AI Bias and Youth

[Steven Vosloo]: The three examples we can give as around persuasive misinformation and disinformation. I’ve used the word persuasive because increasingly it’s impossible to tell whether something text was generated by a human, or an AI, or images. It can also skew children’s worldview. Again, this is about the data that’s been used in the models. But because much of the data comes from the global north, from the US, from Europe, or from the cohort of people who created internet content, a lot of that doesn’t represent the world view of the people of the global south and obviously this is an issue for UNICEF that’s a global organization. And so children don’t see themselves represented in the stories, or in the examples that they are given help with their homework. We’ve also seen research that says when children use large language models to help them write an essay, the bias in the model, and this is unintentional bias throughout the model trying to shape the child’s worldview, but it does shape the positions that they take in writing that essay, kind of whether they’re for something or against it. And so over time this could merge – change how you see the world. I often think that previously we’ve thought about or that there is a concern that algorithms have moved the minority on the edges to extreme positions. But perhaps the new danger is that the majority in the middle or the majority of users get moved to the middle. And with the kind of a more generic homogenous worldview that doesn’t represent everyone.