DataPacts

Rights in an RL Environment: Cloned Software, Seed Data and Contributor IP

Checked

At a glance

An RL environment carries four separate layers of rights: in the software it copies, in the records that seed its tasks, in the tasks and graders contributors write, and in the finished environment a buyer licenses, and a different statute or contract settles each.

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 sets out those layers within AI training data licensing, so that a holder can see which rule applies before anything is copied into an environment.

Every clause is quoted from the party's own published terms with its check date, and every law carries its section and date. The parent page compares builders' documents; this page takes the inputs one at a time.

The Inputs Inside an RL Environment and the Rule Behind Each

Seven inputs can sit inside one environment, and each is governed by its own rule: product terms for copied software, trade-secret, database and data protection law for seed records, assignment clauses for contributor work, and the builder's or buyer's terms for the finished environment.

The table lists each input with the rule or clause on record, as of 2026-11-09 unless a row is dated otherwise.

Input Whose rights it carries Rule or clause on record Date Source
Replicated third-party software the product's owner product API and data terms, e.g. Slack API Terms: apps distributed outside the organization "must not" use API Data to train a large language model checked 2026-11-09 slack.com
Seed records from the holder or its customers the holder; customers; people named DTSA 18 U.S.C. 1839(5); Directive (EU) 2016/943 Art. 4(4); Directive 96/9/EC Art. 7(1) 2016-05-11; 2016-07-05; transposition due 1998-01-01 govinfo.gov; eur-lex.europa.eu
Synthetic records derived from personal data people in the source data, unless anonymous GDPR Recital 26; EDPB Guidelines 02/2026 (consultation draft) applies from 2018-05-25; adopted 2026-07-07 eur-lex.europa.eu; edpb.europa.eu
Tasks, graders and reference solutions the contributor, unless employed or assigned 17 U.S.C. 101 and 201(b); Shipd Contributor Terms s.3.1; Outlier ToS s.4(b) Outlier ToS effective 2026-05-11; checked 2026-11-09 govinfo.gov; shipd.ai; outlier.ai
Outputs of commercial models used to write tasks governed by the model provider's terms Anthropic Commercial Terms s.D.4 (no use "to train competing AI models"); Gemini API terms (no use to develop competing models) 2025-06-17; 2026-03-23 anthropic.com; aistudio.google.com
Open-source libraries and hubs per the license Prime Intellect verifiers library: MIT checked 2026-11-09 github.com/PrimeIntellect-ai
The finished environment per the build or license agreement Proximal MSA s.8.1: Proximal owns Deliverables; DataVendor: a sale licenses rather than transfers last updated 2026-07-06; checked 2026-11-09 proximal.ai; docs.datavendor.ai

The rights of each input travel separately, so a clean grant on the finished environment does not cure a defect in one input.

For the wider picture, see what RL environments are, who builds them and who buys them.

Replicas of Third-Party Software: Builder Claims, Product Terms and Open Questions

Several builders state that they replicate real commercial software for agents to work in, and none of their pages, checked November 9, 2026, says how the owner of the copied product is cleared. Chakra Labs describes "pixel-perfect cloned enterprise environments".

Deeptune builds "training gyms" replicating enterprise software such as Slack and Salesforce; Mercor announced on 2026-07-09 that it will acquire Deeptune, and closing is not confirmed (re-checked 2026-11-17). Plato turns real-world software and data into simulated environments.

Records inside a replica often come from the copied product through its API or an export, and the product's own terms then apply to that data. Four examples of such terms:

  • Slack API Terms: apps distributed outside the organization must not use API Data to train a large language model, and bulk export is prohibited unless an additional agreement allows it; the update was announced on 2025-05-29 and applies to existing apps from 2025-06-30 (checked 2026-11-09).
  • Google Workspace API user data policy: transferring, selling or using user data to create, train or improve a machine learning or AI model is prohibited, beyond a user's personalized model (checked 2026-11-09).
  • QuickBooks Online terms: training or developing AI or machine learning models with information derived from the platform without authorization is prohibited, and Intuit owns de-identified and aggregated data (checked 2026-11-09).
  • Atlassian Developer Terms s.9.2: developers may not sell any End User Data, effective 2025-12-01 (checked 2026-11-09).

What remains open is the copying itself: no statute or ruling on copyright in a software interface or on use of a product's name in a training replica is stated here, and the record shows only that builders publish no clearance method.

Copying a product by logging into it raises access questions. Van Buren v. United States (No. 19-783, decided 2021-06-03) treats "exceeds authorized access" as a gates-up-or-down inquiry, and hiQ Labs v. LinkedIn (9th Cir.

No. 17-16783, opinion 2022-04-18) held that "without authorization" does not apply to public websites while leaving contract, copyright and trespass to chattels claims open. The page on claims a scraper or replica builder can face sets out these claims.

Seed Records and Synthetic Records Placed in Tasks

Seed records are the emails, tickets, documents and accounts that make a replica look real, and when they are a company's own or its customers' records, trade-secret, database and data protection rules follow them into the task. Three rules apply:

  • Defend Trade Secrets Act, 18 U.S.C. 1839(5): misappropriation includes disclosure by a person who acquired the secret under a duty to maintain secrecy; the Act has been in force since 2016-05-11 and 1836(d) sets a three-year limitation.
  • Trade Secrets Directive (EU) 2016/943 Art. 4(4): acquisition, use or disclosure is also unlawful where the person knew or ought to have known that the secret came from someone using or disclosing it unlawfully; in force since 2016-07-05, transposition due 2018-06-09.
  • Database Directive 96/9/EC Art. 7(1): a maker with a substantial investment may prevent extraction or re-utilisation of a substantial part of a database, a right that Art. 7(3) allows to be transferred, assigned or licensed, for 15 years under Art. 10(1); transposition was due 1998-01-01.

The wider rules appear under trade secrets and confidential information in AI data licensing.

Buyers' own conditions point the same way. Mercor's cybersecurity listing states "customer data only with customer agreement", and its hardware engineering listing states "Designs the company owns, not customer IP" (Mercor's statements, checked 2026-11-09).

Personal data and synthetic records need separate care. DataVendor cleans purchased repositories, raw data and Harbor tasks of personal information within 2 business days, and states that tasksets and environments are not cleaned (checked 2026-11-09).

GDPR Recital 26, in force since 2016-05-24 and applying from 2018-05-25, excludes anonymous information, judged by all means reasonably likely to be used, such as singling out.

The EDPB's Guidelines 02/2026 on anonymisation, adopted for consultation on 2026-07-07 and not final as of 2026-11-17, say that inference "would, in particular, be the case for AI models or synthetic data".

Whether synthetic data is personal data is covered in synthetic data and privacy law; who owns synthetic records is a contract term, taken up below.

Task Writers, Graders and Reference Solutions: How Contributor Work Is Assigned

The person who writes a task, a grader or a reference solution is its first author, so a builder or platform holds the work only through employment or a written assignment.

In the United States, 17 U.S.C. 201(b) makes the employer or commissioning person the author of a work made for hire, and 17 U.S.C. 101 limits that status to an employee's work within the scope of employment or a specially commissioned work in nine listed categories with a signed written agreement.

UK CDPA 1988 s.11(2) makes the employer first owner of an employee's work in the course of employment, subject to agreement, and German UrhG s.69b gives the employer the economic rights in programs written by employees, while a freelancer's code needs an express grant.

The detail for employees, contractors and founders is covered under employee, contractor and founder IP assignments.

The table compares published contributor terms, as of 2026-11-09.

Platform or hub Operator on record Clause What the contributor keeps Source
OpenReward General Reasoning, Inc. "Your environment code stays on GitHub"; no license terms on the release page not stated gr.inc
Outlier Scale AI, Inc. and Smart Ecosystem, Inc. (dba Outlier AI) ToS s.4(b): contributor "irrevocably assign[s] all right title and interest throughout the world" in Work Product; ToS effective 2026-05-11 nothing in the Work Product outlier.ai
Prime Intellect Environments Program Prime Intellect, Inc. stack stated as open-source; verifiers library MIT; no license stated per environment not stated primeintellect.ai; github.com
Shipd Datacurve AI, Inc. per the Contributor Terms (privacy policy names Datacurve, Inc.) Contributor Terms s.3.1: irrevocable worldwide assignment of all Work Product on creation or submission nothing in the Work Product shipd.ai/terms

Open-source routes work differently. Prime Intellect states that its entire stack is open-source and that its verifiers library is MIT-licensed, so environments built on that library stay reusable by anyone, and OpenReward states that environment code stays on GitHub without license terms on the release page.

Who Keeps the Finished Environment: Builder Ownership, License-Back and Exclusive Windows

The published terms split three ways on the finished environment: the builder keeps it and licenses it back, the holder keeps it and licenses it out, or the buyer takes exclusive use for a period. The table compares five documents, as of 2026-11-09 unless a row is dated.

Document Who owns the result What the other side gets Exclusivity Source
DataVendor vendor terms (HUD) the seller "by default"; a sale licenses rather than transfers buyer entitlement survives vendor edits or unpublishing; sales final per the agreement behind the listing docs.datavendor.ai
Labelbox Prism the seller keeps full ownership and continued use buyers pay for stricter exclusivity; kill clause deletion within 30 days; trained models cannot be unlearned exclusive by default prism.labelbox.com
micro1 Data License Master Terms the licensee owns derivatives, models, weights, outputs and synthetic data (s.2.2) worldwide, perpetual, irrevocable, sublicensable license (s.1.2); no-prior-grant warranty (s.5.1) nonexclusive micro1.ai
Proximal MSA (last updated 2026-07-06) Proximal owns Deliverables (s.8.1) and keeps learnings including model training (s.8.3) customer: non-exclusive, non-transferable internal-use license during the Term (s.8.2); IP indemnity by Proximal (s.10.1); materials returned, not deleted (s.6.4) non-exclusive to the customer proximal.ai/msa
Turing Project Lazarus ownership "does not change hands" Turing sells the record under a revenue share license to Turing exclusive for the first 6 months; Turing may license the record to more than one lab in that term; then extend or end the term (checked 2026-11-17) lazarus.turing.com

Published terms only; Plato, Halluminate and Chakra Labs publish no ownership terms for environments built on a holder's software.

Proximal's s.8.2 internal-use license during the Term, beside s.8.1 reuse of Deliverables, is the license-back pattern a holder meets with a builder.

DataVendor words its default two ways: its seller documentation says "the usual starting point is that you keep ownership", and its FAQ answers "Yes, by default" to whether the seller keeps ownership, because a sale licenses the assets rather than transferring them (FAQ checked 2026-11-16). DataVendor lists and hosts assets, and Turing sells the record for the holder.

Exclusivity is priced and worded differently. Two founders quoted by Epoch AI on 2026-01-12 said that exclusivity runs roughly 4 to 5 times the non-exclusive price, a claim and not market data.

Turing's license, exclusive to Turing for the first 6 months, leaves ownership unchanged, and Labelbox Prism is exclusive by default. Exclusive vs non-exclusive data licenses: price, lock-in and carve-outs has the full comparison.

Who owns models, weights, outputs and synthetic data trained on the licensed inputs is a separate clause. micro1's s.2.2 gives them to the licensee, which makes synthetic records a contract question and not a copyright one.

The ownership of derivatives and trained models is covered on a separate page, linked at the end of this page.

Letting Your Software or Records Into an Environment: Benefits, Risks, Downsides and Alternatives

Letting a builder copy software or seed tasks with a company's records can earn license income, but the rights in each input have to be cleared first and the finished environment usually stays with someone else.

Benefits

  • Ownership can stay with the holder under a license: DataVendor, Turing and Labelbox Prism state that ownership stays with the seller.
  • A named internal-use license back is available from at least one builder, Proximal under s.8.2 of its MSA.

Risks

  • Customer records and third-party secrets can travel into tasks, which is why a rights audit before licensing data to AI companies comes first; Mercor's listings already carve out customer data and customer IP.
  • Personal data is not cleaned from tasksets and environments by DataVendor (checked 2026-11-09).
  • The counterparty can change: Mercor announced on 2026-07-09 that it will acquire Deeptune, closing not confirmed as of 2026-11-17.

Downsides

  • Several builders and data programs publish no license or payout terms, among them Plato and Halluminate (checked 2026-11-09).
  • On DataVendor, sales are final with no refunds, and the buyer's entitlement survives edits or unpublishing; under Labelbox Prism, models already trained cannot be unlearned.

Alternatives

  • License the records as records, under a non-exclusive license that keeps ownership.
  • Keep contributions open under a permissive license, as the MIT verifiers library does.
  • Not licensing at all, which leaves the software and records with the company.

The complete process is in the complete seller guide, including not selling.

Where Environment Rights Meet Selling, Contract Clauses and Code Ownership

This page follows rights input by input; selling the environment, the full set of license clauses, environment types and builders each have their own page.

This page summarizes statutes, court records and published terms as of the dates shown; it is information, not legal advice.

Frequently Asked Questions

Who owns an RL environment built on a company's software or records?

Whoever the contract names: under Proximal's MSA the builder owns the deliverables, while on DataVendor the seller keeps ownership and licenses use. Proximal's MSA was last updated 2026-07-06.

Can a builder copy a commercial software product to train agents?

Builders state that they do, but no builder page on record says how the copied product's owner is cleared, and the product's own API terms can bar training on its data. Slack's API Terms are one example (checked 2026-11-09).

Do task writers own the tasks they write for RL environments?

Not on the platforms on record: Shipd's Contributor Terms s.3.1 and Outlier's ToS s.4(b) have contributors irrevocably assign their work product. Outlier's ToS took effect on 2026-05-11. Pay and terms across such platforms appear under AI training jobs and platforms compared.

Is synthetic seed data outside data protection law?

Only if it is anonymous under GDPR Recital 26, and the EDPB's draft Guidelines 02/2026 state that inferences can be drawn from synthetic data. The guidelines were adopted for consultation on 2026-07-07 and are not final.