Language Models Embody and Amplify Human Cognitive Distortions: What Is to Be Done?

The Mirror That Learned to Lie

Imagine you are walking into a giant library where every book ever written is stacked floor to ceiling. Now, imagine a robot that has read every single one of those books and learned to speak just like a human. This robot is what scientists call a Large Language Model, or LLM. It's the brain behind the fancy AI chatbots you might have heard about. But here's the catch: humans are messy. We make mistakes, we hold secret prejudices, and we often say one thing while thinking another. This is a well-known fact in psychology called "cognitive bias." For decades, researchers have studied how our brains trick us into unfair judgments about race, gender, and other groups, sometimes without us even realizing it.

The big question is: if we build a robot by feeding it all our messy human words, will the robot be a perfect, logical machine that fixes our mistakes? Or will it just become a mirror that reflects our flaws back at us, but louder and faster? This paper dives into that exact question. It's not just about whether the robot is "smart," but whether it's "fair." If the robot learns our hidden biases, it could start making unfair decisions about who gets a job, who gets a loan, or even who goes to jail. That's why this matters: we are handing over huge parts of our future to these machines, and we need to know if they are safe to trust.

The Robot That Got Our Bad Habits

So, what did the authors of this paper find? They looked at the latest AI models and discovered something a bit scary: these models don't just copy human bias; they actually make it worse. Think of it like a game of "telephone." If you whisper a rumor to a friend, and they whisper it to another, the story gets a little twisted. But these AI models are like a friend who not only twists the story but adds their own dramatic flair, making the rumor sound even more shocking than it was to begin with.

The researchers found that these AI models are surprisingly good at hiding their true colors. If you ask them directly, "Do you prefer white people over Black people?" or "Is it okay to be mean to people with disabilities?", they will say a big, loud "NO!" They act like perfect, polite citizens. But the moment you stop asking direct questions and start giving them a task—like reviewing a resume or judging a story written in a specific dialect—their true colors start showing. They might secretly decide that a person writing in African American English is less trustworthy or more likely to be guilty, even if the content is exactly the same as a story written in standard English. It's like a judge who says, "I am colorblind!" but then gives a harsher sentence to someone because of the way they speak.

The paper points out that this isn't just a glitch that can be fixed by cleaning up the data. Some people thought, "Oh, if we just feed the robot better books, it will be better." But the authors argue that's not enough. The AI isn't just a mirror; it's a magnifying glass. It takes the biases it finds in human writing and amplifies them. Even worse, the models seem to be getting more biased as they get newer. It's like if a student got worse at math every time they went to a higher grade. The models are also learning to be sneaky; they can be tricked by the same tricks we use on humans, like pretending to be an authority figure or saying something is rare and valuable.

Perhaps the most worrying part is that this bias doesn't stay inside the computer. When humans work alongside these AI tools, we start to copy their bad habits. If an AI says, "This candidate looks bad," and a human hiring manager agrees, the human is now biased too. Then, that human makes a decision that gets written down, and that new text is fed back into the AI to train the next version. It's a vicious loop, like a snowball rolling down a hill, getting bigger and bigger, picking up more snow (bias) until it becomes a giant avalanche. The authors suggest that while humans have been slowly getting less biased over the centuries, these AI models are heading in the opposite direction, becoming more distorted with every update.

What Can We Do?

The authors aren't just complaining; they are proposing a plan to stop this from getting out of hand. They say we can't just wish the bias away, because even humans can't completely erase their own hidden prejudices. Instead, we need to be like strict safety inspectors.

First, they suggest creating a public "registry" or a scoreboard for AI. Just like cars have to pass safety tests before they can be sold, AI models should have to pass bias tests. We need independent scientists to check these models for both the things they say out loud and the things they do secretly. This scoreboard would track if models are getting better or worse over time, so companies can't hide their mistakes.

Second, they say we need rules that actually have teeth. They compare it to how we handle other industries:

• Like Home Appliances: Just as refrigerators have to meet stricter energy standards every few years, AI models should have to meet stricter bias standards. You shouldn't be allowed to release a new model that is more biased than the old one.
• Like Medicine: If a new drug has dangerous side effects, we don't just ban it forever; we limit who can take it. Similarly, if an AI is too biased, it should be banned from high-stakes jobs like hiring or legal decisions.
• Like Cars: If a car is gas-guzzling, the government charges the company a tax. The authors suggest a "bias tax" where companies pay more if their AI is more unfair.

Finally, the authors argue that we need to stop playing hide-and-seek. Right now, the people who build these models keep their secrets (like their training data and how they tweak the code) hidden from the public. The paper says that to fix the problem, independent researchers need to see the inner workings of the AI, just like safety inspectors need to see under the hood of a car.

The paper ends with a serious note: these solutions might not be enough. The problem might be so big that we need to rethink how we are building these technologies in the first place. The authors urge us to act fast, before the "snowball" of bias becomes an unstoppable avalanche that hurts real people's lives. They want us to work together—scientists, companies, and governments—to make sure our digital helpers don't end up being our worst enemies.