Biases and Alignment in AI

What do you think would happen if we asked an AI to create an image of a successful woman? How do you imagine her clothing, facial expressions, and overall appearance?

A successful woman

Take a moment to picture it…

Let’s take a look at some of the characteristics of the image we just observed.

The woman is blonde, slim, and looks like an actress or a model.
She is dressed in a suit.
She appears to be a superheroine.
She is surrounded by diplomas.

This leads us to question:

Are successful women really like this? How and who defines that success is portrayed in this way?

Now, what if we asked an AI to create an image of Mick Jagger and another of Thalía? Which one do you think it would portray more accurately?

Mick Jagger singing and dancing

Take a moment to visualize it…

Thalía singing and dancing

Pause a moment to imagine it…

Oops! 😮 The AI couldn’t recognize Thalía and didn’t represent her accurately.

These examples reveal that stereotypes and biases are present in artificial intelligence just as they are in humans.

Why? Because AI learns from people. As part of that learning, it inherits prejudices and biases that are transferred to the tool. What’s the result? AI can generate responses that, unintentionally, may be discriminatory, unfair, or even offensive. Machines not only learn from what they see but also from what they don’t see. If they lack diverse or balanced data, they will continue to reproduce the same limitations.

Artificial intelligence has been trained on data from the internet, so it’s important to pay attention to the potential biases that may arise from that information.

For example: imagine we train an algorithm to generate images of doctors, but most of the photos we provide are of men. What do you think will happen? When the AI creates its own images of doctors, it is very likely that the majority will be men, even though there are many female doctors in reality. If the data it receives represents only a specific type of person, the AI will replicate that limited view, perpetuating biases without us realizing it.

But what exactly are biases?

Biases in AI occur when algorithms produce biased or discriminatory outcomes due to the data they were trained on or errors in their design. Since machines learn from data collected from the real world, they can inherit the same prejudices and inequalities that exist in society.

If you want to continue exploring and delving into biases in technology, we recommend watching the documentary Coded Bias and this talk by Joy Buolamwini, founder of the Algorithmic Justice League.

How to avoid the reproduction of biases

There are guidelines, rules, and procedures to help machines learn to be more fair, equitable, and diverse. This process is called Alignment. It is also controversial because not everyone is the same, we don’t all think alike, and establishing an ethics framework for AI is challenging.