- AI is the new electricity. It is fundamentally and radically changing the fabric of our world, from the way we conduct business, to how we work and live, make decisions, and engage with each other, to how we organize our society, and ultimately, to who we are. Organizations are starting to adopt AI to increase efficiency, better engage customers, and make faster, more accurate decisions.
- Like with any new technology, there is a flip side, a dark side, to AI – machine biases. If unchecked, machine biases replicate, amplify, and systematize societal biases. Biased AI systems may treat some of your customers (or employees) differently, based on their race, gender, identity, age, etc. This is discrimination, and it is against the law. It is also bad for business, including missed opportunities, lost consumer confidence, reputational risk, regulatory sanctions, and lawsuits.
Our Advice
Critical Insight
- Machine biases are not intentional. They reflect the cognitive biases, preconceptions, and judgement of the creators of AI systems and the societal structures encoded in the data sets used for machine learning.
- Machine biases cannot be prevented or fully eliminated. Early identification and diversity in and by design are key. Like with privacy and security breaches, early identification and intervention – ideally at the ideation phase – is the best strategy. Forewarned is forearmed. Prevention starts with a culture of diversity, inclusivity, openness, and collaboration.
- Machine bias is enterprise risk. Machine bias is not a technical issue. It is a social, political, and business problem. Integrate it into your enterprise risk management (ERM).
Impact and Result
- Just because machine biases are induced by human behavior, which is also captured in data silos, they are not inevitable. By asking the right questions upfront during application design, you can prevent many of them.
- Biases can be introduced into an AI system at any stage of the development process, from the data you collect, to the way you collect it, to which algorithms are used, to which assumptions are made, etc. Ask your data science team a lot of questions; leave no stone unturned.
- Don’t wait until “Datasheets for Datasets” and “Model Cards for Model Reporting” (or similar frameworks) become standards. Start creating these documents now to identify and analyze biases in your apps. If using open-source data sets or libraries, you may need to create them yourself for now. If working with partners or using AI/ ML services, demand that they provide such information as part of the engagement. You, not your partners, are ultimately responsible for the AI-powered product or service you deliver to your customers or employees.
- Build a culture of diversity, transparency, inclusivity, and collaboration – the best mechanism to prevent and address machine biases.
- Treat machine bias as enterprise risk. Use your ERM to guide all decisions around machine biases and their mitigation.