AI mastering vs a human engineer: what actually differs

Let me be fair to the machines first, because a producer arguing against automation sounds exactly like a producer protecting his income. AI mastering is genuinely good at what it does. Upload a mix, wait thirty seconds, get back something louder, brighter and more even that will not embarrass you on Spotify. For a demo, a reference bounce, a track you are releasing to a mailing list of eighty people, or a single you need up by Friday, it is the right tool and paying a person would be a waste of your money.

So the question is not whether AI mastering works. It is what it cannot do, and whether that thing matters for the record you are making.

What an algorithm is actually doing

An AI master analyses your file, compares its frequency balance and dynamics against a large set of commercial releases, and moves yours towards the average of that set. That is the whole trick, and it is a good trick. It catches a boxy low-mid, it tames a harsh top, it gets you to a competitive loudness without obvious damage.

But notice what that means. It is pulling your record towards the middle. If your mix is unremarkable, the middle is an improvement. If your mix has something deliberately odd about it — a vocal further back than is fashionable, a drum sound that is supposed to be ugly, a chorus that opens up because the verse was small on purpose — the algorithm has no way to know the difference between a choice and a mistake. It sands both down.

The difference in one sentence

A human engineer asks what the song is supposed to feel like, and an algorithm cannot ask. That sounds like a slogan, so here is what it means on a real day of work.

• You say the chorus needs to feel like the roof coming off. That changes how the verse is handled, not just the chorus — the lift has to be built out of the thing before it.
• You say the record is meant to sound close and a bit claustrophobic. Half the usual moves are now wrong, and an AI will apply them anyway because they are what most records want.
• You say track four is the emotional centre of the EP. It gets treated differently from the singles either side, so the sequence has a shape.
• You say the low end is scaring you. A person tells you whether the problem is the master, the mix, or the bass part itself — and the last one is very common, and no mastering of any kind will fix it.

What a human catches that software does not

Most of the value is not in the processing at all. It is in someone listening to your record properly, at length, on systems that do not flatter it, and telling you the truth. Half the masters I have been sent to 'fix' did not need mastering: they needed the vocal automated four decibels in two places, or a bass part re-recorded, or one section deleted. Software will happily master a problem to a professional loudness and hand it back to you as a finished product.

The other half is context: how the tracks sit as a record, whether the quiet one needs to be quieter, what happens at the joins. An album is a sequence, not eight independent files, and AI mastering treats it as eight independent files.

So when should you use AI mastering?

Genuinely, use it when:

• It is a demo, a reference, or a rough for the band to argue over.
• You are releasing at volume — a track a month — and consistency matters more than depth.
• The budget is zero and the alternative is not releasing.
• You want to hear roughly what a loud version of your mix sounds like before committing.

And use a person when:

• This is the single, the one you are putting money behind.
• The record has a deliberate character you are worried about losing.
• It is an EP or album that needs to hold together as one thing.
• You suspect something is wrong but cannot name it. That is the most valuable hour you can buy, and no algorithm sells it.

The dishonest comparison to avoid

Every demonstration of AI versus human mastering you will find online, including the ones made by studios like mine, is rigged by level. Whichever version is louder sounds better for the first ten seconds, and that is simply how hearing works. If you are comparing your own options, match the two versions to the same loudness before you decide. Then listen to whether the vocal still means what it meant, and whether the chorus still moves.

If it does, use the cheap one. That is not a trick answer — a good enough master on a released song beats a perfect master on a song you never put out.

← All notes