What you’ll be able to do

Right, so here’s the actual state of play. You can write lyrics with ChatGPT or Claude, generate a full track with vocals and music using something like Suno or Udio, and if you want a specific voice or sound effects layered on, ElevenLabs does that bit. Put those together and you’ve got a song, start to finish, without playing an instrument or knowing what a key signature is.

That’s not nothing. A year or two ago this was clunky and the vocals sounded like a fax machine having a breakdown. Now you can type a mood, a genre and a handful of lyrics, and get something back that sounds like a real song. Not always a good song. But a real one, structurally speaking, verses and chorus and everything.

What you can’t do, or shouldn’t try to do, is get it to sound exactly like a specific famous singer. More on that in a second.

Before you start

Before you generate anything you plan to actually release or sell, sort out what you’re allowed to do with it. This bit’s genuinely important and it’s the bit people skip.

Every tool has its own terms about what you can do with the song you make. Some let you use generated tracks freely on a free plan, some only unlock commercial use if you’re paying. I’m not going to guess which applies to you, because it changes and because getting it wrong matters. Check the terms of whichever tool you use before you release anything or try to make money from it.

Don’t try to make the AI imitate a named living artist’s voice or style so closely it’s basically them. Aside from it being a bit grim, ethically, it can also land you in trouble depending on the platform.

And if you’re planning to put the song on Spotify, Apple Music, or anywhere similar, or you’re a member of PRS or a similar rights organisation, check their policies on AI generated music first. These policies exist and they do get updated, so don’t assume what was true six months ago still is. Five minutes of reading now saves a headache later.

Get set up

You’ll need a couple of accounts, depending on what you’re making.

For lyrics, you can use ChatGPT, Claude, or Gemini. Any of them will do. You don’t need a paid plan just to write lyrics, the free tiers handle that fine.

For the actual track, the two names that come up again and again are Suno and Udio. Both take a text prompt, or lyrics you’ve written yourself, and generate a full song: music and vocals together. You sign up, you get some free generations to try it out, and if you like it enough to use it properly you can look at their paid tiers, which is usually also where commercial rights kick in (again: check the terms for whichever one you pick, this changes).

If you want a specific voice doing spoken parts, or you need sound effects woven in, that’s where ElevenLabs comes in. It’s a different kind of tool, more about voice generation and audio than songwriting, but it’s worth knowing about if your song needs something Suno or Udio can’t quite give you.

So the basic setup is: an assistant for words, a generator for the track, and optionally ElevenLabs for extras. You don’t need all three every time. Loads of people just go straight into Suno with a rough idea and let it write the lyrics too.

Try it yourself

Here are three prompts, each doing something different. Copy them, tweak the details, see what comes back.

First, writing the lyrics themselves. This one’s for ChatGPT or Claude.

Write song lyrics about missing a friend who's moved abroad.
Make it wistful but not miserable. Structure it clearly with
labelled sections: Verse 1, Chorus, Verse 2, Chorus, Bridge,
Chorus. Keep the chorus short and repeatable. Aim for around
250 words total.

Second, generating the actual track. This one goes into Suno or Udio, using either the lyrics you just got or your own.

Genre: indie folk, acoustic guitar led
Mood: warm, a little melancholy, hopeful by the end
Tempo: mid-tempo, around 90 bpm
Instruments: acoustic guitar, light strings, soft drums entering
in the chorus
Vocals: a single voice, gentle, not overproduced

[paste your lyrics here, with Verse 1 / Chorus / Verse 2 etc
clearly marked]

Third, refining something you’ve already got. Say the first version came back too polished, too pop, when you wanted it rougher.

I've got a song that came out too clean and radio-friendly.
I wanted it to sound more like a demo, something recorded in
a bedroom, a bit imperfect. Suggest changes to the mood and
instrument description I should feed back into the generator
to get that rougher, more intimate feel. Keep the same genre
and lyrics, just adjust the production style.

That third one you’d actually put to ChatGPT or Claude first, to help you rewrite your prompt, then take the improved prompt back into Suno or Udio. Bit of a loop. Works well though.

A note on the second prompt: marking the structure (verse, chorus, bridge) genuinely helps the generator. Without it, you often get something that meanders, or repeats the same section three times because it’s lost track of what it’s building. Labelling it costs you nothing and makes a real difference.

Check the result

Once you’ve got a track back, don’t just accept it because it exists. Actually listen properly and ask a few things.

Does the mood match what you wrote in the prompt? If you asked for melancholy and got something that sounds like a car advert, that’s a sign to go back and be more specific, maybe drop the tempo or change the instruments.

Does the structure hold together? Listen for whether the chorus actually repeats in a recognisable way, whether there’s a clear shift into the bridge, whether it just sort of trails off at the end. Generators sometimes lose the thread on longer songs.

Do the lyrics actually sit on the melody properly, or does it sound like words crammed in where they don’t fit? This happens more than you’d think, especially with lines of uneven length.

And the honest one: would you want to send this to a friend? Not “is this technically a song” but would you actually share it. If the answer’s no, that’s useful information, not a failure. Go back, change one thing (the tempo, the mood word, the instrument list) and generate again. Most people who get a good result generated four or five versions first, not one.

If it doesn’t work

Sometimes you get gibberish vocals, or a track that ignores half your prompt, or something that just sounds wrong for no clear reason. A few things to try.

Simplify the prompt. If you’ve stacked in ten adjectives and three genres, the tool might be struggling to weight them. Strip it back to genre, mood, tempo, one or two instruments, and try again.

Check your lyrics are formatted clearly. Unlabelled walls of text confuse the structure. Go back to labelling verses and choruses properly.

If the vocals sound strange or off-key, that’s often just the tool having a bad generation. Regenerate. It’s not you.

Now, the bit worth being careful about. Don’t feed an AI tool anything private or sensitive as part of a song, real addresses, other people’s personal details, anything you wouldn’t want stored or reused. Same goes for anyone else’s copyrighted lyrics or melody, don’t paste in someone else’s song and ask for “something like this but different”, that’s a shortcut to a result you can’t legally use anyway. And don’t ask a tool to clone a specific real person’s singing voice. If a generation sounds oddly close to a known artist by accident, that’s a reason to change your prompt, not a bonus.

If you’ve tried adjusting the prompt a couple of times and it’s still not landing, it’s often faster to start a completely fresh prompt than to keep patching one that’s gone wrong. Sunk cost and all that.

Keep exploring

If you enjoyed the lyric-writing side of this, there’s a guide on this site called “How to get AI to write a poem” that goes deeper into that. If you’d rather curate than create, “How to get AI to make a playlist” covers using AI to pull together existing music instead of generating new tracks. And if your song needs a video to go with it, “How to get AI to make a video” picks up from where this one leaves off.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.