Upload a stereo mix to LANDR or run iZotope's Master Assistant, wait half a minute, and a finished master comes back. It often sounds competitive. That result is easy to mistake for judgement, as if the tool listened to your track and decided what it needed. It did not. Understanding what actually happens in those thirty seconds is the difference between using AI mastering well and letting it quietly flatten your record.
Three stages, only one of them is "AI"
An AI mastering pass is three steps stacked on top of each other: analysis, target construction, and processing.
Analysis measures the uploaded mix. The tool reads frequency balance, dynamic range, stereo image and loudness. This is ordinary signal measurement, no machine learning required.
Target construction is the step that genuinely needs a trained model. The tool has to answer one question: what does a released, mastered track in this style usually look like? iZotope is unusually transparent about how this works. Master Assistant compares your track against ten genre class curves, each built from analysis of chart-topping songs in that style. It assigns every curve a percentage of similarity to your audio, then blends those percentages into a single custom target curve. Your track is rarely "100 percent hip hop." It might come back as a weighted mix of several classes, and the target is the blend.
LANDR markets its version as "micro-genre" classification, a model trained on a large corpus of mastered tracks. The branding differs, the principle does not. The model recognises a style and pulls up the statistical profile associated with it.
Processing moves your measured input toward that target. In Ozone, the equalizer compares your input curve to the generated target curve and places EQ bands wherever the two diverge. A dynamics stage nudges your dynamic range toward the target's range, a width stage does the same for the stereo image, and a limiter raises the whole thing to the loudness target.
So the "intelligence" in AI mastering is genre recognition. The mastering itself is conventional DSP: EQ, compression, stereo processing and a limiter, aimed at a number instead of an ear.
Where this design quietly hurts you
Once you see the mechanism, the failure modes are predictable.
It cannot diagnose a mix. A human engineer listens and says the vocal is harsh around 3 kHz, or the kick and bass are fighting for the same 80 Hz. AI does none of that. It matches a global tonal curve. A broad EQ move toward an average spectrum cannot un-mask a buried vocal, and it cannot hear a phase problem between two tracks. Whatever problems you upload, the tool processes them as faithfully as it processes the good parts. If the low end is muddy going in, it is muddy going out, now sitting under a target curve and a limiter.
The limiter is where most damage happens. To hit a loudness target, the limiter applies whatever gain reduction is needed. On a mix that is already loud or already compressed, that stacks compression on top of compression. The result is the familiar squashed sound: pumping, harsh sibilance, a master that fatigues fast. The tool is not being careless. It is doing exactly what you asked, which is to reach a density target, and your mix simply had no room left to give.
It pulls toward the average. This is the part worth sitting with. The target is a statistical profile of released tracks in a genre. If your mix is deliberately unconventional, a wide ambient piece, a track built around restraint, an arrangement with intentional space, the tool reads that as deviation from the norm and corrects it. A polished, distinctive mix can come back sounding worse, because "average" was never the goal and the tool has no way to know that.
The loudness target is mostly a myth
A lot of anxiety around mastering targets one number: -14 LUFS. It is worth retiring that habit.
Streaming platforms normalize playback. Spotify, YouTube, Tidal and Amazon play back around -14 LUFS integrated. Apple Music sits near -16, Deezer near -15, and Pandora does not use LUFS for this at all. If your master is louder than the platform reference, it gets turned down on playback. If it is quieter, it gets turned up.
The consequence: after normalization, a dynamic master and a squashed master play at the same loudness. The squashed one gained nothing for all that limiting. It just lost its transients. Chasing maximum loudness for streaming is chasing a number that the platform cancels out.
A more useful sanity check than a LUFS target is PLR, the peak-to-loudness ratio, which describes how much dynamic room a master keeps. Well-mastered streaming tracks commonly land roughly in the 8 to 14 range. That is a guideline, not a rule, but it is a better thing to watch than a single integrated loudness figure. Master for dynamics and headroom, leave a true-peak ceiling around -1 dBTP, and let normalization handle the level.
Using AI mastering without getting burned
None of this means AI mastering is bad. It means it is a tool with a known shape.
Use it as a fast reference. Master a rough version to hear your mix at competitive loudness, then decide whether the mix itself needs work before you commit to anything.
Fix the mix first. The single most reliable way to get a good AI master is to feed it a good mix. The tool repairs nothing, so do the repairs yourself.
Watch the limiter. If the AI master sounds flat, pumping or fatiguing, the loudness target is too aggressive for your mix. In tools that allow it, lower the target loudness or raise the dynamics setting.
Treat Master Assistant as a learning tool. Solo and bypass each module it builds. Seeing which EQ moves and how much compression your genre target asks for is genuinely instructive, whether or not you keep the result.
On the human-versus-AI question, the honest answer is that most working artists use both. Blind comparisons are commonly reported to favour human masters as less flat and less compressed, though a controlled published study is hard to point to, so treat that as a widely held impression rather than proven fact. The practical split is about stakes. AI mastering is fast and cheap, well suited to demos, reference checks, social uploads and a steady stream of releases. A human engineer costs more and is worth it for the records that matter: albums, vinyl, the single you are pushing hard. The difference you are paying for is not better EQ. It is someone who can listen to your mix and tell you what is wrong with it.