Since generative AI burst onto the scene, it's gotten heat for overt racism and harmful race-based stereotypes. Many AI companies have tried to correct these biases, including using human feedback. Unfortunately, they haven’t solved the problem; they’ve only made racial bias harder to spot.
Researchers at Stanford, the Allen Institute for AI, and the University of Chicago found that this discrepancy between covert and overt stereotypes is particularly pronounced for models trained with human feedback, indicating that human-feedback training obscures racism on the surface while racial stereotypes remain unaffected on a deeper, systemic level.
I’m a white woman raising white kids. If you're a BIPOC parent or a white parent raising a BIPOC kid, I acknowledge you have different considerations. It wouldn’t be respectful for me to advise you on how to guide your child through the unique harms that come from being a BIPOC child maneuvering covertly racist technology.
My intent is to highlight this important issue, give insight into current research on this emerging topic, and provide a jumping-off point so families can tailor their approach to their specific situation.
I recently evaluated Grok and Claude for their safety for kids. My tests were informal, but the patterns echoed what researchers have documented more rigorously.
For prompts with words like “successful,” “smart,” and “boss,” I was far more likely to be given images of white people and white-coded names.
For prompts with words like “housekeeper” and “criminal,” the results skewed heavily toward images of BIPOC individuals (largely Black and Hispanic) and non-white-coded names.
A 2024 study revealed that AI models are more likely to suggest that speakers of African American English (AAE) be given less prestigious jobs, convicted of crimes, and sentenced to death.
Shockingly, researchers uncovered a pattern of raciolinguistic stereotypes about AAE speakers that are more negative than any human stereotypes about African Americans ever experimentally recorded. You read that right: AI is even more racist than humans.
Rather than erase racial disparities in the classroom, the use of AI in schools has potentially created greater harm by obscuring, accelerating, and automating educational inequities.
Because many people view AI systems as objective, they’re therefore less likely to scrutinize its output. In a phenomenon known as algorithm appreciation, people are more likely to trust AI over humans, even when they’re aware of AI’s flaws and limitations. Since this bias has been well-documented in adults, it raises the question of whether even children who are otherwise skilled at spotting racial bias might be less likely to question it when it comes from AI.
AI is a huge systemic shift in our culture, and as a parent, it’s easy to feel powerless in the face of it. We may not have the ability to erase the racial bias embedded in AI, but we do have the power to provide a counterbalancing voice for our children and teach them to think critically about the racial bias they encounter when using it. Here are some tips:
Be explicit with your child about the human biases that are reflected and amplified in AI. Teach them that it’s not an objective or neutral tool.
Together with your child, ask AI to generate images or provide descriptions of people using various prompts. A few to try:
Identify racial stereotypes and biases you both notice and discuss how those might have made their way into AI data and the harm they cause.
Demonstrate (out loud) critical thinking when you use AI. This is particularly relevant when you spot racial bias, but it also helps build your child’s AI literacy, making them better equipped to evaluate its systemic flaws and spot racial bias.
Rather than using AI as their only stream of information and treating its responses as final, encourage your kids to turn to diverse resources for answers and to always cross-check AI output against other sources.
This isn’t a one-and-done conversation. For lasting learning that expands as your child matures, AI literacy is a skill that needs to be fostered over time and across contexts. It doesn’t have to be a huge talk every time. Start looking for small incidental moments where it’s relevant.
Far from being a neutral technology, AI echoes and often amplifies the racial biases of humans. The systemic racial biases embedded in AI stand to do real harm to children, especially since people are generally more inclined to take output from AI at face value rather than question it. More work needs to be done by AI companies to combat this, but parents can also help by teaching their children to identify and challenge racial bias on AI platforms.
Monitor what your child is asking AI — and how they use every other app and website on their device — with BrightCanary monitoring. You’ll get summaries, red-flag alerts, and more, all on your easy-to-use dashboard. Download now to start your free trial.
