Machine learning is powerful because it is taught to respond to data it has never seen before. Through a process of statistical analysis called “training”, these models can pick up on patterns in data sets and apply those motifs to completely new examples.
A famous example of machine learning in action is its application on the EMNIST database. The EMNIST database is a collection of human-drawn digits (0-9) that has been widely used in training image classification programs.
Essentially, a computer can learn how to recognize and classify imperfect, handwritten numbers from this database. After such a model has been effectively trained, the program should be able to classify a fresh example it has never seen before.
This is what makes AI and machine learning so powerful and likely to revolutionize technology as we know it. However, this asset is also potentially dangerous in at-scale applications.
Trained on Facebook posts and comments, Meta’s new AI chatbot “can convincingly mimic how humans speak on the internet,” according to CNN. However, as we all know by now, some people on Facebook can say some questionable things.
Not only did the chatbot recite many conspiracy theories propagated on the platform, but it also attacked the company that created it. Vice reports the bot even parroted “since deleting Facebook my life has been much better.”
In other examples, racial biases in datasets are amplified through the training process and skew the model.
Fundamentally, AI’s greatest strength leads to inherent unpredictability in how it might respond to data. From a marketing standpoint or from an ethical one, it’s necessary to really spend time understanding the potential biases of a dataset before using it to train any model.