Salesforce researchers explain how they build AI systems with ethics baked in from the start. They tackle biased training data, monitor models for new biases that emerge over time, and give customers tools to detect problems before deployment.
AI has really come out of the lab the last couple of years and into real world applications. The consequences or unintended consequences of AI has become much more apparent to people. And they are seeing that potentials for bias and lack of transparency and explainability can have real implications for everyday life.
I think one aspect that makes Salesforce Research different to other research teams is that we try to infuse ethics into our entire process. We really start from the beginning of the research. Should we be researching this? We want to see what could the potential downsides of it be. We want to start including those kinds of thoughts into the entire process of our research.
We always want to make sure that we're not working in an echo chamber. We want to learn from other companies what they're doing and from human rights groups so that we're constantly learning and evolving our practice and ensuring that what we are building gives back to society.
There's a lot of interest and I guess a lot of positive and negative emotions around artificial intelligence today. We have to be very thoughtful about how we are going to deploy AI in the real world and how we as product makers, policymakers, perhaps, are going to let technology interact with our lives.
Every day to build ethical AI. We have to think about those unintended consequences of how the AI might be used beyond what our intention is. We do know that some of the open source training data sets that are available today are not necessarily representative or fair. They may be too heavily skewed towards one gender or one group. And so we think about how do we deal with those training data sets? How can we improve them? We may remove the bias or prejudice from the training data in the beginning when we first launch it, but then it gets exposed to new training data and it learns new things over time. So we must constantly monitor it to understand what it's doing and make sure that it remains fair and transparent. We also build a number of tools for our customers to help them identify if there's bias in their training data and to make more informed decisions. We really want to empower our customers to use our tools responsibly.
We have over 150,000 companies and organizations on our platform, over 40,000 nonprofits. I think we're large enough to have a lot of collaboration between the researchers to really bring the impact that we want to have to both the academic community as well as to our products and customers.
