A fully AI-generated song has topped a Billboard chart. An AI metalcore act has pulled 85 million Spotify streams. A fake band fooled a million listeners before admitting no humans were involved. Whatever you think about it — and metal fans have opinions — the technology behind those numbers is not locked in a label's basement. It runs on your phone, and the interface is a text box.
This is a practical guide to making a metal song with it. Not a think piece about whether you should — we've written that, and you can read our honest forecast on where AI music is actually heading. This is the how.
Step 1: Pick your subgenre before you type anything
The single biggest mistake beginners make is prompting "epic metal song." Metal is not one sound. A generator fed a vague prompt returns the statistical average of all metal — which is to say, something no metal fan would play twice.
Decide what you actually want first. Thrash: fast tempos, palm-muted riffing, aggressive shouted vocals. Doom: slow, down-tuned, mournful, Sabbath-descended. Melodic death metal: Gothenburg-style twin guitar harmonies over harsh vocals. Metalcore: breakdowns, screamed verses, clean sung choruses. Djent: polyrhythmic, mathematically precise, eight-string low end. Power metal: soaring tenor vocals, gallop rhythms, fantasy grandeur. Symphonic: orchestra and choirs stacked over the riffs.
If you need a refresher on what defines each lane, our ranking of the 100 greatest metal bands doubles as a subgenre map — every entry names the style it defined.
Step 2: Choose your generator
Three tools dominate AI song generation in 2026. Suno is the market leader and the engine behind most of the AI acts you've read about, from the Velvet Sundown to Billboard-charting Xania Monet. Udio is its closest rival, famous as the tool behind "BBL Drizzy." And on mobile, apps like Sonx wrap the same text-to-song workflow into a phone app and add voice cloning and AI music videos on top — one prompt in, a finished song with visuals out. (Disclosure: Sonx is operated by our parent company, as noted on our about page. It's also genuinely what we used for the examples in this guide.)
The workflow below is tool-agnostic. Every serious generator takes a style description and lyrics, and returns a full mixed track in under two minutes.
Step 3: Write a prompt with four ingredients
Good metal prompts specify four things: subgenre, tempo, instrumentation, and vocal style. That's it. Here are three that work, taken from our own testing.
For melodic death metal: "Melodic death metal, 190 BPM, twin harmonized lead guitars, double-kick drums, growled harsh male vocals with a clean sung bridge, dark and triumphant, Gothenburg style."
For doom: "Doom metal, 65 BPM, massively down-tuned fuzzy guitars, slow crushing riffs, mournful clean baritone vocals, funeral atmosphere, long sustained notes."
For metalcore: "Modern metalcore, 150 BPM, djenty palm-muted riffs, a huge half-time breakdown, screamed verses and a soaring clean chorus, polished production."
Notice what these don't include: band names. Most generators block "sounds like Metallica" prompts outright, and the workaround — describing the sound instead of naming it — produces better results anyway. You know the vocabulary; you've been reading it on album reviews for years.
Step 4: The lyrics
You have two options: let the AI write them or write them yourself. Do the second one. AI-generated metal lyrics default to a word cloud of shadows, flames, and rising darkness — grammatically fine, memorably nothing. The AI acts that broke through treat lyrics as the human layer: Xania Monet's creator insists the words are entirely hers, with the AI handling only voice and production.
Metal gives you more lyrical room than any genre — war, myth, grief, politics, sci-fi, personal rage. Write three verses and a chorus, mark them with [Verse] and [Chorus] tags so the generator maps structure correctly, and put your hook in plain words. A chorus someone can shout at traffic beats one that merely sounds ominous.
Step 5: Generate, discard, regenerate
Your first generation will be mediocre. So will the next four. This is normal and it is the actual workflow: the creator of "BBL Drizzy" described his method as simply reiterating the track until it became something people would enjoy. Generate ten to twenty takes. Keep the one where the riff surprises you. Regenerate sections if your tool supports it — a great verse with a limp chorus is a regenerate, not a keeper.
A useful filter: play the best take next to a real song in your target subgenre. If the drop-off is embarrassing, the prompt is too vague — go back to step 3 and add specificity. The gap between a lazy AI metal track and a good one is almost entirely in the iteration count.
Step 6: If you release it, know the rules
You can upload AI-generated songs to Spotify and every major platform — none of them ban AI music. But three rules are non-negotiable. First, label it: platforms and distributors increasingly require AI disclosure, and audiences punish acts that hide it far harder than acts that admit it — ask the Velvet Sundown. Second, never clone a real artist's voice without permission; that is what got the most-streamed AI track in history wiped from every platform. Third, don't buy streams — the overwhelming majority of AI-music takedowns are for stream fraud, not for being AI.
What to actually expect
Honesty section. The AI acts with millions of streams are outliers riding playlist algorithms, novelty press cycles, or both — we ranked the biggest AI artists and the pattern across all fourteen is distribution, not sound quality. Your first AI metal song will most likely be heard by you, your group chat, and an algorithm deciding whether anyone else cares. That's the same deal every demo tape ever got.
Which is exactly why the barrier falling matters. The kid who couldn't afford a guitar, an interface, and a rehearsal space can now hear their riff idea out loud tonight, for free, before deciding whether it's worth becoming a musician over. Whether that's metal's death or its strangest new subgenre is being decided right now — by whoever's typing the prompts.
