AI systems rely on human-created work and human judgment, but those contributions can disappear into training datasets, feedback pipelines and platform-held content. The case for paying people is strongest when a system makes material use of an identifiable contribution: record its provenance, obtain permission where required, credit the contributor where practical, and provide fair compensation. That does not mean every person whose work might have appeared in a model’s training data is automatically owed a micropayment, or that an agent should pay a royalty every time it answers.
Who helps an AI agent?
“AI agent” can describe very different systems, from a model that drafts a response to software that takes steps on a user’s behalf. Neither label changes the fact that people contribute at several points in the system’s life. Those contributions are not interchangeable, and neither are the legal or economic claims they may create.
As an Amazon Associate I earn from qualifying purchases.
- Creators and rightsholders make books, images, code, recordings and other material that may be included in training data. Their relationship to a model developer may involve copyright or other rights, a license, or a dispute over whether a particular use is permitted.
- Workers and experts label examples, rank possible answers, correct outputs or provide specialist judgment. This is labor performed for a task; paying for it does not, by itself, settle rights in the underlying works.
- Evaluators and operators assess how an agent behaves after deployment, report failures and help improve its performance. Their contributions may be ongoing rather than a one-time dataset delivery.
The U.S. Federal Trade Commission describes pretraining data as potentially scraped, licensed or obtained from existing services. It also describes human output-ranking in reinforcement learning from human feedback as labor-intensive work that is often outsourced. Those are distinct inputs and call for distinct compensation questions. FTC, Partnerships Between Cloud Service Providers and AI Developers (2025).
What does the cost estimate tell us—and what does it not?
A 2025 position paper by Nikhil Kandpal and Colin Raffel examined 64 large language models released between 2016 and 2024. Under the authors’ wage assumptions, it estimates that paying people to produce the models’ training datasets from scratch would cost 10 to 1,000 times the cost of training the models. This is a modeled replacement-labor estimate, not an invoice, a measured amount already owed, or a universal valuation of every work used in training.
#1 Best Overall
The paper’s argument is deliberately provocative: “This position paper aims to assign a monetary value to this labor and argues that the most expensive part of producing an LLM should be the compensation provided to training data producers for their work.” The range makes the scale of human input hard to ignore, but it cannot by itself establish who qualifies for payment, which use triggers it, or how a particular person’s contribution should be valued. Kandpal and Raffel, “Position: The Most Expensive Part of an LLM should be its Training Data” (2025).
Human judgment still matters after training
Human involvement does not stop when a model is trained. In a study of 86 practitioners working with deployed systems across 26 domains, Melissa Pan and co-authors report that 74% of surveyed production agents depended primarily on human evaluation, 68% executed at most 10 steps before human intervention, and 70% relied on prompting off-the-shelf models instead of tuning model weights. These are findings about the study’s sample, not prevalence estimates for all agents or industries. They show why compensation discussions should include the people who evaluate and operate systems, not just creators whose works may have been used in training. Pan et al., “Measuring Agents in Production” (2026).
What are the main ways to pay?
There is no single payment mechanism that solves every case. Direct pay addresses labor; a license addresses permission to use defined material; a levy or marketplace is a proposed way to pool or organize compensation. They can complement one another, but they are not substitutes for determining what rights and work are involved.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| Approach | Who it can serve | What triggers payment | Main limitation |
|---|---|---|---|
| Direct pay for work | Annotators, raters, experts and other people doing defined tasks | Work performed, such as labeling, ranking or correcting examples | Payment for a task does not automatically license every underlying work. The FTC describes human ranking work as labor-intensive and often outsourced. FTC (2025). |
| License specific material | A rightsholder or contributor able to authorize a defined use | Use under agreed terms, such as a fee or other consideration | Availability and rights vary; a license for one work or use does not settle claims for all data. Licensing is one route for obtaining training data, while UK policy discussions include broader questions about rights and licensing. FTC (2025); UK DSIT (2025). |
| Collective levy or pooled compensation | A broader class of rightsholders or contributors, as defined by a policy scheme | A levy or other pooled contribution tied to covered uses | Consultation responses in the UK report discuss levy systems as a proposal; this is not a universal payment system already in force. UK DSIT (2025). |
| Opt-in contribution marketplace | People who choose to offer material or other contributions under a platform’s terms | A contribution is licensed through the marketplace | benchturn describes its own approach as opt-in, with provenance and informed consent and payment when contributions are licensed. That is the company’s description, not independent evidence of broad adoption or outcomes. benchturn. |
The UK Department for Science, Innovation and Technology’s 2025 report records consultation responses concerning rights reservation, licensing and levy systems. Those responses illustrate competing policy proposals, not a settled global rule that every AI developer must pay every contributor in the same way. UK DSIT, Report on Copyright and Artificial Intelligence (2025).
Rank #3
What would a fair payment system need to solve?
A workable system has to connect an identifiable contribution to a use and a payment without imposing costs that overwhelm the value being distributed. That is a design challenge, not a problem solved simply by adding a payment button to an agent.
- Provenance: keep records that show where material or feedback came from and how it entered a system.
- Permission and rights: distinguish consent to contribute from the legal authority to license a work, and record the scope of any authorization.
- Valuation: set terms for different kinds of contributions without pretending that every item or hour has the same value.
- Privacy and confidentiality: avoid exposing sensitive material or confidential work while documenting its use.
- Reliable claims: guard against duplicate, fraudulent or misdirected claims, and give contributors a way to check use and earnings.
- Proportionate administration: make sure identification, accounting and transaction costs do not consume the compensation intended for contributors.
These requirements also explain why “pay everyone whose work may have been in training” is not a complete policy. A payment scheme needs rules for eligibility, covered uses, evidence, dispute resolution and the difference between a labor payment and a rights license.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Can an agent make the payment itself?
Technically, an agent may be able to carry out a payment on a user’s behalf. That capability does not show that compensation is routine, that a contributor has authorized a transaction, or that payment is legally required in a particular case. The UK Department for Business and Trade’s 2026 report describes consumer uses of agents as early and bounded, and discusses risks involving error, manipulation, transparency, incentives and accountability. It states: “If an AI agent steers, pressures or misleads consumers in ways that harm their economic interests this is likely to be unlawful.” That statement is from a UK government report about consumers and should not be treated as a general legal rule for every jurisdiction. UK DBT, Agentic AI and consumers (2026).
A trustworthy compensation process therefore needs more than transaction capability: it needs clear authorization, an auditable record of why a payment is made, transparent incentives and accountability when something goes wrong.
What should change now?
Developers and platforms should treat human input as something to document and govern, not as an invisible by-product of model building. For defined labor, pay the people doing the work on clear terms. For rights-controlled material, identify the relevant rightsholder and obtain permission where required. For broader classes of contributions that cannot readily be licensed one by one, consider transparent collective approaches while recognizing that their scope and legitimacy require policy choices.
The central case is not that every answer from an AI agent should trigger a royalty. It is that when a system materially depends on a person’s identifiable work or judgment, the relationship should make permission, provenance, attribution where practical and fair payment visible rather than leaving them hidden inside the system.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
