Skip to content
HOW IT WORKS

AI Music Artifacts: What A Cleanup Pass Can And Cannot Fix

Your track came out of the generator with a glassy, metallic top end and drums that click instead of snap, and no amount of re-prompting shifts it. That is not your export settings. AI music artifacts are produced by the generator's own architecture: Deezer's research team traced the tell-tale spectral peaks to the deconvolution modules inside the decoder, and showed that a criterion built on those peaks alone detects AI-generated music with accuracy surpassing 99% in several scenarios. A cleanup pass works on that texture — and on nothing else, which is the half nobody writes down. We read both relevant papers, DistroKid's AI-upload rules, YouTube's Gen AI metadata field and Suno's terms at source on 4 November 2026. Here is what changes at each stage, the three gates a release has to clear, and which one audio processing reaches.

Architectural, not your file3 gates, 1 reachableSources read 4 Nov 2026No label is changed by audio
KEY TAKEAWAYS

The short version

Four findings, each from a document or a dataset you can open yourself.

01

The short version: the artifacts are architectural. Deezer's Fourier analysis attributes the spectral peaks in AI music to deconvolution modules in the generator's decoder — the model's architecture, not its training data or weights, validated on Suno and Udio output.

02

Older detectors fall over on a speed change. On speed-modified Suno v5 the Fourier criterion scores 0.745 AUC and 0.008 F1, while July 2026's scale-invariant model holds 0.986 and reaches 0.997 trained on the attack.

03

What a cleanup chain does is untested. That paper's limitations say “Other common manipulations (equalization, dynamic range compression, codec compression) remain untested” — nobody has published what EQ and compression do to a detector's score.

04

Processing reaches one of three gates. DistroKid's four AI conditions are all rights rules, and YouTube's AI designation is a delivered metadata value — all, partly, none or unknown. Only automated screening sees your audio.

THE DEFINITION

What an AI music artifact actually is

Start with what it is not, because the word “artifact” is borrowed from audio repair and brings the wrong mental model with it. In repair work an artifact is damage added to a real recording — hum, clicks, a dropout, tape hiss — and the standard technique is to profile a quiet passage so the tool learns what the noise sounds like alone. A generated track gives you neither: no clean original underneath, and no silent passage to profile, because what you want to remove was never separate from the music. It was produced in the same pass.

So an AI music artifact is better described as a signature of how the audio was made. The five families below are the ones creators actually complain about, and they do not share a cause: two are properties of the generator, one is an arrangement problem wearing a signal-processing costume, and one is something you did to your own file on the way out. Treating them as one thing is why so many cleanup attempts either do nothing or flatten the track.

TABLE 1 — FIVE ARTIFACT FAMILIES, AND WHAT EACH ONE RESPONDS TO
FamilyWhat you hearWhere it comes fromWhat reduces it
Spectral peaks (“shimmer”, metallic high end)A thin, glassy top end, worst on earbuds and cymbalsDeconvolution in the generator's decoder, per Deezer's Fourier analysisNarrow, targeted reduction in the affected bands
Smeared or brittle transientsDrum hits that click instead of snap, or arrive soft and bluntReported consistently by ear; no published measurement we will citeTransient shaping and dynamic high-frequency control, usually per stem
Too-perfect repetitionLoops that recur with no micro-variation, so the ear reads them as machine-madeHow the material was generated, not damage added afterwardsAlmost nothing, in processing. It is an arrangement problem
Vocal roughness and sibilanceHarsh “s” sounds and a papery edge on held notesReported by ear; the voice is where listeners notice firstDe-essing and surgical EQ, before any loudness stage
Lossy-export damageA hollow, swirly quality under cymbals and reverb tailsYour own export settings, not the generatorNothing. Re-export losslessly instead

Note the third column. Only the first row has a published cause behind it; the rest are consistent listener reports, and we will not invent a mechanism for them — which is the difference between this page and the ones ranking for the subject, several of which assert exact bands and percentages with nothing cited underneath. Our AI audio cleanup guide walks the whole raw export layer by layer.

Flat vector diagram on a violet-black background: five rounded-square nodes in a row joined by a single line, with a jagged violet waveform already filling the first node on the left and continuing through every node afterwards, lower in amplitude but never flat, the last node drawn in grey
The jagged band is in the file at node one. Every stage after it reduces; none of them removes.
THE PUBLISHED CAUSE

Why they are built in, not a bad export

The one part of this subject that has a peer-reviewable answer, and no page on the first results page for it cites the paper.

In A Fourier Explanation of AI-music Artifacts, published to arXiv on 23 June 2025, four Deezer researchers — Darius Afchar, Gabriel Meseguer-Brocal, Kamil Akesbi and Romain Hennequin — take the familiar “checkerboard” problem from image generation and run it through audio. Their argument is that the deconvolution modules generative models use to build a waveform up to full resolution leave small but distinctive peaks in the frequency spectrum, and that those peaks are a consequence of the architecture itself rather than of the training data or the learned weights. They validate it on open-source models and then on commercial output from Suno and Udio.

The practical consequence is blunt: you cannot prompt your way out of it, and you cannot export your way out of it. If the peaks come from the shape of the decoder, then the same artifacts arrive on your best generation and your worst, at every bitrate the tool offers, in every genre. The only things that change them are a different decoder design or processing applied afterwards.

Our reading of arXiv:2506.19108, abstract and claims, 4 November 2026

The second consequence is the uncomfortable one: the same paper reports that a simple criterion built on those peaks detects AI-generated music with accuracy surpassing 99% in several scenarios, on par with deep-learning approaches. That is not a classifier guessing from vibes but a measurement of a structural fingerprint — which is why “it sounds human to me” has never been a useful test. Our detector write-up covers the tools built on this, and sunowatermark's pillar on how AI music detectors work goes deeper into the signal side.

THE FLOW

The five stages, and what changes at each

Five stages, one signal pass. The fourth column is the one that makes this a flow rather than a black box.

TABLE 2 — THE CLEANUP FLOW, STAGE BY STAGE
StageWhat changesWhat it cannot change
01 GenerateNothing yet. Export the highest-quality file the tool offersThe artifacts are already in the file at this point
02 ProcessThe file is analysed and the artifact bands identifiedYour rights, your metadata or your disclosure
03 ReduceSpectral peaks, harsh high end and transient roughness come downArrangement, melody, structure, or a repetition problem
04 ReviewYou decide whether the creative core survived the subtractionNothing — this stage is a judgement, not a process
05 PrepareHand-off to mastering, then to a distributorA distributor's rights rules or a platform's AI label

Two stages do more work than their position suggests. Stage 01 is where most quality is lost and nobody notices, because the decision is made before you are thinking about cleanup: export lossless if the tool offers it. A compressed export has already thrown away detail that stage 03 would otherwise have had to work with, and you cannot get it back by processing harder.

Stage 04 is not a formality either. Subtraction has a point past which the track stops sounding like itself, and you have traded one artificial quality for another. Compare the two files at matched loudness, because the louder of two almost always sounds better for reasons that have nothing to do with the cleanup. And what no stage in that table touches is the paperwork — which is where most rejections actually happen.

Table of five AI music artifact families with what you hear, where it comes from and what reduces it. Spectral peaks come from deconvolution in the generator's decoder and respond to narrow targeted reduction; smeared transients respond to transient shaping; vocal roughness responds to de-essing and surgical EQ. Too-perfect repetition responds to almost nothing in processing because it is an arrangement problem, and lossy-export damage responds to nothing because the fix is to re-export losslessly
Three of the five respond to a pass. Two do not, and no amount of processing is the fix.
THE HONEST SCOPE

The three gates every release meets

Conflating these three is the single most expensive mistake in this niche, and it is the reason people buy processing expecting it to do something it cannot.

TABLE 3 — THE THREE GATES, AND WHICH ONE PROCESSING REACHES
GateWho decidesWhat it readsCan audio processing change it?
1. The distributor's published rulesThe distributor, from what you tell itRights, impersonation, spam and infringement. DistroKid's AI article lists four conditions and none is about audioNo. Processing cannot change who owns a recording
2. Automated delivery screeningThe distributor's and the stores' automated systemsThe audio itself, plus whatever classifier sits in the pipelineThis is the only gate audio processing reaches at all
3. The platform's AI labelWhoever delivers the metadata — you, or your distributorA declared field — YouTube reads DDEX ERN 4.3.2+ or a CSV “contains_ai” value of all, partly, none or unknownNo. It is declared on delivery; no audio tool writes it

Gate 1 is published and it is not about sound. DistroKid's help article on AI uploads, last updated 19 August 2026, opens “Yes—DistroKid accepts music created with AI tools, but there are some rules” and gives four: own 100% of the rights, no impersonation of a voice or likeness, no mass-generated spam, no infringement. Notice what is absent — no audio-quality condition anywhere in the list — and that it closes by noting streaming services may still reject or remove a release that fails their own guidelines. Nor is DistroKid an outlier: several distributors accept AI music openly, which we map on platforms that accept AI music.

Gate 3 is the one worth being precise about, because it is where most of the marketing in this category is quietly dishonest. A platform's AI label is a declared field, not a measurement of your file. YouTube's partner documentation sets out how: Gen AI information is delivered through DDEX ERN 4.3.2+ or a CSV template, and the contains_ai column takes one of four values — “all”, “partly”, “none”, or “unknown” to reset a value declared wrongly before. Deliver nothing and YouTube reads it as unknown unless it can determine otherwise from its own signals, in which case it may designate the content fully or partly Gen AI itself.

So no audio tool — ours, anyone's — changes how a platform labels your release: the label is typed into a delivery field by you or your distributor. Processing acts on gate 2 and the texture underneath it. The practical version of the declaration side is on disclosure that works.

Table of the three gates a release meets. The distributor's published rules read rights, impersonation, spam and infringement, and processing cannot change them. Automated delivery screening reads the audio itself, and is the only gate audio processing reaches. The platform's AI label is a declared field read from DDEX ERN 4.3.2 or a CSV contains_ai value, and no audio tool writes it
Two of the three gates read declarations, not audio.
THE LITERATURE

What detectors measure, and what nobody has tested

The most recent robustness paper is three months old, and its limitations section is more useful to you than its results.

Improved Robustness in AI-Generated Music Detection (arXiv 2607.27454, submitted 29 July 2026, revised on the 31st) starts from the awkward fact that detectors which are near-perfect on raw output degrade sharply once the audio has been altered — and the alteration it studies is a change of speed or pitch, which moves every one of those spectral peaks along the frequency axis. Its own detector is invariant to frequency scaling, trained on speed factors from ×0.7 to ×1.4.

TABLE 4 — SUNO V5 UNDER SPEED MODIFICATION (ARXIV 2607.27454, TABLE 2)
DetectorTraining conditionAUCF1
Afchar et al. (the Fourier criterion)Clean0.7450.008
SpecTTTra-α, as releasedClean0.6920.302
Proposed scale-invariant modelClean0.9860.732
Afchar et al. (the Fourier criterion)Trained on the attack0.8170.675
Proposed scale-invariant modelTrained on the attack0.9970.986

The 0.008 F1 in the first row is not a typo, and it is the most instructive number here: a detector that is excellent on untouched audio can be reduced to almost no useful positive detections by a transform you could apply in any editor. Note what closes the gap — the two “trained on the attack” rows. Detection is an arms race with a short cycle, so a figure measured in 2025 tells you little about the model running in 2026.

“Other common manipulations (equalization, dynamic range compression, codec compression) remain untested”.

arXiv:2607.27454, Limitations, read at source 4 November 2026

That sentence describes almost everything a cleanup and mastering chain does. The literature has measured a speed change, which no sane release workflow applies, and has not measured EQ, dynamic-range compression or codec compression, which every release workflow applies. The training set was also limited to three generators — Suno v3.5, v5 and Udio v120.

Which is why you will not find a number on this page for how much any processing moves any detector's score. Nobody has published one, our own release-side testing is described on the Undetectr review rather than smuggled in here as physics, and a page quoting a precise percentage for an unmeasured transform is quoting itself.

Table of detector performance on Suno V5 under speed modification from arXiv 2607.27454 Table 2. Trained on clean audio, the Afchar Fourier criterion scores 0.745 AUC and 0.008 F1, SpecTTTra-alpha scores 0.692 and 0.302, and the proposed scale-invariant model 0.986 and 0.732. Trained on the attack, the Fourier criterion scores 0.817 and 0.675 and the proposed model 0.997 and 0.986
The paper's own figures, for one attack. Not a measurement of what a cleanup pass does.
DON'T CONFUSE THESE

Cleanup is not mastering, and the order matters

TABLE 5 — THREE THINGS PEOPLE RUN INTERCHANGEABLY, AND SHOULD NOT
ProcessWhenWhat it doesWhat it does not do
Artifact cleanupFirstReduces the synthetic signatures in the source texture — spectral peaks, harsh highs, rough transientsSets no loudness target and fixes no arrangement
MasteringAfter cleanupSets final loudness, balance and width. Spotify publishes −14 dB LUFS and a −1 dB true-peak ceilingDoes not remove artifacts; on a rough file it amplifies the edges you wanted gone
Remaster or regenerationInstead of both, or not at allHands back a different render of the song, which may be cleanerA new generation, not a repair — the performance can change under you

Run them in that order. Mastering is a loudness and balance stage, so pointing it at a file with harsh spectral peaks raises those peaks along with everything else — the limiter does not know which part of the top end you wanted and which part the decoder left behind. Cleanup first gives mastering a better starting point; it does not replace it. And be careful whose target you are using: Spotify is the one major platform that publishes a loudness figure at all, and the per-platform tables you see everywhere else are third-party measurement presented as specification, which we go into on Suno mastering.

The third row catches people out. A remaster is not a repair pass — it hands you a different render, and the performance you liked can change underneath you. If the vocal take was why you kept the track, that is a real risk rather than a footnote, and our route comparison puts numbers on what each option costs you.

Table of three processes run interchangeably: artifact cleanup goes first and reduces synthetic signatures but sets no loudness target and fixes no arrangement; mastering comes after cleanup, sets loudness against Spotify's published minus 14 dB LUFS and minus 1 dB true-peak ceiling, and on a rough file amplifies the edges you wanted gone; remaster or regeneration hands back a different render and is a new generation, not a repair
Mastering a rough file amplifies what you wanted gone.
THE LIMITS

What a cleanup pass cannot fix

A weak arrangement. The too-perfect repetition in Table 1 is the clearest case. No processor has an opinion about whether your second chorus should differ from your first, and a loop reads as machine-made because nothing in it moves. That is fixed by editing, re-generating a section, or playing something over the top — not by a signal pass.

Detail that was never there. Subtraction cannot restore information. If the generator did not produce a convincing cymbal, cleanup gives you a quieter unconvincing cymbal. The same applies to anything a lossy export discarded on the way out, which is why the last row of Table 1 reads “nothing”.

Marks a generator deliberately appended. A different layer, with different rules. Suno's terms of service, revised 10 August 2026 and effective 3 September 2026, state that you agree not to “remove, alter, obscure or circumvent any fingerprint, watermark or metadata Suno appends to an Output”, and that obtaining a copy of an Output other than through a download channel Suno makes available is prohibited. An artifact pass works on the audible texture the decoder leaves behind as a by-product of generation — not the same object as a mark somebody added on purpose, and a distinction to keep straight before reaching for any tool. The identifier layer is covered on the watermark page.

A platform's AI label, and the audience problem. Gate 3 is declared, so nothing in the signal reaches it. And the honest closing caveat is bigger than all of this: what most AI releases are short of is not cleanliness or distribution, it is listeners. A clean file clears a step. It does not find you an audience, and anyone implying otherwise is selling you something.

PRACTICE

Running the pass without wrecking the track

Keep the original, always. Process a copy. This sounds obvious until the third pass, when you have lost track of which file still had life in it — and a generator's output is not always reproducible later.

One pass, not a stack. The instinct when something still sounds synthetic is to add a second cleanup stage, then a third; the result is a dull file with the same repetition problem. If one targeted pass has not fixed it, the problem is in a row of Table 1 cleanup does not own.

Work per stem where you can. A broad high-shelf cut that tames a metallic hi-hat also takes the air off the vocal. Where stems exist the narrow fix beats the wide one; where they do not, dynamic processing that acts only when the harshness appears beats a static cut that is always on.

Then the release steps, in order. Clean, master, check the rights conditions at gate 1, declare accurately at gate 3, submit. For the artifact and screening part — the step this site exists for — Undetectr is the pass we point creators at, and the button below goes straight to it. It works on the texture and the screening step, and has no effect whatsoever on how a platform labels your release.

And when the file is finished, the next wall is the one in the section above. Streaming discovery is closed enough that a storefront and a sync route are worth more to most people than another upload: paid placements in TV, film, games and ads are where the money talk in this niche actually is, and played.fm's sync route is built for pitching into them, with selling direct to your own listeners beside it. Worth noting the contrast, too: Bandcamp bars music generated wholly or substantially by AI, so “just sell it on Bandcamp” is not available to this catalogue.

FINAL SIGNAL

Ready to run the cleanup pass?

Move your AI track through a pass built for generation artifacts and delivery screening — not for changing any platform's AI label, which no audio tool can do.

▸ Independent · We recommend Undetectr

QUESTIONS READERS ASK

AI music artifacts FAQ

What causes AI music artifacts?

For the metallic, shimmery high end, Deezer's researchers trace it to deconvolution modules inside the generator's decoder — architecture, not training data or weights, validated on Suno and Udio output. That is why a better prompt or a cleaner export does not remove it. Other complaints, such as blunt transients, are reported consistently by listeners with no published cause we will cite.

Can you remove AI artifacts completely?

No. A cleanup pass is subtraction: it lowers energy in the bands where the artifacts live and softens rough transients. It cannot restore detail the generator never produced, and it cannot reach a repetition problem, which is arrangement rather than signal.

Does cleaning a track stop it being detected as AI?

Nobody can honestly tell you by how much, because the published literature has not measured it. July 2026's robustness paper tested speed and pitch changes and says in its limitations that equalization, dynamic range compression and codec compression “remain untested” — precisely what a cleanup and mastering chain performs.

Does a cleanup pass remove Suno's watermark?

That is a different layer with different rules. Suno's terms of service, effective 3 September 2026, say you agree not to “remove, alter, obscure or circumvent any fingerprint, watermark or metadata Suno appends to an Output”. An artifact pass targets the audible texture the decoder leaves behind as a by-product of generation.

Will a distributor reject my track because of artifacts?

Not for the artifacts as such. DistroKid's AI article accepts music made with AI tools on four conditions — rights, impersonation, spam, infringement — and none of them is an audio-quality rule. What creators do report is automated screening flagging and rejecting individual releases.

Does exporting at a higher quality reduce artifacts?

It does not reduce the ones the generator created, because those are in the audio before you pick a format. It does stop you adding a second, avoidable layer: a lossy export discards detail a later cleanup stage would have worked with.

SOURCING

Sources, and what we could not read

  • A Fourier Explanation of AI-music Artifacts — Afchar, Meseguer-Brocal, Akesbi and Hennequin, arXiv:2506.19108, 23 June 2025. The deconvolution account of the spectral peaks, the architecture-not-weights claim, validation on Suno and Udio, and the detection criterion surpassing 99% accuracy in several scenarios.
  • Improved Robustness in AI-Generated Music Detection — Dugelay, Barand, Laouar, Campeas, Afchar and Hennequin, arXiv:2607.27454, 29–31 July 2026. Table 4 above is its Table 2, read in the full HTML version, along with the ×0.7–×1.4 speed range, the three generators used, and the limitations sentence quoted verbatim.
  • DistroKid, Can I Upload Music Made With AI Tools to DistroKid? — help article 41182362733715, last updated 19 August 2026, for the opening sentence and the four conditions.
  • YouTube Help, Gen AI disclosures for music partners — for DDEX ERN 4.3.2+, the four contains_ai values, the Fully / Partly / No Gen AI designations, and the rule that an omitted value is read as unknown unless YouTube determines otherwise from its own signals.
  • Spotify for Artists, loudness normalization for −14 dB LUFS, the −1 dB true-peak advice and the three listener presets, and Suno's terms of service (revised 10 August 2026, effective 3 September 2026) for the fingerprint-and-watermark clause and the download-channel clause.
  • What we could not read. DistroKid's help centre returns 403 to a plain request from our environment, so that article was read through its public JSON endpoint, which serves the same body and update date. One ranking page, ceek.com's “Behind the Glitch”, now 404s and is named unread rather than summarised. We ran no audio tests for this page: every figure above is someone else's published number, with a link to it.

Detection research on this subject is moving in months, not years, so treat Table 4 as a snapshot of July 2026 rather than a standing result. Sources read at origin 4 November 2026. Last updated 4 November 2026.