The short answer

Researchers generated 100 tracks for each system and genre across Afrobeats, K-pop, Dance Pop and Heavy Metal, then compared them with equally sized human-produced reference groups using 72 computational audio features. Their results showed two different kinds of flattening: Lyria 3 produced less variation within genres, while Suno showed weaker acoustic separation between genres.

That does not mean every Suno song sounds alike or that Lyria cannot make distinctive music. It means creators should treat model defaults as a real creative force — one that can be stronger than the prompt alone.

What the new research actually found

The August 2026 paper The Algorithmic Flattening of Sound audited Suno and Lyria 3 across four genres. For each system and genre, the researchers generated 100 tracks and compared them with human music corpora of the same size. They measured rhythm and timing, timbre and spectral shape, dynamics and other acoustic characteristics.

The important part is that the two systems did not flatten music in the same way.

The researchers also ran minimal genre-name prompts to expose each model's default tendencies. Their conclusion was especially relevant for creators: prompt-following alone did not fully explain the flattening. That suggests some of what you hear comes from the model's learned priors rather than your instructions.

What “homogenization” means in plain English

When creators say AI music “sounds the same,” they often mean something vague: familiar chord movement, similar vocal phrasing, predictable drops, glossy mixes or recurring instrument choices.

The study asks a narrower question: does generated music show less acoustic diversity or weaker genre separation than human-produced reference music?

That distinction matters. You can have two tracks with different melodies and lyrics that still share the same underlying model fingerprint: similar dynamics, spectral balance, rhythmic feel, vocal presentation or arrangement logic.

How model defaults sneak into your music

Every generative model has tendencies. If your direction is weak, those tendencies make more of the decisions for you.

The danger is not using defaults. Defaults can help. The danger is mistaking the model's taste for your own sound.

Use the Model Default Test before you build the final song

This is a simple way to hear what your chosen system tends to supply automatically.

The point is not to “beat the model.” The point is to identify where the model is making creative decisions you did not consciously choose.

How to keep your AI music from collapsing into the average

1. Make fewer vague genre decisions

“Reggae,” “trap,” “Afrobeats” or “cinematic” tells the model a category. It does not tell the model what makes your version of that category distinctive. Define tempo feel, drum behaviour, instrumentation, arrangement density, vocal character and emotional tension.

2. Choose reference qualities, not just reference labels

Instead of chasing a named artist, define the qualities you actually want: dry drums, sparse bass movement, live-room percussion, restrained vocal reverb, unstable synth texture, abrupt transitions or a specific energy curve.

3. Preserve unusual decisions

If a generation gives you one genuinely strange and useful moment, do not automatically regenerate the entire track. Protect the part that gives the song identity.

4. Compare versions instead of assuming newer is better

Generation tools make it easy to confuse iteration with improvement. A cleaner result can also be a more generic result.

5. Add human-controlled elements where they matter most

MIDI, live instruments, edited drums, custom vocal performance, arrangement changes, manual effects and DAW-level restructuring can interrupt repeated model habits.

6. Build a repeatable sound identity

Your sound should survive changes in model version. If every release depends on whatever the generator happens to prefer that month, you do not yet control the identity strongly enough.

A better workflow: direct → compare → intervene → preserve

Direct: define the sound before generating.

Compare: judge each result against your target, not against the previous generation.

Intervene: change the specific areas where the model's defaults are taking over.

Preserve: keep the unusual details that make the song recognizable as yours.

This is the same reason I treat “finding your sound” as a creation system instead of a better-prompt contest.

Continue to Find Your Sound →

One more creator lesson: AI detection and sameness are not the same thing

The study also found that a standard classifier could distinguish AI-generated tracks from human tracks with very high accuracy using the measured audio features. That does not mean listeners can always identify AI music by ear, nor does it prove there is one universal “AI sound.” In fact, the researchers found Suno and Lyria were acoustically distinct from one another.

So the useful creator question is not “How do I hide that AI was involved?” It is “How much of this sound did I actually choose?”

Related Jack Righteous guide

If you want the broader human-plus-AI workflow, continue here:

Human + AI Music: The Jack Righteous Creation Workflow →

Frequently asked questions

Primary source

Slendebroek & Metaxa — The Algorithmic Flattening of Sound: Computational Evidence and Justice Implications of AI Music Homogenization

The paper was submitted August 6, 2026 and is listed by arXiv as forthcoming in AAAI/ACM AIES 2026. This article translates its findings into practical creator workflow guidance; it does not claim the study applies to every model, genre or generated song.