Your Trusted AI Questions, Answered

What is an acceptable level of bias to strive for in your data?

In an ideal world, all data and subsequent AI solutions would be 100% bias free.

Unfortunately, humans are biased creatures, and data reflects patterns in human behavior. Fortunately, there are ways to identify potential biases in underlying data and mitigate it both at the data level and at stages throughout the solution.

It ultimately comes down to determining the impact of your solution. Early on, ask key questions such as: Who will be affected by this solution and how? What kind of disparate impact may exist? What processes will be put in place to monitor impact over time? How might we mitigate any biases?

These will help establish an understanding of what biases may exist in your data and how they can influence downstream decisions, informing any interventions.

Resources such as the IBM Fairness 360 toolkit as well as built in resources in data science platforms such as H2o.ai provide accessible ways to measure and mitigate bias.

Using explainable and interpretable models over black-box models will also provide insight into how your solution is making decisions.

While we can and should strive to eliminate bias in data, it is crucial to understand the effect solution output will have and measure and mitigate any potential disparate impacts. Advances and tools in the field have made doing so both standard practice and accessible.

– Hannah Arnson, PhD // Director of Data Science