DataPacts

Consent and Compensation for Creators: Opt-in, Revenue Splits and Settlements

Checked

At a glance

Creators are asked, named and paid for AI training only on some routes: an opt-in license or collective can deliver all three, while opt-out regimes, platform settings and unlicensed training often deliver none, and a settlement pays without having asked.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

This page takes the creator side and weighs it against the other side. It belongs to the reference on AI training data licensing, where not licensing is a valid outcome for a creator and for a catalog holder.

DataPacts Editorial Team checked each figure, docket and statement here against primary sources dated 2026-11-09, and re-read the STIM study wording on 2026-11-17; the page on how we source, date and check every figure sets out the method.

Creators and their organizations ask for three things: to be asked before training, to be named where their work is used, and to receive most of the money a license earns.

  • Consent: STIM's AI license covers only works whose creators explicitly opted in, and from 2026-06-01 its members approve works one by one or exclude the AI rights category.
  • Credit: STIM requires its licensees to use an attribution provider such as Sureel, while calling attribution "a relatively new and untested method" with analogy-based distribution as the fallback (checked 2026-11-17).
  • Compensation: the Authors Guild states that a fair split gives 75 to 85 % to the author (2024-12-12), an advocacy position and not a market rate.

The fee split between a publisher and an author is covered in depth under book publishers and authors licensing books.

Only opt-in routes ask creators first, attribution appears in a few licenses, and payment ranges from a stated split to nothing at all. The table compares nine routes as of 2026-11-09, ordered from most to least creator say, and each cell is what the primary source states.

Route Example on record Creator asked first? Creator credited? Creator paid? Source
Individual opt-in addendum HarperCollins nonfiction license (Authors Guild, 2024-11-19) yes, separate addendum not stated yes, 50 % of USD 5,000 per title authorsguild.org
Collective, opt-in work by work STIM AI license, from September 2025; work-by-work approval from 2026-06-01 yes attribution provider required; analogy-based fallback royalties at training, use and output; no rates stim.se
Collective, opt-in via authors' society CLA Generative AI Solution, authors via ALCS, publishers via PLS yes not stated not published cla.co.uk; pls.org.uk
Collective with opt-out, possible extended effect Kopinor agreement, 2025-12-19, s.16 and s.3 members via Kopinor; others only if extended not stated NOK 45,000,000 for 2026 to the collective; distribution not stated kopinor.no
Exception with machine-readable opt-out DSM Directive Art. 4(3); AI Act Art. 53(1)(c) no; creator must reserve no no eur-lex.europa.eu
Platform setting YouTube third-party training, off by default yes, creator turns the setting on no no: YouTube "isn't facilitating payments" support.google.com
Library contributor fund Shutterstock Contributor Fund (2022-10-25); Getty Images 10-K FY2025 not stated; Shutterstock inclusion automatic not stated yes, no rate published submit.shutterstock.com; sec.gov
Class settlement Bartz v. Anthropic, final approval 2026-07-20 no, training preceded the claim no approximately USD 3,000 gross per work settlement docket
Training without a license held fair use Bartz Dkt. 231 (purchased books), Kadrey Dkt. 598 no no no court PDFs

As of 2026-11-09. "Not stated" means the primary source says nothing on the point; no row is a recommendation.

No per-member payout under the CLA, CCC or STIM collective AI licenses is on record. Collective terms, rates and eligibility sit under collective licensing for AI training.

Opt-in Versus Opt-out: Who Carries the Burden of Saying No

Under opt-in the user of the work must obtain a yes, and under opt-out the creator must say no in a form a machine can read, so the burden moves from developer to creator. HarperCollins, STIM and CLA are opt-in routes, and each starts with a creator's separate agreement.

The EU exception works the other way. Article 4(3) of Directive (EU) 2019/790 lets a rightholder reserve rights over online content only in a machine-readable way, and Article 12(3)(c) requires that rightholders can at any time easily and effectively exclude their works from extended collective licensing.

Article 53(1)(c) of Regulation (EU) 2024/1689 obliges general-purpose model providers to respect such reservations from 2025-08-02. Norway's Kopinor agreement of 2025-12-19 lets a rightholder opt out with effect for future training (s.16), and extends to non-represented holders only if approved under Copyright Act s.63 (s.3).

In Germany, the OLG Hamburg held on 2025-12-10 (5 U 104/24) that a natural-language reservation was not shown to be machine-readable in the second half of 2021; the BGH pronouncement in I ZR 281/25 is scheduled for 2026-12-17.

An opt-out acts only forward: under the Kopinor agreement, termination does not affect models already trained. A setting can also be off by default and still unpaid, as YouTube's third-party training setting shows, since YouTube states that it "isn't facilitating payments" between third-party companies and creators (checked 2026-11-09).

The rules for reserving rights are set out in EU text and data mining exception and the opt-out.

Credit: Where a Creator Is Named in a Training License

Credit is the least developed of the three demands: no book or press license on record names the creator in model outputs, and attribution appears mainly where it drives payment.

STIM requires an attribution provider, with analogy-based distribution where attribution is not a reliable basis, and RSL Collective publishes an attribution license under which attribution links replace payment, next to a paid license with royalties per AI result and unpublished rates (checked 2026-11-09).

Bria states that each licensed asset carries an identifier linking it to source, creator and license terms, and that works exceeding relevance thresholds earn ongoing compensation, a claim with no rate (checked 2026-11-09). Fairly Trained announced 9 certified generative AI models at its launch on 2024-01-22 (post by Ed Newton-Rex).

How attribution could meter payment per use is the subject of training data attribution and per-use payment models.

Revenue Splits Offered to Creators, as Each Party States Them

The splits on record run from 50 % to the creator in a publisher or platform deal to 85 % in one licensing platform's standard terms, and every figure is the stating party's own. The table lists ten statements alphabetically by stating party.

Stating party Route Creator share as stated Date Nature Source
Authors Guild recommendation for author and publisher 75 to 85 % to the author 2024-12-12 advocacy position authorsguild.org
Authors Guild on HarperCollins publisher opt-in license 50 % of USD 5,000 per title, not set against advances 2024-11-19 license term reported by an authors' body authorsguild.org
Credtent licensing platform, standard terms 85 % to creatives, rights holders and data providers, quarterly checked 2026-11-09 platform claim credtent.org
GEMA PLAI catalog "anteilig und angemessen" (proportionately and appropriately), no figure checked 2026-11-09 collective statement gema.de
Getty Images contributor compensation "on a recurring basis", no rate FY2025 10-K filing sec.gov
Newsflare AI licensing of uploaders' video 50:50 revenue share checked 2026-11-09 platform claim learn.newsflare.com
Shutterstock Contributor Fund compensation where contributor IP was used, no rate checked 2026-11-09 library claim submit.shutterstock.com
STIM collective AI license fixed fee plus a share of the AI service's income, no rate checked 2026-11-09 collective statement stim.se
Troveo content owner licensing royalty each time content is used, no rate; USD 50M+ "to creators globally" vs USD 20M+ "Paid to Licensors" checked 2026-11-09 platform claim, contradiction kept troveo.ai
Wirestock commission-based projects 50 % commission per license 2026-07-13 post platform claim wirestock.io

As stated by each party; a split is a share of the intermediary's receipts, not of the buyer's price; no figure is a market rate.

Three cautions apply. A split says nothing about the size of the fee it divides, Troveo's two totals on its own pages differ and are both shown, and the Authors Guild called the HarperCollins split "far too much to the publisher" (2024-11-19). Program terms for individual creators are covered under creators selling content for AI training.

A settlement pays creators for training that already happened without their consent, so it delivers compensation but neither consent nor a license for the future. In Bartz v.

Anthropic (N.D. Cal. 3:24-cv-05417) the court gave final approval and entered judgment on 2026-07-20 (Dkt. 680) to a non-reversionary USD 1.5 billion settlement on a Works List of 482,460 works, with a gross estimate of approximately USD 3,000 per work.

The only per-work figure from a distribution is approximately USD 2,203.56 per Claimed Work across 449,731 Claimed Works, in the initial distribution set out in Class Counsel's status report of 2026-09-02 (Dkt. 688); the settlement website still expected first payments on or before 2026-11-15 (checked 2026-11-17), and the figure excludes Anthropic's final USD 450 million payment and is not a final total.

The per-work share is divided among the copyright owners by elected default splits for non-education works, or by contract or publishing agreement, so an author's share can be less than the per-work figure.

A settlement is a different agreement from a license. UMG stated on 2025-10-29 that the companies settled the copyright litigation against Udio and added new recorded-music and publishing licenses, with terms not disclosed. Claims, rulings and timing are tracked under the Bartz v. Anthropic settlement.

What Creators and Their Organizations Have Said on the Record

Creators' positions on record come from authors' and collecting societies' statements, a government consultation, a European Parliament resolution and lawsuits, and each states what its source measured. The table is chronological and attributed.

Date Who Position or finding Forum or document Source
2024-11-19 Authors Guild HarperCollins 50-50 split "far too much to the publisher" Guild statement authorsguild.org
2024-12-12 Authors Guild a fair split gives 75 to 85 % to the author; standard trade contracts grant no AI training rights Guild statement authorsguild.org
filing date not stated YouTube creators (putative class) Millette v. OpenAI over transcripts, member of MDL No. 3143, pending as of 2026-11-09 N.D. Cal. 5:24-cv-04710 jpml.uscourts.gov
2025-11-11 GEMA judgment on 9 song lyrics against OpenAI, not final LG München I, 42 O 14139/24 justiz.bayern.de
2025-12-10 a photographer (Kneschke) appeal dismissed; BGH pronouncement scheduled 2026-12-17 OLG Hamburg 5 U 104/24 bundesgerichtshof.de
2026-03 UK consultation respondents 81 % of 11,520 responses preferred option 1 (strengthen copyright, licensing required); 10 % option 0 Report on Copyright and AI assets.publishing.service.gov.uk
2026-03-10 European Parliament non-binding call for sectoral voluntary collective licensing, remuneration for past uses and a rebuttable presumption of use P10_TA(2026)0066 eur-lex.europa.eu
2026-10-29 Authors Guild and authors (class plaintiffs) summary-judgment motions briefed, OpenAI opposition filed; no ruling as of 2026-11-09 S.D.N.Y. 1:23-cv-08292 (MDL 3143) archive.org (Dkt. 2312)
checked 2026-11-17 STIM (citing Goldmedia for GEMA and SACEM; CISAC and PMP Strategy) 71 % of members likely negatively financially affected, losses up to 27 % by 2028; up to 24 % of creators' incomes globally at risk STIM FAQ stim.se

As of 2026-11-09 unless a row states another date; each row is the named party's own position or finding.

The survey-style figures in the last row are the cited studies' own, reported through STIM, and this page states no independent survey of creators' consent or payment preferences. The government response to its consultation is covered under UK copyright and AI training rules after the 2026 report.

The Case Against Asking and Paying Every Creator

The case against per-creator consent and payment rests on three points found in primary sources: courts have held some training lawful without a license, governments prefer a voluntary market to mandated payment, and most training data is not licensed at all.

In Bartz (Dkt. 231, 2025-06-23) the court held training on purchased books "exceedingly transformative" and fair use, while pirated library copies were not excused.

In Kadrey v. Meta (N.D. Cal. 3:23-cv-03417-VC, Dkt. 598, 2025-06-25) the court said plaintiffs are "not entitled to the market for licensing their works as AI training data"; it called market dilution "potentially winning" but undeveloped and said the ruling "does not stand for the proposition that Meta's use ... is lawful".

Advocate General Szpunar's opinion in C-250/25 of 2026-11-03 finds the training questions inadmissible and says, in an obiter passage that is not binding, that training is as a rule permitted under the EU text and data mining exception.

Japan's Copyright Act Art. 30-4, in force since 2019-01-01, permits such uses unless they unreasonably prejudice the owner.

The US Copyright Office's pre-publication Part 3 report of 2025-05-09 (p. 106) says to let the licensing market develop without government intervention, with targeted extended collective licensing where market failures are shown, and (p. 52) that knowing use of pirated works weighs against fair use without being determinative.

The UK report (Section G, para. 63) proposes no intervention in the licensing market at this stage, although it calls the opt-out exception "no longer the government's preferred way forward".

On the third point, OpenAI states that less than 1 % of its pre-training datasets come from licensed third-party datasets (para. 159, PIPEDA-2026-002, 2026-05-06), a statement the Canadian regulators could not verify.

The objection that licensing every creator is impractical is an argument made on behalf of developers, not a finding of any source here, and DataPacts gives no verdict on it. Each point carries its limit; the legal position is set out under fair use and AI training in US law.

License, Join, Reserve, Claim or Do Nothing: Benefits, Risks, Downsides and Alternatives for Creators

For a creator, a license or collective can secure consent and some payment, but every route has a cost, and doing nothing is a valid choice. The lists describe each route without recommending one.

Benefits

  • A say before use under an opt-in addendum or collective (HarperCollins, STIM).
  • A stated share, such as 50 % of USD 5,000 per title (2024-11-19) or 85 % under Credtent's terms (checked 2026-11-09).
  • Credit where attribution drives royalties (STIM, RSL Collective).

Risks

  • A reservation that is not machine-readable may not count (OLG Hamburg, 2025-12-10).
  • Trained models keep what they learned after an opt-out (Kopinor agreement).
  • The intermediary sets the split (Newsflare, Wirestock).

Downsides

  • No per-member collective payout is published.
  • Settlements pay only for the past and are divided among owners.

Alternatives

  • An individual opt-in license.
  • An opt-in collective.
  • A machine-readable reservation under Directive (EU) 2019/790 Art. 4(3).
  • A claim in a class settlement.
  • Not licensing at all and keeping platform settings off.

Other options are compared under alternatives to selling data to AI labs, including not selling at all.

How the Creator View Fits the Ethics of Selling Data

The creator view is one of four ethics questions about licensing data for AI training, next to the employee view, the privacy of people in the data and reputational risk. The parent page covers all four under ethics and risk of selling data for AI training.

What This Page Does Not Cover

This page does not cover creators using AI tools to earn money or the copyright status of AI outputs. Contributors paid for newly produced data belong under human data for AI. The page is information, not legal advice.

Frequently Asked Questions

Should AI companies pay creators whose work trains their models?

Creators' organizations say yes and ask for most of the fee, while US courts have so far held some training to be fair use without payment, so the answer depends on the route and the jurisdiction. The Authors Guild states a fair split of 75 to 85 % to the author (2024-12-12), and Bartz (Dkt. 231) and Kadrey (Dkt. 598) held training fair use on their records.

Is opt-in or opt-out better for creators?

Opt-in gives the creator the first word and usually comes with payment, while opt-out leaves training allowed until the creator reserves rights in a form machines can read. STIM covers only opted-in works, whereas the EU exception requires a machine-readable reservation under Article 4(3).

Does a settlement give an AI company a license to the works?

No settlement on record is described as a license for future training; UMG paired its Udio settlement with separate new licenses, which shows the two are distinct agreements. The Bartz settlement pays for past conduct.