90,000 AI Tracks a Day: What That Number Really Means

A number like this counts what got uploaded, not what got listened to. It is a supply figure, and on its own it tells you nothing about demand, quality, or whether your own track has any chance. What it does tell you is cheaper and more useful: pressing submit has become so cheap that the bottleneck has moved from making a track to getting a track found, and that shifts where your effort belongs. Below: four questions to ask any platform statistic, a five-layer funnel from upload to paid play, what the listening-side filter announced for late 2026 actually changes, and a one-page disclosure file you can keep for every track you release.
Start with what the number is counting
The figure that has been circulating for Deezer in recent reporting is around ninety thousand fully AI-generated tracks arriving each day. Treat that as a reported figure with a date on it, not a fixed constant: platform numbers like this move monthly, and the honest move is to re-check the source before you repeat it. The reading method below works regardless of whether the current figure is higher or lower.
The single most useful thing to understand is that this is a count of submissions, not of listens, not of money, and not of anything a listener did. Almost every wrong conclusion people draw from it comes from skipping that distinction.
| What people read it as | What it actually measures | Why the gap matters |
|---|---|---|
| "People want AI music" | People can make AI music cheaply | Supply and demand move independently; a flood of supply often means the price of attention fell |
| "AI music is taking over the catalogue" | AI music is taking over the intake queue | Intake and catalogue are different denominators by orders of magnitude |
| "There is no room for my track" | There is a lot of competition for the default discovery paths | Default paths are not the only paths; direct links and owned audiences route around intake volume |
| "The platform must be flooded with plays" | The platform is flooded with files | Most uploaded material, human or generated, reaches very few plays |
| "Quality must be terrible" | Nothing about quality is measured | A detection tag is a claim about origin, not a judgement about the music |
Read that table as a translation aid. Every time you see the number quoted, ask which row the writer is standing in. Most hot takes sit in the left column and argue from it as though it were the middle one.
Four questions to ask any upload number
These four questions work on any platform statistic, not just this one. If you cannot answer all four, you do not yet know what the number means, and you should treat it as a signal of direction rather than a measurement.
- Who counted it? A platform counting its own intake is reporting on itself. That is not dishonest, but it does mean the measurement method, the threshold, and the timing all belong to the party with the strongest interest in the story. Note who published it and where.
- Counted when? Intake figures are snapshots. A number from one quarter says nothing about the next, and a figure quoted a year later without a date is close to worthless. Write the date next to the number every single time.
- What counts as AI? This is the question that changes the number most. Fully generated, partially assisted, cloned voice, stems run through a model, and a human performance with light processing are five different things, and a statistic that adds them together is answering a different question than one that separates them.
- What is the base? A daily count means nothing without the total daily intake behind it, and a percentage means nothing without the absolute number beside it. Ask for both. A rising share of a shrinking total is a different story from a rising share of a growing one.
| Question | If the source cannot answer it | Where to look instead |
|---|---|---|
| Who counted it | Treat the figure as a claim, not a measurement | Platform newsroom or the original press statement |
| Counted when | Assume it is already stale | Publication date of the article quoting it |
| What counts as AI | Assume the widest possible definition | Platform help pages on tagging and disclosure |
| What is the base | Compare only against figures from the same source | Platform transparency or annual reporting |
Two of these questions are cheap to answer and one is expensive. The expensive one is always the definition question, which is exactly why it is the one most often skipped.
The five-layer funnel from upload to paid play
This is the framework worth keeping, because it turns a scary headline into five ordinary checkpoints, four of which you can act on. Every track you release passes through all five layers in order, and most tracks stop somewhere in the middle.
| Layer | What happens | Where tracks stop | What you control |
|---|---|---|---|
| 1. Submitted | A file and its metadata enter the intake queue | Format, duration, and metadata problems | Everything at this layer |
| 2. Cleared | Rights, identity, and duplicate checks run | Conflicting metadata or a rights question | Your paperwork and your consistency across releases |
| 3. Tagged or not | Detection decides whether the origin is flagged | Nothing stops here; a tag changes routing, not existence | Whether you disclosed anything at all |
| 4. Reached a listener | Search, algorithmic placement, editorial, or a direct link | The overwhelming majority of all material | More than people think, and it is the real work |
| 5. Paid | Plays cross whatever threshold the platform sets | Low play counts, or disputed attribution | Almost nothing directly, which is why layer 4 matters |
The important shape here is that volume collects at layer 1 and revenue happens at layer 5, and the layers in between are where the filtering happens. A statistic about layer 1 tells you the queue is long. It says nothing about how many tracks reach layer 5, and it says nothing about whether yours can.
It also tells you where not to spend your time. Arguing about layer 1 is arguing about other people's uploads. Layers 2 and 3 are paperwork and honesty, and they are cheap to get right. Layer 4 is the only hard one, and it is hard for the same reason it has always been hard: nobody owes you an audience.
What the listening-side filter changes
In October 2026 Deezer announced a plan to roll out a global setting, before the end of the year, that lets listeners hide music identified as fully AI-generated from their search results. Fully generated is the operative phrase: material made by a person with tools assisting parts of the process has been described as outside the scope of that setting. Deezer had already been excluding fully generated material from algorithmic recommendation and editorial playlists, and it has said its own detection tool has been running since early 2025.
Verify the current wording before you rely on any of it, because rollout plans change. What matters more than the exact date is the direction, and the direction is clear: origin is becoming a routing decision rather than a hidden property.
| What is in scope | What has been described as out of scope | What that means for your workflow |
|---|---|---|
| Tracks identified as fully generated | Assisted work where a person directed the result | Keep a record of what you actually did, not just what tool you used |
| Search visibility when a listener opts in | Listener who leaves the setting off | Do not build a plan that depends on one platform's default |
| Recommendation and editorial placement | Direct links and links you send yourself | Owned distribution is unaffected by placement rules |
The practical read: a filter like this does not remove your track, it removes one path to it. Paths you bring yourself are not affected by anyone else's toggle, which is an argument for keeping a plain link to your work somewhere you control.
Your disclosure file: ten items per release
Whatever the rules turn out to be on any given platform, the side that wins disputes is the side with a dated record. This is a checklist you can run in a few minutes per track, and it costs you nothing to maintain.
- Date of generation and the tool used, including the version if you can see it.
- The exact prompt or brief you submitted, saved as text rather than remembered.
- What you changed afterwards: edits, arrangement, lyrics, mixing, mastered elsewhere.
- What you wrote yourself: lyrics, melody, structure, or none of those.
- Whether any vocal is synthetic, and whether it imitates a specific performer.
- Whether any sample or stem came from elsewhere, and under what terms.
- The export settings: sample rate and bit depth, format, and file name.
- Release date and the distributor used, plus the submission reference.
- Where the source files live, so you can produce them later if asked.
- One sentence of plain description of how the track was made, written at the time.
That last item is the one people skip and the one that gets used. A sentence written on the day you made the track is worth more than a paragraph reconstructed a year later under pressure.
Copy-paste disclosure template — fill it in and store it next to the master file:
Track: [title] Released: [date] via [distributor] Origin: generated / assisted / performed (pick one) Generation Tool and version: [tool, version] Date generated: [date] Brief submitted: [paste the prompt or brief] Takes generated: [number] Selected take: [identifier] Human contribution Lyrics: written / edited / generated Melody: written / edited / generated Structure: written / edited / generated Editing: [what you changed after generation] Mix/master: [where it was finished] Source material Samples or stems from elsewhere: none / [list] Terms for each: [license or note] Synthetic vocal imitating a named performer: no / [explain] Files Export: [format, sample rate, bit depth] Stored at: [path or link] Submission reference: [distributor reference] One-line description, written at the time: [This track began as a generated instrumental; the melody and lyrics were written afterwards and the arrangement was edited by hand before mastering.]
Keep one of these per release in a single folder. When a platform asks a question, or when someone contests a track, the answer is a file you already have rather than a memory you have to rebuild. This is the same discipline as keeping a claim-ready evidence file, applied at release time instead of after a dispute.
A claim ledger for the numbers you repeat
If you are going to quote a statistic in your own writing, in a pitch, or in a description of your work, keep a ledger entry for it. This takes thirty seconds and prevents the most common failure, which is a figure that gets more confident as it travels further from its source.
CLAIM: [the number, in words] Source: [who published it, and the URL] Published: [date of the source] Checked: [date you looked] Definition: [what counts as AI, in the source's own words] Base: [total, or "not stated"] Quote-ready: yes / no If no, why: [missing definition / missing base / undated / secondhand]
The "quote-ready" line is the whole point. A figure that fails it is still useful to you as a sense of direction; it just does not belong in a sentence where someone will read it as a measurement. Our own rule here is simple: if the base is not stated, the number gets described as "reported" and dated, never as settled.
Re-checking the number yourself in ten minutes
You do not need anyone's permission or any special access to sanity-check a headline figure. This is a routine you can run whenever a number starts circulating.
- Find the original source, not an article about it. Platform newsrooms and press statements are usually short and easy to read.
- Write down the publication date next to the number.
- Find the sentence that defines what counts as AI. Underline it.
- Find the total, or note that no total is given.
- Convert to a share only if you have both numbers, and never convert in your head.
- Search for a more recent statement from the same source and compare.
- Note whether the newer statement changed the definition, not just the number.
- Record all of it in the claim ledger above.
- Decide whether the figure is quote-ready, and write that decision down.
- Set a reminder to repeat this in three months, because intake figures age badly.
Step 7 is the one that catches the interesting cases. A number that jumps between two statements often jumped because the definition changed, and a definition change is a much better story than a number change.
What actually changes for you
Very little of this alters the daily work, and it is worth being blunt about that. The job is still: write a clear brief, generate a few takes, pick one, fix what is wrong with it, finish it properly, and document what you did. None of that gets harder because someone else uploaded a lot of material.
Two things genuinely do shift:
- Paperwork went up in value. When origin becomes a routing decision, a dated record of how a track was made stops being optional tidiness and becomes part of the release. The disclosure template above is the whole cost of that.
- Default discovery went down in reliability. When intake is crowded and placement rules are in flux, a plan that depends entirely on one platform's recommendations is a plan with a single point of failure. Anything you send directly is unaffected by it.
Cost has not changed either. Making the track was never the expensive part; the four-layer budget model breaks that down, and the honest answer on earnings is unchanged by any intake figure.
When a headline number makes you want to change plans
| Reaction | What the number supports | What to do instead |
|---|---|---|
| "I should stop, the market is saturated" | Nothing; saturation of intake is not saturation of attention | Ship one finished track and measure your own reach, not the queue |
| "I should hide that I used a tool" | Nothing; disclosure is the durable position | Fill in the disclosure file and keep it dated |
| "I should upload much more volume" | Only if your bottleneck is layer 1, which it usually is not | Find which layer your tracks stop at before adding volume |
| "I should switch platforms" | One platform's intake is not another platform's | Read each platform's own terms, then decide |
| "I should check my licenses again" | Yes, this one is actually reasonable | Read the license before you ship, not after |
Common questions
How many AI tracks are uploaded to streaming platforms each day?
The figure most often quoted for Deezer in recent reporting is around ninety thousand fully generated tracks a day, but intake numbers like this move quickly and definitions vary between sources. Treat any such figure as dated and reported rather than settled, then run the ten-minute re-check above before repeating it. The important point is not the exact number: it is that intake volume is now large enough to shape platform policy.
Does a high upload number mean listeners are hearing more AI music?
Not necessarily. Upload counts measure submissions, and most submitted material of any kind reaches very few plays. A large intake number tells you about the cost of making files, not about what listeners chose. If you want to know what listeners are hearing, you need play-side data, which platforms rarely publish in useful detail.
Will my track be removed because it was made with a generator?
A tag on origin has been used to change routing rather than to delete material: what has been described publicly is exclusion from recommendation and editorial placement, and an optional listener setting for search visibility. Removal is a different mechanism, usually tied to rights questions rather than to how the audio was produced. Read the current terms of each platform you distribute to, because they are not identical.
Does using AI for part of a track count as fully generated?
That depends entirely on the definition in force, which is why the definition question matters more than the number. Assisted work has been described as outside the scope of the listening-side setting announced for late 2026, but do not rely on a summary: read the wording, and keep your own record of which parts you wrote and which parts the tool produced.
Should I disclose AI use on my own releases?
Yes, and not because a platform forces you to. A dated, honest record is the position that survives a policy change, because policies move in one direction and your file can move with them. Use the disclosure template above; it takes a few minutes per release and costs you nothing to maintain.
What should I actually do differently after reading a number like this?
Two things. Keep a per-release disclosure file so your paperwork is ready before anyone asks, and reduce how much of your plan depends on one platform's recommendation system. Everything else about the work stays the same: brief, takes, pick, fix, finish, document.
Where does licensing fit into all of this?
Licensing is upstream of every platform rule and unaffected by intake volume. Read the terms of the tool you generated with before the track ships, and keep the receipt or the terms page with the disclosure file. Reading an AI music license and the pre-upload checklist both cover this in detail.
Bottom line: the number is real enough and it is about supply, not demand. Take it as a signal that paperwork and owned distribution went up in value, keep a dated record of how each track was made, and spend your remaining effort on the layer where your tracks actually stop.