Published 2026-07-17 · by Brian Crumrine

Releasing an AI-produced album in 2026 - what the platforms won’t tell you

I’m Brian. I write and produce music under a few artist names - Wrong Aisle for concept albums, Brian 200 for interior stuff, Snacks Not Found for holiday EDM, Subjects for songs about specific people in my life. I also produce for RT Wright, a retired teacher with a 114-song catalog he’s been carrying around for forty years. My day job is software engineering. I document music releases the way I document and test software.

This spring I finished a Wrong Aisle album called Gift Wrapped. Twelve tracks, one per TikTok Live gift, priced ascending - a Rose up to a Universe. Built for the community that trades those gifts. Submitted to SoundOn, TikTok’s own distributor, on July 9. As of this writing it hasn’t shipped. Universe got flagged, my appeal turned into a policy rejection I couldn’t argue back, and I can’t resubmit until end of month. Here’s what I learned about how SoundOn actually works that isn’t in any docs.

Why this album, why SoundOn

Gift Wrapped is a TikTok-native concept record. Every track is about a specific TikTok Live gift, in the genre that gift’s personality demands - bossa nova for the Rose, chiptune for the GG, K-pop pastiche for the Finger Heart, trap for the Panda, 2-tone ska for the Doughnut. Every one played completely straight. (The Panda got retired from TikTok’s gift lineup. I loved the song too much to cut it.) Audience for this specific album is a niche: TikTok Live hosts, gift senders, sellers who need reveal-moment soundtracks.

But this album isn’t why I’m on SoundOn. The bigger reason is that SoundOn is the only free, self-serve path into TikTok’s Commercial Music Library - a curated pool separate from the general Sound Library, licensed for commercial use so brand and shop content can post without licensing risk. TikTok Shop affiliates post product-promotion videos for commission - commercial content, general library locked out, CML required. Same for any TikTok surface where commercial licensing matters.

SoundOn distribution also licenses your tracks for CapCut (ByteDance’s video editor), the other big commercial-context surface in the ecosystem. Distribution establishes the license - any track shipped through SoundOn is legally usable in CapCut. But actual placement in CapCut’s discoverable library and template pool appears to run through a separate curation gate. SoundOn’s Breakthrough promotional program lists “placement in a CapCut template” as an explicit activation, so it’s not automatic from shipping. Being on SoundOn is the prerequisite for that placement ever happening, though. Every other distributor doesn’t put you in position at all.

Every other distributor (DistroKid, TuneCore, CD Baby, RouteNote) delivers to the general library instead of the CML. DistroKid in particular is fast and reliable - I use it for non-CML releases and it’s the ten-year-old interface that just works. So SoundOn wasn’t just the right call for this album. It was the pipeline test. Gift Wrapped is proof-of-concept for how I get catalogs in front of Shop affiliates and other ByteDance commercial-context surfaces. I have catalogs coming that are designed around those audiences specifically. The friction I was about to eat isn’t a cost of shipping this album. It’s a one-time cost of building the road.

The rejection

The album went in July 9. Four days later, July 13, this came back:

The release was not approved because it doesn’t comply with our delivery policies.

The track Universe was not approved due to the following reason(s):

Audio: suspected use of unauthorized music samples, remixes, covers (please contact SoundOn if you have permission to make covers and use music samples)

One track out of twelve. Universe - the closer, the maximalist prog-EDM finale, the top of the price ladder. The rest of the album was implicitly approved.

Before appealing I ran the master through Shazam and a second recognition service. Both clean. That’s the two-minute test worth doing before any argument. Real song comes back, the flag is legitimate - re-render or leave it off the album before you submit. Nothing comes back, the flag is fuzzy - probably a partial fingerprint match against something in the platform’s reference catalog. Festival EDM trips this more than bossa nova because supersaws and big-room chords sit in the same crowded space as ten thousand other tracks.

Nothing came back. I appealed.

First thing that isn’t in any documentation: fingerprint filters false-positive more on crowded genres than sparse ones. Pre-scan your masters with Shazam before submitting, especially anything festival-EDM adjacent. The tool is free. The submission slot is not. That last part matters more than I understood at the time.

The appeal

The evidence was strong. I shared a Suno-hosted link for the winning Universe take - my third or fourth attempt on the same prompt, from my account, with full lyric and style prompts visible. The keeper was dated July 5, four days before I submitted the album to SoundOn, with the earlier attempts sitting above it in the take list. Third-party-hosted provenance nobody could argue with, and incidentally the exact shape of iteration you’d expect from someone actually working on a song rather than laundering someone else’s. I attached screenshots, pasted the full prompts as text, mentioned the negative Shazam results, and asked them to share the matched work if there was one.

SoundOn support replied:

After careful consideration, we would like to inform you that based on our current platform policies and the guidelines established by our partners, we are unfortunately unable to accept this type of track at this time.

Read that carefully. The samples suspicion is gone. The reply doesn’t defend it, doesn’t repeat it, doesn’t mention it. It’s been replaced with something else. “This type of track.” “Guidelines established by our partners.” Partners meaning TikTok.

I asked for clarification. The reply:

Regarding the eligibility criteria, please note that all releases must be 100% compliant with platform policies. This includes restrictions on the use of samples without authorization, AI-generated content, or any material that may not meet originality and licensing standards.

There it is, in writing, from SoundOn support: AI-generated content restricted. This is not on their public FAQ. It’s not on their pricing page. It’s not in their license agreement (I’ve read the whole 19-page thing looking for it). It’s also not consistent with eleven other tracks on the same album - same tool, same account, same submission - approved without comment.

The operative rule underneath the policy statement: their filter flags occasionally, and disclosed AI gets rejected under a catch-all clause. Undisclosed AI that doesn’t trip the filter gets approved. A very different policy from “AI content restricted.” A policy about disclosure, not about the music.

So my honest, thorough, evidence-heavy appeal was the thing that killed Universe. The filter had raised a beatable question. Disclosure converted it into an unbeatable one. Ten of eleven approvals came from not being asked. Volunteering was the mistake.

Second thing that isn’t in any documentation: for AI-produced music, disclosure via appeal is worse than the appeal being denied on the merits. If a track flags at review, don’t argue. Delete the release, re-render the flagged track, and resubmit the whole album. If the re-render flags too, delete and resubmit without it. There’s no policy-level path back once “AI-generated content” is written next to your release in someone’s ticket queue.

The quota trap

The instruction was to delete the rejected album and resubmit without the flagged track. Fine - the eleven approved tracks were still good.

Except the release editor doesn’t let you remove tracks. Twelve slots, twelve required audio files. So I went to delete the whole thing and re-upload as eleven.

Except: quota reached. I could not upload again until July 31.

I asked SoundOn to reset it, since I was following their earlier instructions. The reply confirmed the mechanic in writing:

Based on how our system works, submitting a release already counts toward your monthly upload quota, even if the release is later deleted or rejected. Because of this, your account has reached the upload limit for this month.

Read that carefully too. The free-tier upload limit is 5 songs per month per their FAQ. What that email tells you is the limit isn’t 5 songs. It’s 5 attempts. A rejection burns a slot. A deletion burns a slot. A re-upload after fixing a flagged track burns a slot.

So the real free-tier ceiling: submit a release you’re not sure will pass, and your whole month can go to one album that never ships. My twelve-track album that got eleven approvals consumed my July quota. The eleven that passed cost me the same as a rejection would have.

Third thing that isn’t in any documentation: every SoundOn submission is expensive. Pre-scan every master, pre-check every metadata field, treat each upload as if it’s your only one for the month. Because it might be.

What the platforms won’t tell you

Four things I learned about how SoundOn actually works, none in any public policy, FAQ, or license agreement:

The filter is inconsistent. Eleven identical productions approved, the twelfth flagged. Same tool, same mastering, same submission, same credits. If yours flags, it’s not because you did something wrong. It’s because their filter did something arbitrary.

Disclosure is the trigger, not the technology. Their submission form has no field for AI disclosure. Their License Agreement is silent on it (all 19 pages, zero mentions). TikTok’s own Music Terms of Service explicitly permit AI-generated music with disclosure - captions, hashtags, or the platform’s AI-generated content label - but that policy governs music in user-uploaded videos, not tracks delivered into TikTok’s library through a distributor. SoundOn, as the distributor pipeline into that library, has no equivalent public policy. Their April 2026 partnership with ACRCloud targets derivative works (sped-up, slowed, and pitched uploads), not AI screening. But disclose AI production in an appeal and a human reviewer applies a “restricted content” clause that appears nowhere else in the flow. The unwritten policy is “don’t ask, don’t tell.” Stable equilibrium until you break it.

Support is polite but structurally powerless. Every reply was courteous, quick, and clear. What support couldn’t do: remove a single track, partially approve one, or reset a quota consumed by their own team’s instructions. Front-line support at SoundOn is a communication layer, not a decision layer. That’s how the company scoped the role, not a criticism of the people doing it. Ask for clarity, not exceptions.

The free tier is a hard ceiling dressed as a soft paywall. Five attempts per month, with rejections consuming quota, isn’t a policy about upload volume. It’s a policy about surface area. Support hints at “plans and services available for creators with higher distribution needs,” but as of this writing I can’t find any paid tier actually available to subscribe to. So release more than one album a quarter on the free tier and you just don’t release it. There’s no purchasable exit.

And it’s not just SoundOn. In January or February 2026, LANDR - a different distributor - silently didn’t deliver my AI tracks to TikTok and YouTube. I found out because the tracks never showed up on those platforms. LANDR’s reasoning at the time was that those platforms don’t allow AI music, which was at odds with TikTok’s own Music Terms of Service then and now. Same industry-wide pattern: undocumented enforcement, no proactive notice, different distributors answering the same question differently and changing their answers over time.

But the industry has been moving fast since. As of May 2026, DistroKid requires AI disclosure at upload - a checkbox for whether lyrics, vocals, melody, or instrumentation were AI-generated. That disclosure flows through DDEX metadata to Spotify (AI Credits launched April 2026), Apple Music (Transparency Tags, May 2026), and Deezer (which combines self-reporting with auto-detection, and excludes fully-AI tracks from editorial playlists regardless of what you declare). SoundOn hasn’t built any of that. Two paths taken in the same year: transparent labeling with continued distribution, or silent enforcement without disclosure. Most of the industry is choosing the first. SoundOn’s ticket queue is the outlier.

None of it is written down. All of it is enforced. The rules exist in the ticket queue, not in the documentation. Every artist coming to these platforms learns them by paying for them. Not conspiracy. Not malice. Just the ordinary fact that platforms in transition write their real policies in support tickets, and by the time you know the rules you’ve already paid tuition.

On AI music being real music

The temptation, reading the above, is to argue SoundOn is treating AI music unfairly. Not going to make that argument. It presumes AI music needs the platform’s blessing to be real, and it doesn’t. So I’m not defending it. Just telling you what making it looks like.

“The Ambush,” another Wrong Aisle track, took three versions to render right. Version 1 buried the drop in a trailing clause of the style field. The model reasonably averaged the states it was given and produced a track that was mostly chill with a slightly darker middle. Version 2 overcorrected: I led the field with “dark dynamic contrast” and the model treated that as the global palette, painting the whole song harsh from bar one. Version 3 wrote the field in time order - innocent café world first and dominant, ambush as a single late event - and hard-coded a full bar of dead silence before the drop, which is the single strongest dynamics cue the model responds to.

That’s craft. Diagnose the failed take, form a theory about why it failed, revise the input, re-test. Same iterative practice a producer runs on a session musician who isn’t playing the part right. Tool is compressed. Judgment isn’t.

The album’s concept - every TikTok gift as its own genre, played completely straight, with the parasocial-gifting-economy adult layer underneath - didn’t come from the model. The lyric for Rose (“send me once, send me small, one coin, that’s all”) didn’t come from the model. The decision to close the cover art on a shut gift box with a rose floating above it rather than a lid blowing off, because a closed box says “there is something enormous in here” more effectively than an eruption - didn’t come from the model.

I make music with AI the way a photographer makes pictures with a camera. Camera does the mechanical part. Photographer decides what’s in the frame, and why, and when to press the button. Everyone accepted that argument for photography a century ago. AI music is where photography was in 1900. Not going to tiptoe for a decade waiting for the world to sort it out.

What I chose not to do

Tools exist that remove AI signatures from generated audio. They work. Using one would have made the appeal succeed. I didn’t. Three reasons:

First, it wouldn’t have worked reliably. The recording exists on Suno’s servers regardless of what I strip from the file, with the prompts and timestamps intact on their side. Distributor-side fingerprint partnerships are tightening - SoundOn’s April 2026 ACRCloud deal is exactly that kind of infrastructure. Even if a stripped file passed today, a retroactive scan later could match my master to Suno-hosted records I don’t control. And in my case specifically, I’d already sent SoundOn the Suno link with the appeal - the paper trail is on their side too. The audio doesn’t stop being Suno-generated just because the file metadata forgets.

Second, the risk asymmetry is terrible. Right now, if a track flags, I’m an artist whose track got filtered - recoverable, no lasting mark. Get caught stripping signatures once and I’m an account that deliberately defeated screening. You can’t win an arms race where the other side owns the detectors, controls the archive, and can re-scan retroactively.

Third: stripping concedes the position it claims to defend. That’s the one that matters. Making AI music pass as human surrenders the claim that AI music is legitimate. The honest way to defend that claim is to make good work openly and let the receipts stand.

If you’re in this position now, the tempting move is the wrong move. Do the thing that lets you keep making music openly for the next ten years, not the thing that gets one album approved this week.

What actually works

If you’re releasing AI-produced music on SoundOn right now, the practical guide, in order:

Pre-scan every master. Shazam plus one other recognition service, every track, before it goes anywhere near a submission form. Especially anything in a crowded sonic space - festival EDM, big-room house, mainstream pop drops. If anything matches, that track isn’t ready. Re-render or leave it off.

Pre-check your style-field language for Suno-side collisions. Certain words trigger Suno’s own render-time filter, which silently swaps the style field for something generic and produces a different track than you spec’d. “Skank” (the ska technique, same as a band name), “triggerman” (the sampled New Orleans bounce beat), any well-known artist name, any famous song title. This is a Suno problem, not a distributor problem - your prompt never leaves Suno, and the mastered file that ships to SoundOn has none of your prompt language in its metadata.

Don’t volunteer to SoundOn. Their form doesn’t ask. Their license doesn’t require it. The undisclosed status quo is what allows their filter to approve your work. Silence isn’t dishonest when the platform designed silence into the flow. On distributors that do ask - DistroKid, and increasingly others - answer honestly. Their disclosure infrastructure is aligned with the DSPs downstream, which display the information without rejecting the release.

Never appeal with disclosure. If a track flags at review, delete the release and resubmit with the track re-rendered. If it flags again, delete and resubmit without it - SoundOn’s editor doesn’t let you drop single tracks from a release. Appeals with AI provenance evidence convert filter-level flags into policy-level rejections that can’t be argued back.

Treat every submission as expensive. The quota is 5 attempts per month, not 5 songs. Rejections and deletions consume quota. A five-track EP with one flagged track can burn your whole month if you delete-and-resubmit.

Own your surfaces. The album’s home is a page on your own site, not the DSP. Distribution channels come and go and their rules change without notice. Your surface is the one thing that can’t be pulled at the discretion of a filter. This is what saved me from writing this piece angry instead of clear.

The receipts

Everything above is verifiable.

Public documents referenced: TikTok Music Terms of Service, which contains a dedicated AI Music section permitting AI-generated music in user-uploaded videos with disclosure requirements (captions, hashtags, or the AI-generated content label). SoundOn’s April 2026 partnership with ACRCloud, announced as “Derivative Works Detection” targeting manipulated audio (sped-up, slowed, pitched, mashup uploads) rather than general AI screening. SoundOn’s License Agreement (all 19 pages) mentions AI, generative content, labeling, or disclosure zero times.