Imagine asking your AI assistant to explain a complex concept, but tailoring the explanation to your specific background. This is the promise of personalization in AI, and it's a tricky problem. How do you make sure the AI uses relevant context, like your age or profession, without resorting to harmful stereotypes? Researchers at UC Berkeley have developed a clever technique called Context Steering (CoS) that tackles this challenge head-on. CoS works by measuring how much influence a piece of context has on the AI's response. For example, if you tell the AI "I'm a toddler," it will adjust its explanation of Newton's Second Law accordingly, using simpler language and analogies. But what's really innovative about CoS is its ability to control this influence. You can dial up the personalization for things like movie recommendations, where understanding your preferences is key. Or, you can dial it down to mitigate bias. Imagine an AI tasked with answering questions about different demographics. CoS can be used to ensure the AI doesn't fall prey to stereotypes, promoting fairness and accuracy. The researchers tested CoS on various tasks, including personalized movie summaries and bias detection in question answering. They found that CoS could effectively tailor responses to individual preferences while reducing harmful biases. They even used it to quantify implicit hate speech in online text, demonstrating its versatility. CoS offers a powerful new tool for shaping AI interactions, making them more personalized and less biased. While challenges remain, like handling multiple contexts simultaneously, CoS represents a significant step towards building AI systems that are both helpful and fair.