Bias in generative AI: Challenges and mitigation strategies

- [Instructor] So in this video we'll discuss algorithmic bias and generative AI, a complex but not insurmountable challenge. According to News Guard, factually incorrect, AI generated articles have increased by more than 1,000% just last year. UCS researchers have discovered bias in up to 38.6% of so-called facts that were used by generative AI models when generating new content. What's fascinating is how differently bias manifests in traditional ML versus generative models. Traditional ML models might show bias through measurable disparities in accuracy or performance across different groups, but generative models can perpetuate subtle cultural biases that are hard to quantify, but equally harmful, if not more so. A study by Edith Cohen University analyzed AI generated images from Midjourney and focused on how it depicts Olympic teams from various countries. In his study, he found that men were depicted five times more often than women with 82.86% of images showing male athletes. And in analysis of more than 5,000 AI images that were created by Stable Diffusion's text-to-image model, Bloomberg found that images associated with higher paying job titles featured people with lighter skin tones, and that results for most professional roles were male dominated. Here's what makes addressing this particularly challenging. Bias often originates in the training data itself, and it's not always obvious. Teams preparing data sets might not even realize they're introducing bias. This is why prudent companies are proactively implementing robust frameworks for assessing and mitigating bias before they release their AI models. Ensuring ethical AI development requires a structured approach that integrates ethics into the development process rather than treating it as an afterthought. Let's explore key strategies organizations are using to build responsible AI systems. One, there's the team integration and testing. The most effective AI companies embed ethics specialists directly within their development teams. Instead of addressing ethical concerns at the end of development processes, these specialists proactively identify and mitigate bias early on. Some best practices include daily bias detection, which is where they incorporate ethical checks in daily standups to identify potential biases in real time. And then they're also doing proactive issue identification. This is where they do a regular review of data sets and algorithms to catch potential ethical concerns before they become systematic problems. I've seen this work particularly well at a FinTech company that made bias detection part of their daily standups. Along with technical progress, they also highlighted potential bias concerns they've encountered. This proactive approach helped them identify and address issues before they became embedded in the system. Then there is the community engagement layer. A diverse range of perspectives is critical for ensuring AI models are fair and inclusive. Engaging with communities affected by AI systems helps organizations better understand ethical risk. This includes diverse user training where you ensure that models are evaluated across different demographics and use cases. It also includes active dialogue where you're having regular discussions with stakeholders to identify concerns and gather feedback. And lastly, you can consider impact assessments, which is a continuous evaluation of how AI decisions affect different groups. Real-time bias detection is essential in AI ethics. Companies are now using secondary AI models to monitor outputs for fairness and inclusivity before they're deployed. This approach, often referred to as LLM-as-a-judge, allows one AI model to evaluate the outputs of another, helping to minimize bias. This is similar to having an LLM, let's say GPT-4o, judging the output of a smaller model, for example, GPT-4o mini, or any other model, perhaps Claude Sonnet 3.5. You can prompt the model to specifically check for fairness, inclusivity, and bias in the output before being presented to users. Other examples include AI models monitoring each other, where some organizations use LLMs like GPT-4o to assess the fairness of smaller models before outputs are presented to their users. Or also automated bias flags where algorithms are used to detect potential stereotypes or demographic imbalances in AI generated content. To build ethical AI, bias mitigation must happen at multiple levels. First, there's training data curation, and this is where you're ensuring data sets are diverse, representative, and free from historical biases. Second, there's model architecture design where you're implementing fairness constraints within the model itself. And then third, there's output filtering and controls, which involves applying post-processing techniques to catch biased outputs before deployment. For large enterprises, scaling bias mitigation efforts requires additional measures such as dedicated bias monitoring systems, regular community consultations, and transparent impact assessments and clear incident reporting protocols. By embedding these ethical practices into the AI lifecycle, organizations can build more responsible, fair, and trustworthy AI systems.