Attribution is essential and really difficult. Picture Credit score by Michael Whalen
AI might develop into excellent at figuring out affect with out turning into any higher at paying creators.
Attribution has develop into one of many music trade’s favourite solutions to the AI drawback. The promise sounds cheap: establish the recordings and compositions that helped form an AI-generated tune, then credit score and pay the folks behind them.
As a composer, I would like that system to work. I’m not satisfied that anybody has proven the way it will produce significant revenue for many creators.
My skepticism shouldn’t be solely concerning the know-how or the mathematics. It’s also about who’s shaping the dialog. Too few of the folks growing, negotiating and selling AI attribution seem to have the pursuits of artists and rights house owners on the heart of their work.
I’m not suggesting that nobody cares. I’m saying that incentives matter.

Watch this episode of the Artist Growth Podcast for a “tremendous pleasant” introduction to attribution.
AI firms want entry to music. Labels and publishers need new licensing income. Attribution firms need their know-how adopted. Attorneys, consultants and directors profit when agreements are negotiated and new programs are constructed. Most of those contributors are rewarded when a deal closes.
Artists and rights house owners profit provided that they maintain significant consent and management, the charges are ample, the accounting is trustworthy and the cash really reaches them.
That could be a very totally different commonplace.
The attribution dialog can also be troublesome as a result of a number of separate questions are continually being collapsed into one:
-
- Was a piece included within the coaching knowledge?
-
- Did the AI firm have permission to make use of it?
-
- Did the work affect a specific output?
-
- How a lot cash ought to that affect generate?
-
- Who’s legally entitled to obtain the cash?
-
- What stays after each break up, deduction and administrative price?
Attribution might assist reply a few of these questions. It doesn’t reply all of them.
1. Attribution can establish affect however it can not create permission
The phrase “attribution” is getting used so broadly that it could imply nearly something.
Audio fingerprinting appears to be like for an actual or recognizable match. Detection makes an attempt to find out whether or not a bit of music was generated by AI. Provenance data the place a file got here from and the way it was created or modified. Dataset disclosure identifies the fabric used to coach a mannequin.
Attribution makes an attempt one thing harder: estimating which coaching examples influenced an output and the way a lot every one contributed.
There are additionally two totally different moments when attribution can happen. Coaching-time attribution tries to find out how the unique dataset formed the mannequin. Technology-time attribution examines materials deliberately provided when a specific output was created, corresponding to a reference recording, uploaded monitor or licensed artist mannequin.
Technology-time attribution could also be simpler to doc as a result of the system is aware of what the person provided. Coaching-time attribution is rather more troublesome. A mannequin might have realized patterns from tons of of hundreds or tens of millions of examples, and totally different attribution strategies might produce totally different solutions.
Researchers together with Choi and colleagues and Morreale and colleagues have proposed methods to hint coaching affect. The analysis is essential, however there’s nonetheless no universally accepted trade methodology that produces constant outcomes throughout totally different fashions.
Warner Music Group’s 2026 settlement to accumulate Sureel AI reveals the place the most important firms imagine that is going. Warner mentioned the Sureel registry comprises tens of millions of music property and might monitor the usage of parts of a piece in AI programs.
That will develop into useful infrastructure. It nonetheless doesn’t reply the financial query.
A report would possibly present {that a} recording influenced an AI-generated tune. It can not show that the AI firm had permission to make use of the recording. It can not set up a royalty charge. It can not resolve how the cash must be divided between the composition and recording. It can not power anybody to pay.
Attribution is proof. Permission comes from a license. Fee comes from a contract or regulation.
These distinctions matter.
2. Scale is the quantity that adjustments every little thing
The know-how is difficult, however my central concern is fundamental math.
Meta disclosed that its MusicGen analysis mannequin was educated on roughly 400,000 recordings representing about 20,000 hours of music owned by Meta or licensed for the mission. That’s one helpful disclosure. Many industrial AI companies nonetheless don’t publish a whole coaching manifest, the variety of mannequin variations they’ve created or the variety of generations their customers produce.
The output aspect is increasing even quicker.
In April 2026, Deezer reported receiving nearly 75,000 absolutely AI-generated tracks on daily basis. That amounted to greater than two million tracks a month and roughly 44 % of its every day uploads.
Deezer additionally mentioned absolutely AI-generated music represented only one to three % of streams and that as a lot as 85 % of these streams have been fraudulent.
These figures cowl one streaming service. They don’t embody the a lot bigger variety of generations that stay inside AI platforms, are discarded, are revised repeatedly or are utilized in non-public tasks.
The dimensions creates two totally different denominators.
The primary is the variety of works used to coach a mannequin. The second is the variety of outputs generated from that mannequin. Each numbers could also be monumental.
A creator can not consider an attribution proposal with out understanding:
-
- The overall sum of money coming into the system.
-
- The variety of recordings and compositions sharing that cash.
-
- The variety of outputs being counted.
-
- Which outputs generate income.
-
- What degree of affect triggers fee.
-
- Which deductions are taken earlier than creators are paid.
A proportion can sound beneficiant whereas representing nearly nothing. With out the underlying numbers, the proportion is advertising and marketing.
3. One monitor can produce a crowd of claimants
A recorded tune usually comprises a minimum of two separate copyrighted works.
The musical composition covers the music and lyrics. The sound recording covers the recorded efficiency and manufacturing. The U.S. Copyright Workplace treats them as separate works. Permission to make use of one doesn’t robotically embody permission to make use of the opposite.
That distinction is simple to state and troublesome to manage.
A sound recording could also be owned by an artist, a document firm or a number of events. A composition might have a number of writers, publishers and directors. Their possession percentages might differ by territory. Some shares could also be disputed or lacking.
The mannequin might study from melody, concord, rhythm and lyrics. It could additionally study from a singer’s phrasing, the sound of an instrument, a producer’s selections, an engineer’s work or the timbre of a specific recording.
Voice, identify and likeness create one other set of rights that will not be lined by both the composition or sound-recording license.
The present royalty system already reveals how fragmented fee can develop into. The Mechanical Licensing Collective administers one class of U.S. digital mechanical royalties. Efficiency rights organizations, SoundExchange, labels, publishers and different directors deal with totally different rights and makes use of.
Attribution would possibly establish a supply recording. Earlier than that info can produce a fee, the system nonetheless has to find out:
-
- Who at the moment owns that proper.
-
- Which territories are lined.
-
- Whether or not the use was licensed.
-
- How co-owners divide the fee.
-
- What occurs when possession is disputed.
-
- What occurs to unmatched cash.
An attribution rating shouldn’t be a royalty assertion.
4. Correct fractions can nonetheless add as much as nearly nothing
The toughest query shouldn’t be whether or not a system can calculate affect. It’s whether or not sufficient cash will enter the system to make that calculation economically significant.
The U.S. Copyright Workplace addressed this disagreement in its 2025 report on generative AI coaching. Some commenters argued that funds to particular person creators could be negligible due to the variety of works concerned. Others argued that small funds can accumulate over time.
Each statements could also be true. Neither tells a person creator what to anticipate.
Take into account a hypothetical $100 million annual pool:
-
- Divided amongst 100 million outputs, the typical is $1.00 per output.
-
- Divided amongst 10 billion outputs, the typical is $0.01, or one cent, per output.
-
- Divided amongst 100 billion outputs, the typical is $0.001, or one-tenth of a cent, per output.
It is a stress check, not a forecast.
At 10 billion outputs, the pool gives a mean of 1 cent per output. If an attribution system connects that output to 1,000 supply works, the typical turns into $0.00001 per work.
That quantity should then be divided between the composition and sound recording. Both sides might have a number of house owners. Performers, directors and accumulating organizations might also have claims or deductions.
Affect weighting wouldn’t divide the cent equally. It could in all probability direct more cash towards a small variety of works. That will profit the highest-scoring materials whereas leaving a lot of the coaching catalogue with even much less.
A pure professional rata system additionally transfers the danger of scale to creators. An AI firm can generate extra outputs with out rising the whole sum of money within the pool. Each new output then reduces the typical worth of the prevailing claims.
The corporate’s product turns into extra productive. The creator’s attributable share turns into much less useful.
Extra correct accounting can not restore an insufficient charge. It might probably solely divide the insufficient charge extra exactly.
5. The fee system will likely be in-built contracts quite than code
The primary licensing experiments are starting to point out what an AI fee system would possibly appear to be.
STIM, the Swedish accumulating society, says its AI license compensates rightsholders individually for coaching and generated output. Coaching income is split equally among the many musical works included within the license. Output income is distributed via attribution or, when attribution is unreliable, via a proxy primarily based on how works have been used within the coaching knowledge.
It is a critical try to unravel the issue. It additionally demonstrates the issue.
Equal division is easy to manage, however the math turns into acquainted if tens of millions of works share a modest pool. Proxy distribution could also be essential when attribution fails, however the selection of proxy can decide who receives the cash.
STIM additionally says there isn’t any established market commonplace and that it can not but say how a lot a person creator will obtain.
The Nationwide Music Publishers’ Affiliation mentioned its 2026 agreements with Udio and KLAY present publishers with compensation equal to document labels on the coaching aspect. Equal remedy for the tune and recording could be an essential enchancment over the economics of streaming.
It nonetheless doesn’t inform us whether or not the whole fee is ample, how artists and songwriters take part, or what occurs to settlements, advances and fairness acquired by the businesses negotiating the agreements.
A attainable system inside 36 months
By October 2029, I can think about a licensed AI music market constructed round a number of fee swimming pools quite than one common fund.
Labels, publishers, accumulating societies and distributors will negotiate separate agreements. The small print will differ, however the fundamental construction might embody three fee layers.
Fee for coaching
An AI firm pays for the proper to incorporate music in a mannequin. The fee is the higher of an annual minimal assure or an agreed proportion of subscription, promoting, API and enterprise income.
A brand new mannequin or substantial retraining requires a brand new license and one other fee.
Coaching compensation is owed as a result of {the catalogue} helped construct a industrial product. It mustn’t rely solely on whether or not anybody can hint a specific output again to at least one recording.
The coaching fee is split between a musical works pool and a sound-recording pool. Publishers and accumulating societies administer the composition aspect. Labels, distributors and grasp rights organizations administer the recording aspect.
Performers and session musicians will want a protected share or a lot of them might obtain nothing. Impartial artists and writers will in all probability take part via distributors and collective licenses quite than negotiating with each AI firm individually.
Fee for outputs
A separate pool covers generated music that’s launched, monetized or used commercially.
Outputs with a robust and independently verified connection to particular works obtain a bigger share. When attribution confidence is low, the system makes use of a broader distribution methodology quite than dividing a microscopic fee amongst hundreds of supposed sources.
Artist-specific voice fashions, identify and likeness makes use of, remixes, continuations and deliberate imitations require separate consent and better charges.
A shared rights system
The trade will want a clearinghouse connecting recording identifiers, composition identifiers, possession shares, territories, consent and mannequin permissions.
DDEX already gives requirements for exchanging recording and musical works knowledge and is analyzing metadata for AI-generated music. That provides a place to begin, however it isn’t a whole world rights database.
Below a reputable system, AI firms would report their datasets, mannequin variations and industrial outputs. Labels and publishers would combination small quantities and problem quarterly funds. Conflicting possession claims and unmatched royalties could be held in escrow.
Creators would obtain statements displaying which service used their work, which license utilized, how their fee was calculated and which deductions have been taken.
That’s the attainable future.
It’s not the identical because the possible consequence.
The folks on the desk will decide the result
That is the place my deepest skepticism begins.
The folks negotiating these agreements typically management the knowledge used to guage them. A creator might by no means see the whole coaching dataset, the whole license price, the attribution formulation, the scale of an advance, the worth of an fairness place or the way in which cash is allotted between recordings and compositions.
AI firms profit from authorized certainty. Labels and publishers achieve new licensing revenue. Attribution suppliers profit when their know-how turns into a part of the system. Attorneys, consultants and directors are paid to create and handle the agreements.
The artist or songwriter will be the final particular person to study what was negotiated of their identify.
This doesn’t imply each participant is appearing in unhealthy religion. It means their pursuits will not be robotically the identical.
Too few folks within the attribution dialog seem to start with the query that issues most to creators:
Does this method depart the artist or rights proprietor materially higher off?
An organization might describe a deal as “honest” as a result of the music was recognized and a fee was calculated. That doesn’t imply the artist consented to the use, understood the formulation or acquired an economically significant quantity.
The Council of Music Makers addressed this danger in its September 2026 steerage. It urged creators to ask whether or not labels and publishers obtain lump sums, advances, settlements, income shares or fairness; how these types of worth are allotted; and whether or not creator funds are utilized towards previous unrecouped balances.
These will not be secondary particulars. They decide the place the cash goes.
Earlier than supporting an AI licensing system, I might need clear solutions to the next questions:
-
- Was every work licensed for a named service and mannequin?
-
- Can an artist or songwriter decline participation or withdraw after the license expires?
-
- Are the composition, sound recording and performer pursuits all lined?
-
- What’s paid for coaching and what’s paid for output?
-
- Is the fee the higher of a minimal assure or an outlined share of income?
-
- How are settlements, advances, fairness, administrative charges and recoupment handled?
-
- Do creators take part in each type of worth acquired by their representatives?
-
- Can an unbiased get together audit the dataset, attribution methodology, income base and royalty statements?
-
- Can creators right possession knowledge, problem a end result and get well disputed or unmatched cash?
If the solutions are confidential, incomplete or managed solely by the events receiving the cash, the system shouldn’t be clear sufficient.
The actual check
I count on licensed AI firms to pay into coaching and output funds. I don’t count on one common fund, and I don’t assume that the primary agreements will likely be honest.
The possible end result will likely be a patchwork of direct licenses, collective agreements and regional guidelines. Some will likely be accountable experiments. Others might use the language of attribution to make an insufficient deal sound equitable.
I’m not skeptical as a result of I imagine attribution know-how is ineffective. Attribution may develop into an important layer of proof and accounting.
I’m skeptical as a result of a technically correct system can nonetheless legitimize an financial association designed with out significant creator participation.
Too few folks within the present dialog are rewarded for saying that the fund is just too small, the phrases are unfair or the creator’s share is meaningless. Many of the industrial contributors are rewarded for getting the system licensed and working.
A database can inform us who could also be owed. It can not resolve what their work is price. It can not assure that artists and songwriters participated within the negotiation. It can not stop confidential advances, fairness, administrative charges and recoupment from absorbing a lot of the worth earlier than creators are paid.
If the fund is insufficient, attribution will divide an insufficient fee with higher precision. An organization might then level to the precision of the calculation as proof that the system is honest.
Earlier than anybody declares the issue solved, artists and rights house owners ought to have a seat on the desk, significant consent, entry to the monetary phrases, unbiased audit rights and a assured share of each type of worth generated by the settlement.
The ultimate query shouldn’t be merely whether or not AI music may be attributed. It’s figuring out who the attribution system was designed to serve…
Sources and knowledge:
Warner Music Group. Warner Music Group Acquires Sureel AI, June 10, 2026.
Meta AI. Open Sourcing AudioCraft, August 2, 2023.
Deezer. AI Generated Tracks Now Characterize 44 % of All New Uploaded Music, April 20, 2026.
U.S. Copyright Workplace. Copyright and Synthetic Intelligence Half 3 Generative AI Coaching, Could 2025.
U.S. Copyright Workplace. Copyright Registration for Sound Recordings Round 56.
The Mechanical Licensing Collective. How It Works.
STIM. AI License Data and Creator FAQs, 2025 and 2026.
Nationwide Music Publishers’ Affiliation. NMPA Inks Offers With Udio and KLAY, June 10, 2026.
Council of Music Makers. Steering for When Contemplating AI Proposals From Labels and Publishers, September 23, 2026.
DDEX. Recording and Musical Works Knowledge Requirements and AI Working Group Data.
Choi and colleagues. Massive Scale Coaching Knowledge Attribution for Music Generative Fashions through Unlearning, 2025.
Morreale and colleagues. Attribution by Design, 2025.



