Language Models: A Mirror to Ourselves

This article was published in “El Economista”, in December 2024.
by Rubén Castillo (Clintell's Data Lead)

When we think of artificial intelligence, especially language models like ChatGPT, the image that often comes to mind is that of a precise, efficient, and perfectly rational machine. An intelligence devoid of emotions or contradictions, capable of processing vast amounts of information and providing answers rooted in pure logic. But is that really the case? In reality, language models are deeply rooted in one of the most human aspects of us: language. And with it, they are exposed to all our biases, tendencies, and behaviors.

There are tasks where humans are fully aware of our limitations. As Daniel Kahneman and Amos Tversky demonstrated in their research on behavioral economics, humans have clear difficulties in understanding probabilities and tend to make systematic errors in judgments under uncertainty. A curious example of this can be found in forensic statistics, where a rule known as “Benford’s Law” is used to detect financial fraud. This law predicts a specific distribution of leading digits in certain natural datasets, such as financial statements. When humans attempt to fabricate such numbers, they unknowingly deviate from this distribution, providing a clue to potential manipulation.

It’s unsurprising that humans stray from statistical ideals.

But would it surprise us if the same happened with an artificial intelligence model? A striking example was demonstrated by an internet user who asked ChatGPT to generate a random number between 1 and 100. Surprisingly, the model showed a notable tendency to choose the number 42—a nod to the science fiction work The Hitchhiker’s Guide to the Galaxy by Douglas Adams, where 42 is “the answer to life, the universe, and everything.” This behavior might seem anecdotal, but it is, in fact, a perfect example of how language models learn not only logical patterns but also cultural references and popular ideas embedded in human language.

These tendencies stem from how language models are trained. They use data from texts written by millions of people, encompassing a diverse range of human behaviors, both rational and irrational. Every idea, bias, and preference is contained within that data. In this way, what AI reflects back to us is a kind of mirror of our collective words and decisions. When AI appears to be “irrational” or biased, it is simply replicating our own irrationality or biases.

Language models are also a reflection of our social biases.

The words we choose and the way we describe certain situations or people reveal our perceptions and attitudes, and language models learn these biases. If the training data contains a tendency to describe certain groups of people negatively or to associate specific job roles with particular genders, the AI will also learn and reproduce those associations. This serves as an uncomfortable reminder of how our words are not neutral and how biases can be perpetuated when left unchallenged.

This leads to an interesting reflection: language models not only help us automate tasks or improve processes, but they also offer a unique opportunity to look in the mirror and understand the type of information we are generating as a society. They show us how prevalent stereotypes are in our discourse, how consistent (or inconsistent) we are in our ideas, and even how often certain cultural myths are repeated.

In this sense, AI can act as a catalyst for positive change. By recognizing that models replicate our errors and biases, we can also work to correct them. This forces us to review the data we use to train these intelligences and to become more aware of our own limitations and prejudices. We cannot expect AI to be perfect when it is trained on imperfect data, but we can learn from its errors to improve our understanding of how we think and act.

Interestingly, this ability of AI to reflect our behaviors also reveals the beauty and complexity of human language. Language is not just a set of grammatical rules and vocabulary; it is a tool that encapsulates culture, history, emotions, and contradictions. Language models, exposed to this richness, cannot help but reproduce both the best and the worst of our expression. This proves that artificial intelligence, far from being a cold and distant entity, is as full of humanity as the data we feed it.

So, the next time you interact with AI, remember that, in a way, what you’re seeing is a reflection of yourself and all of us. AI is not an impartial machine devoid of feelings; it is a mirror of our words, decisions, and biases. We may not always like what we see, but that is also an opportunity to learn and perhaps change the way we speak, think, and act as a society.