AI Governance — Have the (AI) cake and eat it too!
“Bank fined $175M for charging higher fees and interest rate to more than 30,000 minority borrowers”
“Credit card company fined $75M for selling products to Spanish speaking customers at a much higher interest rate compared to others”
Such headlines (which unfortunately are real) give nightmares to CIO/CDOs. Most of the times, the root cause of these problems is a biased AI model! Executives want to accelerate their adoption of AI to ensure their competitive advantage — but they also want to ensure that they don’t land up in a situation where they are the cause of such headlines. So the multi-million dollar question that needs to be answered is: How to balance these two opposing needs?
Imagine a drum roll and the entry of the knight in shining armor — AI Governance from IBM aka AI Factsheets! It will help you have the (AI) cake and eat it too!
So what exactly is AI Governance? Simply put, AI Governance helps ensure that AI models can be operationalized as quickly as possible without the possibility of leading to the headlines mentioned above. It does this by governing the full AI lifecycle and helps enterprises do three things:
- Know their models
- Trust their models
- Use their models
Knowing your models: IBMs AI Governance capability captures metadata about the model, tracks the data and model lineage and documents the model lifecycle in what is called as Model Factsheets. The metadata captured includes things such as: What was the fairness of the model at development time? What was the quality of the model during validation, what was the fairness of the model in production last week?, etc. Having answers to all these questions at your fingertips helps build confidence in the model and speeds up the deployment of the model in production. In order to explain this point, consider a real world example of what happens in enterprises today: There is a model validation team which validates the model before it is deployed into production. During the validation process they ask questions such as: Why was a deep learning model used and not a simpler model like a decision tree based model? The data scientist who might have built this model months ago will have to dig up his/her notes to find the answer— this might take days. This back and forth delays the deployment of models into production. In this situation having model metadata available to the validator which shows that the data scientist had built a decision tree model but it had poor accuracy and had bias, will avoid the need to get the data scientist in the loop and will speed up the deployment of the model to production.
Trust your model: In order to ensure that biased models do not land up in production, there is a need to have the right enterprise level policies and rules and a way to enforce them across the model lifecycle. Examples of such policies are: A model related to marketing should have fairness above 80% before it can be sent to the model validator for validation. IBMs AI Governance capabilities allows enterprises to define and enforce such governance policies across the enterprise. It also provides the tools to data scientists, model validators and model ops personnel to evaluate the model on different metrics such as fairness, drift, quality, explainability, etc. These metrics can then be used to ensure that the models are thoroughly evaluated along all dimensions at each step of the model lifecycle.
Use your model: The final piece of the puzzle is to ensure that models can be used with confidence by different applications without worrying about the possibility of them acting in a biased manner. For this it is important to monitor these production models and put in guardrails that will ensure that the fairness, drift, quality, etc., of these models does not go beyond pre-defined thresholds.
Thus the trio of know your model, trust your model and use your model ensures that enterprises can leverage the full benefit of AI. In summary, IBM’s AI Governance capability helps organisations achieve the twin opposing goals — speed up the adoption of AI across the enterprise and ensure that the AI models do not lead to bias and undesired outcomes. So if you want to have the (AI) cake and eat it too, give it a try!
