Predicting the next Summer Hit? Kind Of. Unless it’s something new.
How AI Language Models use Linguistic Structures to Predict which Songs will Chart.
In 2015, data analyst Andrew Powell-Morse ran a simple test. He took 225 songs that had each spent at least three weeks at number one on the Billboard charts, across pop, country, rock, and R&B; and hip-hop, and ran their lyrics through the Flesch-Kincaid readability index, a formula originally built to grade the complexity of US Navy training manuals. The result surprised him: chart-topping songs averaged a second to third grade reading level, comparable to an eight or nine year old, and the trend had been getting simpler since 2006.
This was, statistically, one of the most consistent features of a hit.
His findings echo a larger field that treats song lyrics not as art to be appreciated but as data to be modeled. Long before a song reaches the radio, NLP systems are reading its words, scoring its structure, and estimating its odds of success, sometimes before a single human listener has had an emotional reaction on record.
Modern lyric analysis pulls from several layers of language at once. Repetition structure is one of the clearest signals. Researchers studying the UK Official Singles Chart between 1999 and 2013 built a multidimensional measure of lyrical typicality, capturing lexical repetition, complexity, thematic content, and emotional tone across 1,457 songs that reached the top five. They found that typicality (essentially how closely a song's language matched the established norms of its genre) significantly predicted how long a song stayed in the top five, even though it did not predict whether a song reached number one in the first place.
Beyond lyric typicality (word meaning/content), what are some other linguistic features that labels track in order to predict success?
Sentiment and emotional valence. Several research groups have built models that score lyrics for emotional trajectory across a song's runtime, mapping shifts from verse to chorus and identifying which emotional arcs correlate with replay value and chart longevity. A 2024 NLP study using DistilBERT, a compressed version of the BERT language model, was able to predict whether a song would succeed based on lyrics alone with 79% accuracy, a notable result given that the model had no access to melody, production, or artist reputation, only text.
Pronouns. It turns out, word choice at the smallest level matters more than most listeners would guess. Pronoun density, how often a song uses "you" versus "I" versus "we," correlates with genre and audience connection in ways labels track closely. Thematic modeling using techniques like Latent Dirichlet Allocation can sort songs into clusters based on the topics their lyrics gravitate toward, then map those clusters against historical chart performance to flag which themes are currently overperforming or fading.
Syntax. Structural features get equal attention. Songs are scanned for hook placement, how quickly the chorus arrives, how many distinct words appear across the full track relative to its length, and how much of the song is built from repeated phrases versus new ones. Earlier NLP work using support vector machines and decision trees was already showing in the late 2010s that lyric-only models could meaningfully predict whether a song would chart, without any audio features at all. The more recent transformer-based models have only sharpened that signal.
The uncomfortable implication is obvious, if disheartening: if labels can model - with real accuracy - which lyrical patterns correlate with commercial success, the incentive to write toward those patterns becomes very strong. Repetition gets rewarded. Simplicity gets rewarded. Familiar themes get rewarded over unfamiliar ones, because the model has no way to score a genuinely new idea except by comparing it to ideas that have already worked. This is the same homogenizing pressure that data-driven analysis exerts in other creative industries, just applied at the level of individual word choice instead of plot structure or pacing.
In other words, familiar language keeps a song around, (though it does not necessarily launch it). What does launch a song? That seems to be more complex.
Researchers studying lyric typicality have pointed out that NLP tools built on static word dictionaries struggle badly with slang, regional language, and words artists simply invent, the kind of linguistic risk-taking that sometimes defines an era. A model trained on yesterday's hits is, by construction, better at recognizing yesterday's patterns than at recognizing the next one. The algorithm can tell you with real confidence what already worked. What it cannot do is tell you what has never been tried before, which is usually the thing that ends up mattering most.
