Understanding AI Bias

Machine learning is now increasingly available across the entire Microsoft portfolio, ranging from tools targeting the data scientist (Azure ML Studio and Jupyter notebooks) to developer libraries (Azure Cognitive Services and ML.NET), and is embedded in nearly all its applications. However, concern is growing about bias introduced by AI, that is, groups that are incorrectly over- or underemphasized, leading to unequal or unfair treatment.

At its core, machine learning works by analyzing large quantities of data, deriving correlations and relationships among many variables (hundreds, thousands, or even more—OpenAI’s GPT-3 purports to use some 175 billion, and a model developed in Beijing allegedly employs over a trillion) and creating a mathematical model describing the input, or training, data. Such a model can then be used with new data to predict future results.

The development cycle for an AI model typically consists of several phases:

Cleansing the data, that is, ensuring that all records have the requisite quality (for example, no blank or invalid fields). This is usually (by far) the most time-consuming phase. (See the sidebar “The Many Dimensions of Data Quality.”)