DataPacts

Training Data Attribution and Per-Use Payment Models

Checked

At a glance

Training data attribution estimates how much individual training examples shaped a particular output of an AI model, and a small number of buyers, platforms and collectives now say they use such estimates to pay data holders per use.

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

The research is nearly a decade old in its modern form, since influence functions were applied to neural networks in 2017, and it has since been scaled to large language models.

No operator recorded on this page publishes an attribution method, threshold or rate behind its payments, so the page keeps the research record and the payment claims apart. Both sit within AI training data licensing, where the question for a licensor is whether a payment tied to influence is measurable at all.

What training data attribution estimates, and what it leaves out

Attribution answers a counterfactual: how would this output change if a given training example had been left out or weighted differently? The field separates three questions:

  • Influence on one output: how much one example moved one prediction (influence functions, TracIn).
  • Effect of a subset: how a model trained on part of the data behaves (datamodels).
  • Value to overall performance: what each example is worth to the model as a whole (Data Shapley, a valuation measure rather than a per-output one).

Exact answers need retraining without the example, which the research literature calls infeasible at model scale, so every method is an approximation. Influence functions need gradients and Hessian-vector products from the trained model.

Attribution also differs from similarity search, which finds look-alike examples but, in the words of one paper, does not capture broader contribution, and from provenance, which records who held which right, a topic that belongs to training data provenance and the chain of rights.

Six attribution methods from the research record, compared

Six methods recur in the research behind attribution-based payment, and each estimates something slightly different. The table lists them by first publication date, as of 2026-11-17, from the publisher or arXiv page of each paper.

Method Authors First published, venue What it estimates Stated limit or scale Source
Influence functions Koh, Liang ICML 2017, PMLR 70:1885-1894 effect of up-weighting one training point on a prediction, via gradients and Hessian-vector products theory assumes convexity; authors report approximations still useful for non-convex models proceedings.mlr.press/v70/koh17a.html
Data Shapley Ghorbani, Zou ICML 2019, PMLR 97:2242-2251 value of each training datum to predictor performance Monte Carlo and gradient-based estimates proceedings.mlr.press/v97/ghorbani19c.html
TracIn Pruthi, Liu, Sundararajan, Kale arXiv 2020-02-19; NeurIPS 2020 change in a test point's loss across training, from saved checkpoints any model trained with stochastic gradient descent or a variant arxiv.org/abs/2002.08484
Datamodels Ilyas, Park, Engstrom, Leclerc, Madry arXiv 2022-02-01 model output when trained on a subset of the data needs many trained models; linear datamodels predict well arxiv.org/abs/2202.00622
TRAK Park, Georgiev, Ilyas, Leclerc, Madry arXiv 2023-03-24 attribution for large differentiable models authors: can match methods that train thousands of models arxiv.org/abs/2303.14186
EK-FAC influence functions Grosse, Bae, Anil et al. (17 authors) arXiv 2023-08-07 influence of pretraining sequences on language model outputs up to 52 billion parameters; influence near zero when phrase order is flipped arxiv.org/abs/2308.03296

TracIn is the method behind the query "estimating training data influence by tracing gradient descent", which is the title of the paper by Pruthi, Liu, Sundararajan and Kale.

It follows how the loss on one test point changes across saved training checkpoints, and the authors describe it as applicable to any model trained with stochastic gradient descent. Venues are given only where the publisher page or the arXiv comments field names one, so Datamodels and TRAK carry arXiv dates alone.

What attribution research has shown, and what it has not

The papers show that influence can be estimated at large scale, and they also show that the estimates can be wrong for deep networks and sensitive to small changes in the input. Each result below is the authors' finding, not an independent proof.

Shown in published papers

  • Grosse et al. report that influence functions scaled with EK-FAC reach language models of up to 52 billion parameters, with accuracy similar to traditional estimators.
  • The TRAK authors state that it can match the performance of attribution methods that require training thousands of models.
  • Ilyas et al. report that simple linear datamodels predict model outputs well in their experiments.

Not shown, or shown to fail

  • Basu, Pope and Feizi (ICLR 2021) find influence estimates fairly accurate for shallow networks but "often erroneous" for deeper ones, and call influence functions in deep learning "fragile".
  • Grosse et al. report that influences decay to near zero when the order of key phrases is flipped.
  • Wang, Efros, Zhu and Zhang (ICCV 2023) evaluate text-to-image attribution through customized models tuned to exemplars, a benchmark rather than proof on production models.
  • Royalty frameworks built on attribution, from Wang, Deng, Chiba-Okabe, Barak and Su (arXiv 2024) and from Deng et al. ("Generative Content ID", arXiv 2023, revised 2025), are preprints with no venue listed, and the second argues that similarity-based proxies are poorly suited to royalty distribution.

None of the operators in the next section names a paper, method or independent evaluation behind its payments. That is a statement about what the pages checked do not publish, not a statement about what the operators use.

Per-use and attribution-based payment models as operators state them

Nine operators on record are listed below; eight state a per-use, per-output or influence basis, and none of them publishes a rate tied to attribution. The table is alphabetical, and each row carries its own check date; the wording is the operator's own, so every row is a claim.

Operator Model Payment basis (operator's statement) Rate published Training use Checked Source
Bria AI Licensed Training Catalog and attribution program "Works exceeding relevance thresholds earn ongoing compensation"; "Compensation ties to influence, not catalog size"; attribution vector per generation, under 50 ms no trains its models only on licensed data from image libraries including Getty Images, Alamy and Envato (Bria's claim) 2026-11-09; re-read 2026-11-17 bria.ai/licensed-training-catalog; bria.ai/attribution-technology
Cloudflare Pay per crawl (closed beta) price per crawl request set by the site owner; HTTP 402; Cloudflare is Merchant of Record; payouts monthly; fees not stated set by each site owner not distinguished from other crawling 2026-11-09 (page updated 2026-07-28) developers.cloudflare.com
Getty Images contributor compensation for AI training use partnered on models that "compensate creators on a recurring basis for the use of their content as training data"; over USD 220 million royalties to contributors in 2025, all products no yes, stated in the 10-K 10-K FY2025 filed 2026-03-16 sec.gov
Human Native (Cloudflare) AI data marketplace not recorded; acquisition by Cloudflare announced 2026-01-15, price not disclosed not recorded not recorded 2026-01-15 blog.cloudflare.com
Labelbox Prism data licensing upfront fee at signing plus revenue share; "Models already trained on the data cannot be unlearned" no yes 2026-11-09 prism.labelbox.com/faq
ProRata (Gist) attribution for AI answers contribution scores "can inform compensation"; "Payment depends on the relevant rights, licenses, and commercial agreements" no "never used for model training" 2026-11-09 prorata.ai
RSL Collective Collective License: AI Training (Paid and Attribution) paid: royalties "whenever an AI application uses your content to generate a result"; attribution: attribution links instead of payment no permits AI training 2026-11-09 rslcollective.org/license
STIM collective AI license for music royalties during training, during use of the AI service and when generated music is used; independent attribution technology required no yes, opted-in works only 2026-11-09 stim.se
Troveo licensing aggregator "Every use earns you a royalty"; paid-out claims of USD 50M+ ("paid to creators", content-owners page) and USD 20M+ ("Paid to Licensors", business-data page), both Troveo's no licenses to AI labs 2026-11-09 troveo.ai

Bria is the only operator here whose own pages tie compensation to measured influence on each generation. Its catalog adds that "Unique content earns more" and "Common concepts share revenue", and it calls its engine "patented" without giving a patent number (Bria's statements, read 2026-11-17).

Bria also states that results "can be visually demonstrated and audited", with no procedure published. Program terms and reputation sit on the profile of Bria AI as a data buyer.

Influence, output or access: three bases for paying data holders per use

Per-use payment rests on one of three events: measured influence on an output, the output or answer itself, or access to the content.

Influence is the basis Bria and STIM state. STIM says creators of included works receive royalties during model training, during use of the AI service and when the music it generates is used, and it requires licensees to use independent attribution technology.

Its own AI license page calls attribution "relatively new and untested" (both pages read 2026-11-16). Collective terms are compared under collective licensing for AI training.

Output is the basis RSL Collective states: its paid license promises royalties whenever an AI application uses the content to generate a result, and its attribution license trades training for attribution links (checked 2026-11-09).

Answers are the basis ProRata states: its contribution scores can inform compensation for AI answers, and it says publisher content supplied for attribution is never used for model training, so its model sits in answers, not in training (checked 2026-11-09).

Access is the basis of three others. Cloudflare charges per crawl request at a price the site owner sets, with Cloudflare as Merchant of Record. Troveo states that every use earns a royalty, without a published rate.

Labelbox Prism pairs an upfront fee with revenue share. None of the three measures influence, and crawler charging itself is covered under pay-per-crawl and crawler licensing.

What a licensor cannot check under an attribution-based payment

A licensor paid by attribution depends on a measurement it cannot reproduce, because none of the operators recorded here publishes the inputs. As of the check dates above, the pages do not publish:

  • the attribution method, or a paper behind it;
  • the relevance threshold (Bria) or the score formula (ProRata);
  • the share of revenue paid out;
  • statement frequency and minimum payment;
  • an audit procedure.

Attribution can measure use but cannot reverse it. Labelbox Prism states that models already trained on the data cannot be unlearned, so ending a license stops future training but not influence already inside a model. The contract is where reporting and audit rights are agreed, as set out under audit rights in AI data licenses.

Benefits, risks, downsides and alternatives of attribution-based pay for a data licensor, including not licensing

Attribution-based pay rewards data that keeps shaping outputs, and it moves the risk of low use and of an untested measurement onto the licensor.

Benefits

  • Income can continue while the data keeps influencing outputs, as Bria and STIM state.
  • Distinctive works can earn more under Bria's stated scheme.
  • Research methods exist to estimate influence at scale.

Risks

  • Estimates can be wrong for deep networks.
  • The measurement is the buyer's own and unpublished.
  • Per-use income depends on the buyer's product succeeding.

Downsides

  • No rate is on record, so an offer cannot be compared with a fixed fee.
  • Influence stays in a trained model after the license ends.

Alternatives

  • A fixed upfront fee or milestone payments.
  • An upfront fee plus a revenue share, the Labelbox Prism model.
  • A collective license.
  • Not licensing at all, or blocking crawlers.

Where attribution meets payment terms, provenance records and the courts

Attribution sits between the provenance record that shows who licensed what and the payment terms that turn use into money, and the pages below take each part.

On the court record, in Doe v. GitHub the Ninth Circuit on 2026-09-16 affirmed dismissal of the DMCA section 1202(b) claim over code reproduced without attribution, holding that the tools create new works that never contained copyright management information; the mandate issued on 2026-10-08, and the case is back in the district court as of 2026-11-17.

Frequently asked questions

What is training data attribution?

Training data attribution is an estimate of how much individual training examples shaped a model output. Because retraining without each example is infeasible at model scale, methods approximate that counterfactual, using gradients, saved checkpoints or many trained models. The estimate is a measurement, not a record of rights.

What is TracIn, and what does "tracing gradient descent" mean?

TracIn is an attribution method from Pruthi, Liu, Sundararajan and Kale, posted to arXiv on 2020-02-19 and published at NeurIPS 2020. It follows how the loss on a test point changes across saved training checkpoints, which is the "tracing gradient descent" of the paper's title, and it applies to models trained with stochastic gradient descent.

Can an AI developer prove which training data shaped an output?

No, the research gives estimates, not proof. The papers report failures for deep networks and sensitivity to phrase order, and no operator on this page publishes an independent evaluation of its attribution.

Does any AI company pay data holders each time their data is used?

Several operators say so, as of their check dates, and none publishes an attribution rate. Bria states a basis in influence, RSL Collective in output, STIM in training, service use and generated music, and Cloudflare in each crawl request.

How this page is sourced and dated

Papers were read on their publisher or arXiv pages, operators' terms are kept as claims with check dates from 2026-11-09 to 2026-11-17, and the court record is as of 2026-11-17. The approach is set out under how we source and date figures. The page is information, not legal or financial advice.