Hurdle #1: How to account for poor decision making and limited data transparency in GenAI?

Companies rely on accurate data to make key decisions every day.  Can you count on GenAI to provide consistent predictions for critical business decisions?  How do you ensure results are accurate and free of biases?

One current example of bias is Google’s Gemini AI and its inaccurate depiction of historical facts. Facing major backlash, Google had to pause work on Gemini AI. An example of poor decision making occurred in 2023 when a prankster tricked a car dealer’s AI Chatbot into selling him a new car for $1.

The beauty of GenAI is the ability to provide creative and dynamic answers to requests, but users should be provided additional information about responses like confidence numbers as well as references.  In addition, GenAI platforms should provide developers with the confidence to transparently debug their machine learning models. In this article, I explore Microsoft’s answer to these types of questions.

User and Developer Confidence

For end users and developers alike, remember the adage “Trust, but verify” as it relates to GenAI.  For end users leveraging Copilot within products like Edge Browser, each GenAI response is followed with “Learn More” links to referenced articles for verification.  If using GitHub Copilot for developer support, a code referencing feature can be enabled to show you referenced repositories.  While there is still work to be done, these small features go a long way to improve user confidence in results.

Microsoft Interpretability Tools

To address poor decision making in machine learning models, Microsoft has created a set of interpretability tools for use by data scientists and developers to understand and improve their machine learning models.  Here is a high-level overview of each tool:

• InterpretML:Includes an open-source Python toolkit that integrates interpretability techniques (ex: SHAP, LIME, and PDP) which provides an API and interactive dashboard to compare and visualize the different AI results/explanations:

• Explainable Boosting Machine (EBM): Opposite of a “black box”, the EBM is a glass box model which lets you see exactly what’s going on inside and how it works while combining that with a fairly accurate prediction.  Boosting refers to adding many small pieces of information to get the “bigger picture” view and is beneficial when understanding and debugging your machine learning predictions.
• Azure Machine Learning: Azure’s cloud platform supports model interpretability throughout the ML lifecycle.  It is used to host InterpretML as well as other tools such as Azure Machine Learning Studio.

Some key differentiators for Microsoft AI Interpretability tools include:

1. Support for a variety of interpretability techniques and algorithms
2. Support for both black box and glass box models
3. Flexible tools that can be standalone Python libraries or integrated with Azure Machine Learning

Microsoft Copyright Commitment

In addition to these interpretability tools, Microsoft is putting their “money where their mouth is” and announced the Copilot Copyright Commitmentfor its customers in September 2023.  At a high-level, Microsoft will defend commercial customers and pay for any adverse judgements if sued for copyright infringement for the use of the Azure OpenAI Service Outputs.  Of course, there are guardrails and mitigations clients must follow to be eligible, but this is a positive step to add confidence in leveraging GenAI within corporate America.

Up Next

In the next article, I will explore Microsoft’s answer to CIOs as they are trying to decide how much to invest in generative AI technology and allocate resources effectively.