We Analyzed 6,798 AI Song Generations: 1 in 10 Prompts Copies a Real Artist’s Work

9.8% of AI song prompts tried to copy a real artist's style, voice or existing lyrics: about 1 in 10. That is the main finding from 6,798 song generations made by 5,280 people on OnMusic between June 13 and October 5, 2026.

• Song generations analyzed: 6,798 (internal and test accounts excluded)
• People: 5,280
• Prompts checked before generation: 6,505
• Prompts that tried to copy a real artist's style, voice or lyrics: 9.8%
• Songs made from the user's own lyrics: 31%
• Median time to a finished full-length song (Lyria 3 Pro): 41 seconds

Most people aren't trying to break rules. They describe the sound they want the fastest way they know: "a song like [famous artist]" or "in the voice of [singer]". Of the 6,505 prompts our pre-generation check reviewed:

• 5.0% included existing song lyrics
• 4.5% asked for a named artist's style
• 1.2% asked for a named artist's voice

Some prompts did more than one of these, so the categories add up to more than the 9.8% of prompts flagged at least once. Taken together, about 1 in 20 prompts named a real artist (5.0%).

Naming an artist doesn't help. Licensed AI music models are built to avoid imitating specific artists, so OnMusic rewrites these prompts into neutral descriptions (genre, tempo, mood and instruments) before generating. A prompt like "melancholic indie folk, fingerpicked acoustic guitar, soft female vocals, 90 BPM" gets you the feel without the risk. Our guide to whether AI music is legal explains why this matters if you plan to publish the song. New to AI music? Start with how AI music generation works.

We compared the four models with more than 300 generations each. "Completed" means the song finished and was playable. "Rejected" means the model provider declined the prompt.

• Lyria 3 Pro (Google): 4,597 generations · 80.6% completed · 8.6% rejected · median render 41s (90th percentile 76s) · average song 94s
• Lyria 3 Clip (Google): 1,023 generations · 84.1% completed · 10.9% rejected · median render 19s (90th percentile 28s) · average song 46s
• ACE-Step: 543 generations · 99.8% completed · 0% rejected · median render 17s (90th percentile 37s) · average song 80s
• CassetteAI: 398 generations · 100% completed · 0% rejected · median render 19s (90th percentile 30s) · average song 82s

What this means:

• Google's Lyria models make the most polished full songs, but they're the strictest. They rejected about 1 in 10 prompts, usually for naming real artists, quoting copyrighted lyrics or tripping safety filters.
• ACE-Step and CassetteAI almost never refused a prompt and were among the fastest, with median renders under 20 seconds. They suit quick instrumentals and background music.
• Longer songs take longer. Lyria 3 Pro songs are about twice as long as Lyria 3 Clip songs, and take about twice as long to render.

If a model rejects your prompt, remove artist names and quoted lyrics, describe the sound instead, or switch to a less restrictive model. You can try all of them in the AI song generator.

For a wider comparison of tools and pricing, see the best AI music generators of 2026, or compare Suno alternatives and Udio alternatives.

31% of songs were made from lyrics the user wrote, not from a text description alone. Writing your own lyrics avoids the main copyright risk. If you need a starting point, an AI lyrics generator can draft verses for you to edit, and lyrics to song turns finished lyrics into a full track.

• Source: OnMusic song generation records from June 13 to October 5, 2026, read once on October 5, 2026.
• Sample: 6,798 song generations from 5,280 people. Internal and test accounts were excluded (95 generations). The five most active accounts made 2.6% of all generations, so no single user skews the results.
• Prompt flags: before each generation, an automated check reviews the prompt and labels requests that name a real artist's style or voice, or that contain existing lyrics. 6,505 prompts went through this check. The labels come from an AI classifier, so a small number of false positives and misses is expected.
• Model reliability: 91 failures caused by our own systems (a file-retrieval bug) were excluded, because they say nothing about the models. Render time runs from request to finished song, including queue time. Only models with more than 300 generations are shown.
• Privacy: only aggregate counts were used. No prompt text, lyrics, names or account details were published or kept.

• One platform. This data comes from OnMusic users. People on other apps may write prompts differently.
• The flags are automated. The artist and lyrics labels come from an AI check, not a human review.
• Speeds change. Render times depend on provider load and will shift as models are updated.

Who is this for? Creators who want to know why an AI music tool rejected their prompt; journalists and researchers who need real-world data on how people use AI music generators; and anyone choosing a model who wants speed and reliability numbers from real use, not marketing claims. You may cite these figures as "OnMusic AI Song Generation Study (2026)" with a link to this page.