AI Lyric Generators: The Weakest Link in AI music Products
Ezra Sandzer-Bell- Feb 19
- 6 min read
Updated: Feb 28
Lyric generators are one of the internet’s favorite past times. I can remember experimenting with them as a kid in the 90’s, decades before artificial intelligence entered the picture.
In the late 2010's, sites like These Lyrics Do Not Exist made some improvements to the fill-in-the-blank and paint-by-numbers approach. Then, as large language models (LLMs) went mainstream in the early 2020s, AI copywriting tools like Jarvis began rebranding their web apps for lyric generation use cases.
OpenAI released ChatGPT in December 2022 and social media influencers started using the chatbot to create music, in combination with text-to-speech sites like Uberduck. Most of those songs were parodies based on celebrity deepfakes.
A few more years have passed and fine-tuned ChatGPT plugins like Lyric Writer have surfaced with a more "on the rails" user experience.
Nevertheless, as of 2025, AI lyric generators still lack the emotional resonance or "soul" that fans expect from their favorite artists. In this article we'll explore why that might be the case.
Table of Contents
The Anthropic case: AI lyrics and copyright law
There's a major problem with AI lyric generators that could prove difficult to resolve. It's tempting to think that the problem is technical or even spiritual. People often say that machines could never understand human emotion enough to offer real lyrical depth.
That may or may not be the case, but there's evidence suggesting that data and copyright law are a major roadblock to improving on this issue.
Large language models rely on good data to produce good outputs. That's why big tech companies like OpenAI and Anthropic scraped so much of the internet to train their models. The problem is that they did this without paying any licenses.
Music publishers Universal Music, ABKCO and Concord Publishing sued Anthropic in October 2023 over what it claimed were copyright infringing outputs. That lawsuit reached a head in January 2025 when the courts ruled that Anthropic needed better guardrails to prevent infringing outputs.
Some are interpreting this as a bad sign for the music industry. They were hoping to win the case against unlicensed inputs during the model training phase itself. But the courts seem to be siding with the transformative fair use argument.
At the end of February 2025, Amazon announced that Alexa+ would include an integration with Suno. The AI songs seem to be here to stay.
AI song generators Suno & Udio default to ChatGPT
Let's rewind a bit. Major AI song generators like Suno and Udio shocked the internet with unprecedented, high-quality compositions in 2024. Their models seemed to continuously improve as new features rolled out.
Yet for some reason, the AI lyric generation on both services remained stuck in an extremely formulaic state. They spit out four bar end-rhyme schemes and cliches that feel more like nursery rhymes than anything you'd hear in modern music.
Take the example below, where Suno was prompted for trap lyrics and produced "Elixir of the Streets" with the opening line "streets hit harder, venom in veins flows steady". It's difficult to put into words how totally off base these lyrics are from the actual genre they're trying to imitate.
Users speculate that these companies can't improve on the lyrical output. The most compelling explanation boils down to copyright lawsuits. They don't want to give the record industry additional legal ammunition against them in the courts.
Suno and Udio have both been sued by the RIAA and GEMA for musical copyright infringement. Kneecapping their lyric output and sticking to low quality ChatGPT outputs could be part of a legal strategy to deflect any copyright infringement back up the chain to OpenAI.
New data cloud compares AI lyrics to rap lyrics
Near the end of February 2024, a data scientist took to Reddit and released an AI lyric word cloud comparing 60,000 AI generated songs from Suno with 40,000 lyric charts from the "lyric interpretation" website Genius.
That word cloud graphic is shown below, with the most commonly used words from AI lyrics shown on the left and human-made rap lyrics from Genius on the right. You can find the Suno dataset here and the Genius lyric dataset here.
This graphic made the rounds on Twitter here and several users weighed in to confirm that the left column reflected very specific, overused words that surface in their own Suno, Udio and ChatGPT experiments.
Words and phrases like whispers of, neon lights, gentle, heart, softly, and laughter seem ordinary enough on their own. But power users on any of these platforms instantly recognize them. They are overused to an extreme degree in both song titles and lyrics.
The right half of the image slants heavily toward slang like somethin, gang, huh, ayy, and the profanity that rappers actually use in their lyrics.
The "best AI song lyric generators" are so bad it hurts
Our staff is generally bullish on generative AI music software. We tend to side with companies that train their models on licensed data, but we also understand that the state of the art models did not follow those guidelines. That being said, we are not inherently against the use of artificial intelligence in creative music workflows.
Nevertheless, Google's results for the "best AI lyric generators" are pure slop. My brief experiments with each tool led me to believe that all of them were using ChatGPT on the backend. In other words, they're just different outfits for the same underlying tech.
The problem is not limited to the four bar phrasing and ChatGPT AI lyric cliches about "gentle whispers" and "grateful hearts". There's a serious culture problem with each of these generative AI tools as well.
The example below is from Canva's AI song lyric generator, where the placeholder text reads "generate rap lyrics for a hip hop track about climate change".
Could it be any more obvious that the political ideologies of the company employees are being imposed on the musical genres they're creating lyrics for?
OpenAI has heavily censored their model to favor non-offensive language and non-infringing outputs, which ends up being fundamentally at odds with trends in popular music.
This second example from a site called AI Song Lyrics Generator falls into the same trap, with lyrics like "Got a story in your mind but the words are hard to find? Let AI be your guide with creativity by your side". This is the nursery rhyme effect that I mentioned earlier in this article. Users find themselves in a kind of upside down world where the highly sanitized and maximally non-offensive lyrics take on an almost offensive quality. They are so bad that they feel almost like an insult to everyone's intelligence.
Culture rot: The real-world impact of AI lyric slop
There's an argument in favor of AI lyric generators, that says they can be a useful songwriting aid. I think that's perfectly reasonable if it's true for an individual, but the real problem begins with AI song generators are used by bad actors to spam DSPs like Spotify and YouTube with music that lacks human touch.
During the second half of 2024 and into 2025, I began personally encountering AI generated songs in cafes, restaurants and even hotels. After noticing this trend for a while, I started pulling out my phone and compiling video evidence.
The video below highlights half a dozen different locations where this took place:
When I've used Shazam to identify the tracks, they were either not detected or the song credits cited a single name for the composer, engineer and label. Those songs had publishing dates no earlier than Fall 2024 and the "artists" were ghosts - no sign of them anywhere on the internet.
I've asked the baristas and front desk personnel to show me their phone. Every time, the music is being streamed from YouTube playlists with generic titles like "Cozy Cafe | Chill Lofi Music Mix".
What happens when raw AI songs, spit out by a model with zero human touch, become a common feature in the places where we go to socialize?
This will have an impact on culture - it's what media theorist Douglas Rushkoff calls a regression to the mean. The four bar end-rhyme schemes, the false sincerity of the AI vocals, the GPT-lyrics. Everything about the songs are statistically average.
In my opinion, one of the greatest dangers posed by AI song generators is that we end up filling public spaces with low quality music and losing the connection with genuine artistry.
