From Romance to Misogyny: What Do Spanish-Language Songs Really Say?

IPN researcher uses Artificial Intelligence to analyze 60,000 Spanish-language songs, revealing how music can reinforce gender stereotypes, misogyny, and discrimination.

Music is woven into everyday life. It fills celebrations, public transportation, social media feeds, and the headphones of millions of people. Yet beneath catchy melodies and memorable lyrics, songs can also convey messages that normalize social problems such as sexism, misogyny, homophobia, and gender stereotypes.

This is the focus of the research conducted by Tania Gisela Alcántara Medina, a Ph.D. candidate in Computer Science at the Research Center for Computing (CIC) of the Instituto Politécnico Nacional (IPN). Over the past six years, she has used Artificial Intelligence (AI) and Natural Language Processing (NLP) to examine the social messages embedded in Spanish-language music.

The project began as a sentiment analysis study based on a dataset of roughly 300 songs. As the research evolved, however, the need emerged to expand the corpus and explore the broader social narratives conveyed through music.

Today, the database includes approximately 60,000 songs spanning genres such as rock, pop, reggaeton, regional Mexican music, trap, and salsa, primarily from Mexico and Latin America.

Building the collection presented several challenges. Accessing song lyrics required overcoming copyright restrictions by demonstrating that the material would be used exclusively for educational and research purposes. The greater challenge, however, came afterward: determining how to classify the songs.

Who Decides What Is Romantic and What Is Misogynistic?

The research quickly showed that concepts such as romantic, sexist, and misogynistic are strongly influenced by cultural context and personal experience.

To minimize bias, the research team assembled diverse groups of evaluators—including men, women, feminists, non-feminists, and participants from different generations and social backgrounds—to label the songs according to the messages they perceived.

The results revealed striking differences. Songs that some listeners described as romantic were viewed by others as examples of symbolic violence or possessiveness, while lyrics considered overtly problematic by some groups were embraced as "anthems" by others.

One recurring example was "Eres Mía" by Romeo Santos. Some participants interpreted it as a romantic declaration, whereas others argued that the repeated phrase "you are mine" objectifies women and reinforces the idea of ownership.

A similar pattern emerged with "Mátalas" by Alejandro Fernández. Although part of the research team initially considered the song clearly misogynistic, several evaluators—particularly men—did not perceive any problematic elements in its lyrics.

These contrasting interpretations highlighted how deeply musical meaning depends on cultural and generational perspectives.

The study also found that the most harmful songs are not always the most explicit. During an analysis of homophobia in music, the researchers discovered that openly offensive language was often easier to recognize than subtle messages that had become socially normalized.

Alcántara pointed to "El Gran Varón" by Willie Colón as an example. While many listeners outside the LGBTQ+ community saw nothing problematic in the lyric "A crooked tree never straightens its branches," members of the community explained that associating sexual diversity with something "crooked" carries a significant discriminatory message.

These findings underscore a central challenge: AI systems cannot automatically understand social context or cultural nuance. Teaching an algorithm to recognize concepts such as misogyny, sexism, or gender stereotypes requires carefully labeled examples created by human evaluators.

As Alcántara explained, supervised learning depends entirely on the quality of the training data.

"If the data are labeled incorrectly, the model will learn those mistakes. If the dataset contains bias, the algorithm will also learn that bias."

Songs can therefore be classified into categories such as misogynistic, non-misogynistic, psychological violence, hate speech, or gender stereotypes.

Which Musical Genres Show More Violence?

The project currently employs Large Language Models (LLMs) alongside classical machine learning, deep learning, and fine-tuning techniques to analyze thousands of songs.

The results indicate that contextual AI models—which are better able to interpret surrounding information and prior examples—perform more accurately on these complex classification tasks.

Among the patterns identified, reggaeton frequently displayed higher levels of female sexualization and objectification, often portraying women as objects of desire or consumption.

Regional Mexican music commonly reflected themes of possession and control, with recurring expressions such as "you're mine," "you're going to marry me," or "you should stay at home."

Perhaps the most surprising findings emerged from pop music and romantic ballads. Many songs traditionally regarded as timeless love classics contain messages involving emotional manipulation, dependency, or resentment toward women.

The researchers also observed that songs with the strongest misogynistic content generally feature shorter, more repetitive lyrics, whereas longer compositions tend to contain fewer misogynistic elements.

Listening Critically

Despite these findings, Alcántara does not advocate banning any musical genre.

Instead, she believes listeners should develop a more critical awareness of the messages conveyed through music and avoid normalizing violent or discriminatory narratives.

"You may enjoy a song because of its rhythm or melody, but it's also important to recognize the message it communicates."

The broader goal of the research is to encourage discussion about music's cultural influence while demonstrating how AI can serve not only as a technological tool, but also as a resource for social awareness.

One potential application, she explained, would be incorporating misogyny-detection algorithms into platforms such as YouTube to strengthen parental controls by helping identify songs that may be inappropriate for younger audiences.

The datasets created by the research team have already been shared with the international scientific community through collaborative competitions organized by IberLEF, an initiative of the Spanish Society for Natural Language Processing (SEPLN).

The project has since expanded into two major initiatives: Misogyny 2025, which focuses on songs from Mexico and Latin America, and Misogyny 2026, which incorporates Spanish music and examines linguistic and cultural differences between both regions.

Looking ahead, the researchers plan to move beyond lyrics by studying how rhythm and music affect listeners' brain activity through electroencephalograms, in collaboration with the Polytechnic University of Valencia.

Ultimately, the research extends beyond artificial intelligence itself. Its purpose is to better understand how music shapes social attitudes and how AI can help identify harmful messages that often go unnoticed.

As Alcántara concluded: "Sometimes we don't even realize what we're listening to."