Since 1 June 2026, STIM has required a writer’s approval, work by work, before a Swedish song enters its artificial intelligence (AI) licence. Alongside Sony Music’s roughly 260,000 takedowns of AI-generated deepfakes impersonating its artists, that creates two separate catalogue decisions: whether to authorise training and how to respond to an imitation of an artist.
Consent attaches to the work
STIM’s dedicated AI rights category makes writer approval the entry condition for each song. For an independent label, the practical consequence is that an AI licensing discussion needs to identify the works being offered and the writers who must approve them. Treating the catalogue as a single commercial package does not answer the permission question for each song.
This changes the unit of negotiation. A catalogue owner might assess an offer across its whole business, but STIM’s approval mechanism operates at song level. A writer can therefore evaluate the proposed use of a particular work rather than make one decision covering every composition. A label preparing an offer should organise its review around that same unit.
The useful operational record is a decision attached to a song: approval, refusal or a request for further terms. That gives the business a way to distinguish the catalogue it would like to offer from the works whose writers have authorised participation. The distinction matters before anyone forecasts licensing income, because a commercial proposal cannot substitute for the approval that admits a work to this scheme.
Separate training income from output income
STIM’s AI licence distinguishes training royalties from output royalties. That separation gives writers two revenue questions to ask. What payment attaches to allowing a work into training, and what payment attaches to output? An offer described only as an AI licensing opportunity leaves those questions bundled together.
The mechanism matters because permission and compensation perform different jobs. Approval determines whether the work enters the licence. The royalty terms determine how the writer participates financially. Agreeing to participate does not itself explain the value of either revenue component, so a writer should examine both before approving a song.
For catalogue planning, keep the two components separate in the assessment. Ask the prospective licensing partner to explain the basis of each payment and how the writer can understand the resulting allocation. This is a contract-review decision, not a reason to assign an assumed value to every AI use. Without the proposed terms, a label cannot compare the offer with its own objectives for that work.
Before seeking approval, prepare a short review for each song:
- Identify the work and the writer whose approval the licence requires.
- Set out the proposed training royalty and output royalty separately.
- Explain what withdrawing permission would change.
- Record the writer’s decision against that specific work.
Withdrawal changes the next decision
Revocation under STIM’s AI licence runs forward. Writers should therefore assess entry with the withdrawal rule already in view. A later change of mind affects the licence prospectively; it should not be treated as an undo button for the original approval.
That makes the timing of review part of the commercial decision. Resolve questions about the scope of permission and the royalty structure before the song enters the licence. The ability to revoke later gives the writer another decision point, but it does not replace scrutiny at entry.
Independent labels can apply this by keeping the approval date, approved terms and any subsequent withdrawal together in the work’s records. The purpose is practical: the person negotiating the next deal should be able to distinguish an earlier authorisation from the permission available now. A catalogue strategy can change while the history of its licensing decisions remains relevant.
Impersonation needs its own response
Sony Music’s roughly 260,000 deepfake takedowns concern AI-generated impersonations of its artists. That count measures removal activity. It does not provide a royalty rate or a valuation for an independent catalogue, and it should not become an input in a licensing forecast.
The rights question also differs. STIM’s mechanism asks whether a writer approves a song entering an AI deal. An impersonation dispute asks how an artist’s identity is being used. A label should keep those questions separate when reviewing an AI proposal: permission concerning a composition should not be read as an answer to every identity issue.
Dani Deahl’s discussion of AI impersonation addresses differences in state publicity rights and the NO FAKES Act’s proposed federal approach. The proposal’s status matters when choosing an enforcement route: a proposed federal approach is not a current federal remedy. Artists and labels should establish the applicable identity-rights basis before selecting their response.
The actionable catalogue decision is to maintain two records: song-level AI licensing permissions and artist-level impersonation concerns. Use the first to decide which works can enter a deal and on what financial terms. Use the second to organise an enforcement response when an imitation appears. A writer’s training decision and an artist’s identity protection can involve the same business, but each needs its own authority, records and action.
