Artificial Intelligence (AI) is a term that is popping up everywhere right now. The media is talking about how AI can change our future and how complex the solutions AI can provide in a blink of an eye.
However, the hype around AI, data science, and machine learning (ML) in general can be misleading towards some exaggerated news about what AI can or cannot do. It's important to remember that an AI model is built by a human, and, like its creator, it has flaws. AI, as its name suggests, tries to simulate human thinking. To be a little more technical, for any AI model to work, it needs to be a little bit wrong. For example, if you want a computer vision model that can "perfectly" differentiate between dogs and cats, the accuracy of this model can't be 100%.
When a model achieves 100% accuracy, it's called overfitting in data science. The overfit model can't deal with other data besides the data that it was trained with. Therefore, the AI that is overfitting won't do any good.
Another example of how flawed an AI can be is the case of racist or misogynist models. These models happen to interpret the data that is given to them while training. In this case, if the data scientist building the model doesn't pay attention to the diversity of color and gender, the model is doomed to be racist and/or misogynist based on their cultural context.
However, AI can solve many problems. It's not the solution for all of them. Let's think about a problem where the output needs to be extremely precise. The client needs 100% accuracy, and any mistake can lead to a huge disaster. Imagine the case of a chemist that needs to create a chemical compound, such as a medicine.
He has a system like an AI to help him request the ingredients for his mixture. This machine communicates with another machine through IoT and delivers the exact amount of components needed for this medicine. In this case, an AI wouldn't be the best approach to solving this problem.
As we discussed in this article, AI models cannot be 100% precise, otherwise they won't be generalist enough to solve their tasks. They need to have a slight error to achieve their most reliable result. In the example of dogs and cats, sometimes the AI will have a cat classified as a dog or vice versa, and that's okay for it to work properly. But in the case of the amount of chemical compounds for a medicine, that cannot be the case.
If the amount is not precise, some medicine could have adverse effects on the patient, leading to a more serious cause. In this case, an AI deciding the amount of compounds is not reliable, although sometimes it'll get everything correct and there can be one time that it got wrong.
In cases like this, we can rely on automation to help us. Automated processes can be 100% precise and repeat the same process with 100% precision. We always need to consider maintenance here, even in AI cases.
The process will require less computer power and less effort among the development team. Moreover, it'll deliver the required result with precision. In the case of the example, we can imagine a software that receives the name of the medicine needed to be made and retrieves the recipe from the database. It then sends the exact amount of components to the machine that will deliver it to the chemistry.
In conclusion, AI models can really solve many problems that we have, such as predicting the productivity of a company in the next month, recommending products for customers to buy, text autocorrect (you may have already disabled that on your phone), and many more. However, relying entirely on AI and expecting it to do everything for you with perfection, even in the near future, is something rather utopian. Sometimes, as Occam's Razor states: "the simplest solution is almost always the best," and automation and other solutions are still solving many of our daily problems. Some solutions are yet to be discovered.
I've always been passionate about research and development, which led me to become a Data Scientist. What caught my attention the most is how "magical" the concept of what AI can or cannot do is. Therefore, I enjoy sharing my knowledge to demystify the concept of all-knowing and movie-like AI.