How artificial intelligence amplifies human bias

(Credit: © Jakub Jirsak | Dreamstime.com)

Dangerous feedback loops between humans and machines

LONDON — Physicians’ unconscious bias can impact patient care. Hiring managers’ biases can influence hiring decisions. But what happens when you add AI to these scenarios? New research shows that AI systems not only reflect human biases, but also amplify them, creating a snowball effect that gradually strengthens human biases over time. produces.

This alarming finding comes from a new study published in the journal Nature Human Behavior, which explores how AI can shape human judgment in ways that compound existing biases and errors. We’re showing you what we can do. In a series of experiments with 1,401 participants, researchers at University College London and the Massachusetts Institute of Technology found that even small initial biases can snowball through repeated human-AI interactions. We have discovered that it can become large. This amplification effect is significantly stronger than what occurs when humans interact with other humans, and suggests that there is something unique about the way we process and internalize the information generated by AI. suggests.

“Humans are inherently biased, so when you train an AI system on a set of human-generated data, the AI ​​algorithm learns the human biases embedded in the data,” the study says. Co-lead author Professor Tali Shalot explains. , in a statement. “AI tends to exploit and amplify these biases to improve prediction accuracy.”

Let’s consider a hypothetical scenario. Healthcare providers can use AI systems to help screen medical images for potential diseases. If that system has even a slight bias, such as being slightly more likely to miss a warning sign in a certain demographic group, over time, human doctors will be able to incorporate that bias into their own screening decisions. may become unconsciously incorporated into the system. As AI continues to learn from these human decisions, both human and machine judgments can become increasingly distorted.

Researchers investigated this phenomenon through several carefully designed experiments. In one important test, participants were asked to look at a group of 12 faces displayed for 0.5 seconds and judge whether, on average, the faces looked happier or sadder. Ta. The first human participants showed a slight bias, classifying faces as sad about 53% of the time. When a computer program called a convolutional neural network (think of it as an AI system that processes images in the same way the human brain does) was trained on human judgments, this bias was greatly amplified, with a 65% increase in classified the face as sad.

As new participants interacted with this biased AI system, they began to adopt its distorted perspective. The numbers tell an amazing story. If participants disagreed with the AI’s decision, they changed their mind almost a third of the time (32.72%). In contrast, when interacting with other humans, participants changed their opposing opinions only about one-tenth of the time (11.27%). This suggests that people are about three times more likely to be swayed by AI decisions than human ones.

Bias amplification effects appeared consistently across different types of tasks. In addition to facial expressions, participants also completed a motion perception test in which they judged the direction of a dot moving across the screen. We also evaluated other people’s performance on the task and found that participants felt particularly They found that they were more likely to overestimate men’s performance.

“Not only can biased people contribute to biased AI, but biased AI systems can also change people’s own beliefs, so people using AI tools need to understand everything from social judgment to basic cognition. “This could lead to even more bias in this area,” said co-investigator Dr. Moshe Glickman. -Lead author of the study.

To demonstrate the real-world impact, the researchers tested a popular AI image generation system called stable diffusion. When asked to create an image of a “financial manager,” the system showed a strong bias, producing an image of a white male 85% of the time. This is very different from real world demographics. After viewing these AI-generated images, participants were significantly more likely to associate the role of financial manager with white men, demonstrating how AI bias shapes human perceptions of social roles We have proven that we can do it.

When participants were falsely told that they were interacting with another person when they were actually interacting with an AI, they internalized the bias to a lesser extent. Researchers believe this may be because people expect AI to be more accurate than humans for certain tasks, and are more susceptible to AI influence when they know they are working with a machine. Suggests.

This finding is particularly alarming given that people frequently encounter AI-generated content in their daily lives. From social media feeds to recruitment algorithms to medical diagnostic tools, AI systems are increasingly shaping human perception and decision-making. Researchers note that children may be particularly susceptible to these effects because their beliefs and perceptions have not yet been formed.

However, the study wasn’t all bad news. When humans interacted with accurate and unbiased AI systems, their own judgment improved over time. “Importantly, we now know that people make better decisions when interacting with accurate AI, so it’s important to improve AI systems to be as fair and accurate as possible,” Glick said. Dr. Mann says.

Bias in AI is not a one-way street, but a circular path where human and machine biases reinforce each other. Understanding this dynamic is critical if we are to continue to integrate AI systems into increasingly important aspects of society, from healthcare to criminal justice.