test their solutions and make any adjustments necessary. Real-world data will continuously increase its accuracy once the solution has been created. 
  • Anticipating and addressing the negative consequences as soon as possible. It is important to consider the aftermath or possible consequences of an AI solution before releasing it into the world, given the current concerns of social platforms collecting too much data on people and the negative effects this could have. 
  • Focusing on continuous improvement. An AI program is a constantly evolving process due to changing scenarios, new data, and possible outcomes. The goal should be to continue to expand and improve the AI solution. 

The real world, unlike the controlled environment of a lab, is a chaotic setting with constant surprises.  Data scientists recognize that an algorithms’ effectiveness relies on more than the data or the science used for it. It depends strongly on the understanding of human behavior, emotions, and unpredictable situations. While it is not an easy journey, it is the most effective way to make sure you develop pragmatic solutions. 

Artificial Intelligence in the Real 2022 World

Get the best blog stories in your inbox!