Building unbiased artificial intelligence: Inclusion and Diversity across a team developing A.I.

Diversity plays an essential role in teams creating AI systems. It is critical that as AI systems perform more tasks on their own, these systems reflect objectivity and consideration. Datasets and trained models must integrate the range of perspectives of the end users of these systems — otherwise bias, exclusion and discrimination surface and are reinforced. I operate from a belief that you can’t be what you can’t see. Scientists and developers training machines should share that philosophy. A machine that hasn’t seen objectivity simply will not have the ability to be objective. When building the team culture of Crane Ai we made it a precedent to put inclusion at the center of your company — in fact we mapped our hiring process to how we wanted our A.I. system to behave.

I operate from a belief that you can’t be what you can’t see. Scientists and developers training machines should share that philosophy. A machine that hasn’t seen objectivity simply will not have the ability to be objective.

Building the Team

Creating the team wasn’t based on hiring more diverse looking people or a specific amount of a certain gender/race/age/culture . We believe that approach is opposite of inclusion. Instead we focused on teaming with people from different backgrounds with different skillsets who want to come together as a team to make A.I. systems that solve specific problems.

We processed hundreds of applications, conducted many interviews and reached out to recommended candidates over several months. The most important criteria in our process was passion. People who shared our identical passion of innovating through learning, teaching and collaborating as a team. The goal of our AI system was no different — delivering an innovative experience powered by an active learning/teaching process. Among other characteristics we looked for were interests and experience. Over time we discovered that while the overlap was great for team chemistry the range informed details of the AI system that made a significant difference to the end user experience given we were able to test a wider variety of cases internally before shipping product.

CraneAi Research and Engineering Team

Key Wins Resulting from a Diverse Team

Computer Vision

CV is one of the most important components of our AI system. While CNN’s (https://en.wikipedia.org/wiki/Convolutional_neural_network) service many of our use cases we seen shortcomings in key scenarios where we needed to enlist more robust approaches. Annotating data to train our custom model seen the greatest success when the diversity in career experience for the specific domain helped us identify less obvious use cases resulting in increased prediction confidence. Team members who had a greater familiarity with certain patterns were able to annotate and label our data more effectively tripling the output confidence.

NLP/Conversation

In multi-turn conversational interfaces (such as chats) managing out of scope references, who’s being referenced in conversation or scope segues is incredibly challenging due to the nuance of how we say things vs how we communicate in digital-short-form. To overcome that challenge we tapped the team for cultural input to manage the range in classification possibilities. In one case the engineers referred to “login” to “authentication”, while other team members used “sign in” to describe the same process. Such a detail could challenge a machine’s accuracy. Taking input from each team member we conducted several white board sessions and engineered a knowledge graph.

Knowledge Graph

The knowledge graph became one of the key utilities in coupling the backgrounds and perspectives of the team. Our knowledge graph functions similar to a brain in the sense that it stores the things it knows as memories. The nodes in the graph are used to weight the results the machine produces. Beyond rapidly bootstrapping team members with new data it also is critical in enriching vague ideas and providing the machine greater context to make more accurate decisions. Today we frequently debate topics in open forums and add key resources to the graph. At every turn our performance increases and the benefits are epic!


We are active on our journey to inspire inclusion and empower AI teams to become more diverse. We recognize that moving the needle isn’t easy and fraught with red tape, we plan to continue to grow and influence long-term change to help reduce bias and discrimination in AI systems of all kinds.

For more information about Crane Visit our site

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.