How AI mastering works, and where it stops working

Upload a mix, wait a minute, download something louder and shinier. That is the promise of AI mastering, and for a lot of independent releases it is now the default last step. But "AI" hides a fairly understandable process, and knowing what that process is lets you get better results and spot when the machine has made the wrong call.

This guide is about what happens after you press the button. If you want the before, our guide on prepping stems and mixes for AI mixing and mastering covers what services ask you to upload. If you want to do it yourself, start with mastering at home.

The basic loop: measure, compare, correct

Almost every automated mastering tool, whether a web service or a plug-in assistant, does some version of the same four things.

1. Measure the mix

The tool analyses your file: its spectrum (how much energy sits in each frequency band), its dynamics (how far apart the loud and quiet moments are), its stereo image, and its loudness. Loudness is not a vague impression here. It is measured with the method in ITU-R BS.1770, which defines integrated loudness in LUFS and true-peak level. That same standard sits under the meters streaming services use, which is why a mastering tool can talk sensibly about "streaming loudness".

2. Classify and pick a target

Next, the tool decides what kind of record it is listening to, or asks you. Some services infer genre or character from the audio; others let you choose a style or intensity. iZotope's Master Assistant documentation describes the plug-in version of this step: it listens to the track, takes a genre target you pick, or a reference track you upload, and builds a target from it.

The target is the key idea. It is a description of where a finished record of this kind usually lands: a tonal balance curve, a loudness, a dynamic range. The "intelligence" in AI mastering is mostly in how good that target is and how well the tool matches your track to the right one.

3. Set the processing

With a measurement and a target, the tool works out the difference and chooses settings to close it. In practice that means the same four moves a human engineer uses: corrective EQ to nudge the tonal balance, gentle compression, sometimes saturation or stereo adjustment, and a limiter to reach a loudness without clipping. Nothing here is exotic. The chain is ordinary; only the choice of settings is automated.

4. Render and report

The service renders the master, usually with a preview and a few intensity options. Plug-in assistants instead leave the chain open in front of you, and iZotope's guide is explicit that the result is a starting point to adjust rather than a final answer. That difference matters more than any brand: one gives you a file, the other gives you decisions you can change.

What the loudness number really means

Many people judge an AI master by how loud it is. That is the wrong test for streaming. Spotify's own loudness normalisation page says it adjusts playback to about -14 LUFS by default, so a master pushed far louder is simply turned down. In our view, once normalisation is on, heavy limiting buys you no extra loudness and usually costs you punch. For a side-by-side of the products themselves, see our take on AI mastering, honestly.

So when a service offers "low, medium, high" intensity, treat "high" as a creative choice about density and grit, not a way to be heard. Our streaming loudness targets guide and limiting and true peak explainer go deeper on the numbers.

The limits: where automated mastering goes wrong

Knowing the loop makes the failure modes predictable.

• It cannot fix the mix. Mastering processes the whole stereo file. If the vocal is buried or the kick and bass fight, broad EQ on everything cannot pull one element out without changing the rest. That is why the services' own prep advice centres on the mix you send: the input decides most of the outcome.
• It matches averages, not intentions. A target is built from what records of a type usually sound like. A deliberately dark, lo-fi or very dynamic track will be "corrected" toward the middle, because the tool cannot know the darkness was the point.
• Genre guesses can be wrong. If the classification step mistakes a hybrid track for something else, every later setting follows the wrong target. If a tool lets you pick the style or give a reference, use it.
• It hears one file at a time. An album or EP needs songs that sit together in level and tone. Mastering each track separately, by any method, can leave a quiet ballad next to an over-pushed single. Check the sequence as a listener would.
• It does not do delivery specials. Vinyl cutting, Dolby Atmos and stem mastering have their own requirements. See our guide to stem mastering.

A practical way to use it

1. Export a clean, unlimited mix with headroom, as the service asks.
2. If you can pick a style or upload a reference, do it. A reference you chose beats a genre the tool guessed.
3. Compare the master with your mix at matched loudness. Turn the master down until both feel equally loud, then listen for what changed. Louder always sounds "better" for a few seconds.
4. Listen on a phone, earbuds and a car or laptop, not only on your monitors.
5. If the low end or vocal moved in a way you dislike, go back and fix the mix, then re-master. Don't stack a second mastering pass on top.
6. Keep your unmastered mix. You may want a human or a different tool later.

What we think

AI mastering is good at something real: getting a reasonable, translatable, correctly measured master quickly and cheaply. For demos, singles on a tight budget and checking how a mix will hold up, that is useful, and we would rather an artist release with a sensible automated master than sit on a song for months.

But we think the honest way to describe these tools is "automated averaging", and artists should treat them that way. The value of a mastering engineer was never only the EQ curve. It was a second pair of ears asking what you meant, and then protecting that. A target curve cannot ask. So the artist has to bring the intent: choose the reference, judge at matched loudness, and refuse a result that makes your record sound more like everyone else's. Used like that, the tool works for you. Used as a final verdict, it quietly moves your sound toward the middle of the genre, one release at a time.