Arooj Aftab's two-payment rule: pay artists when AI trains on them, and every time it uses what it learned

On Tuesday 20 October, Arooj Aftab posted two Instagram stories that boiled the AI-pay argument down to two words: usage and derivative. Artists should be paid because AI has already trained on their music, she wrote, "aka usage". And they should be paid again every time a model uses what it learned from that music to make something new, "aka derivative work". The posts were reported by Dawn Images on 21 October and by The Express Tribune, with screenshots, on 22 October. AIMusicpreneur set the two points out side by side as a pay model.

Aftab won the 2022 Grammy for Best Global Music Performance for "Mohabbat", and AIMusicpreneur notes she was the first Pakistani artist to win one. When an artist with that profile puts a pay model in two lines, it travels.

What did she actually say?

Both stories reshared slides from United Musicians and Allied Workers (UMAW), a US union of musicians and music workers. Over the first, a slide calling for AI music to be labelled, she wrote that "even if you don't use AI to make music - it is going to use your music to make AI music", and added that the rules should be got right (Express Tribune). The second story, over UMAW's slide on paying human artists, carried the two-payment argument (Dawn Images).

Is that what the union is asking for?

Not quite, and the gap is the interesting part. UMAW's statement of 12 October makes three demands: label GenAI music at every level of the industry, from software developers to retailers; have streaming platforms refuse GenAI tracks outright; and pass the Living Wage for Musicians Act so streaming pays human recording artists directly. It describes GenAI music as trained on existing work without consent, but it doesn't ask for that training to be paid for. UMAW's position is closer to "keep it off the platforms" than "license it".

The bill it backs is a streaming-pay bill. As Rep. Rashida Tlaib's office describes it, it adds a small fee to streaming subscriptions, taxes platforms' non-subscription revenue and pays artists from a non-profit fund in proportion to monthly streams, with a cap per track per month. The page, from September 2025, puts Spotify's average per-stream royalty at around $0.003 and doesn't mention AI at all.

So Aftab took a union campaign about labelling and streaming pay and added an AI licensing model to it. That model is hers.

How does it compare with what the industry is proposing?

It runs straight into the other big artist voice on the subject. In his keynote at the UN's AI for Good summit in Geneva, scheduled for 8 July, CISAC president Björn Ulvaeus argued for paying on the way in. According to ITU's write-up, he rejected tracing each output back to its sources and proposed a collective model instead, with a share of AI subscription revenue paid out to the creators whose work trained the models: "Our work went in. We should be paid for what went in."

Aftab's first payment is Ulvaeus's whole model. Her second is the part he set aside.

It isn't hypothetical, either. Swedish collecting society STIM launched a pilot licence in September 2025 under which creators receive royalties during model training, during use of the AI service, and when music generated by it is used. It requires independent attribution technology to link outputs to the human works that influenced them, with Songfox and Sureel as pilot partners (Sureel providing attribution), and it covers only a limited set of works from members who explicitly opted in. That is, in effect, Aftab's two payments, written into a contract, with consent at the front.

What it means for working musicians

The worry behind all of this is income, now. Musicians online describe the same squeeze, though we could not open those threads to check them, so we cite none here. Anecdotes are not data, but they point at who gets hit first: the hired session musicians, demo singers and library composers whose work is easiest for a model to replace.

That is why the second payment matters more than the first. A training fee is paid once, mostly to whoever controls the catalogue, and it's done. If generated music then replaces the session musicians whose recordings taught the model, a one-off fee doesn't cover the income they've lost. A per-output share at least ties what the artist earns to how much the model is used.

For artists, three practical things follow:

• Read any AI clause for both payments. A licence that pays for training and says nothing about outputs is a buyout, whatever it calls itself.
• Ask who the attribution provider is. Per-output pay only works if something measures influence, and that measurement needs to be independent of the AI company paying out, as STIM requires.
• Opt-in beats opt-out. STIM's pilot only includes works whose owners said yes. That should be the baseline everyone else is measured against.

What we think

Aftab's framing is sharper than most industry proposals because it separates two things that licensing deals tend to bundle: access to the music, and a share of what is built from it. Collective input-only models like the one Ulvaeus describes are easier to administer, and that's their appeal. But they cap what creators can earn at whatever the market pays for training data, while the value of a model keeps growing with every track it makes.

Our read is that the fight over the next few years won't be about whether training gets paid; even the input-only camp, Ulvaeus included, accepts that it should. The fight will be about attribution: whose measurement of influence counts, who audits it, and whether it's built into the licence or left out. STIM's pilot shows it can be written in. The test for every new deal between an AI company and a rights holder is simple: does it pay for outputs, and who measures them? If the answer is "no" and "nobody", artists are being offered a one-time payment for a lasting loss of work.

The losers if per-output pay never arrives are not the stars; they have leverage to cut their own deals. They are the session musicians and working players on those Reddit threads, whose recordings are training data and whose jobs are the first thing a model replaces.