DataPacts

Game and Gameplay Data for AI and World Models

Checked

At a glance

Gameplay data is the record of people playing games, captured as video with the keyboard, mouse or controller inputs behind it, and AI developers license it together with game worlds and builds to train world models and agents.

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

Origin Lab states USD 100 to 100,000 per game (checked 2026-11-09, re-read 2026-11-16) and Mercor states USD 100,000 to 1,000,000 per studio (re-read 2026-11-16). No gameplay licensing deal or court case is on record in the DataPacts deal and case trackers as of 2026-11-16.

This page sets out, within AI training data licensing, what gameplay data covers, who acquires it, the stated terms, the privacy and EULA limits and the option of not licensing.

The DataPacts Editorial Team explains how we source, date and check every figure.

What Gameplay Data Means for AI Training

Gameplay data covers five kinds of material: gameplay video, recordings synced with player inputs, playtest and analytics telemetry, game worlds and 3D assets, and the game build or code itself.

The table lists each kind with the acquirer pages that name it, as of 2026-11-16.

Kind of game data What it contains Acquirer pages that name it Source
Gameplay video full-length gameplay videos and trailers of the studio's own games Mercor Game Studio listing; Sieve "gameplay" Mercor listing feed; sieve.ai/partners
Input-synced recordings gameplay video synced with keyboard, mouse or controller input; input telemetry, frame-level metadata, action annotations Mercor; Origin Lab FAQ Mercor listing feed; originlab.ai/faq
Playtest and analytics telemetry player analytics, playtest telemetry, support tickets Mercor Mercor listing feed
Game worlds and 3D assets environments, character models, meshes and materials; design documents Origin Lab; Sieve "virtual environments, 3D assets"; Mercor "Design documents and 3D assets"; Claru "synthetic/game-engine data" originlab.ai/indie; sieve.ai/partners; claru.ai/partnerships
Builds and code released games; game and engine code with commit history Patronus AI "released games"; Origin Lab "source code and commit history"; Mercor data.patronus.ai; originlab.ai/indie

A studio's own company records, such as chat, tickets and shared drives, form a separate data set that the page on game studios licensing data and assets covers. The full list of data types AI companies buy shows where gameplay sits among the others.

Why World Models and Game Agents Train on Gameplay

Games give AI developers what ordinary video lacks: every frame comes with the action that caused it, so a model can learn what happens next after each input.

General Intuition states that it builds "World models that predict the outcomes of actions" and "Action models that decide what actions to take", and that players on Medal, the platform it was spun out of, upload billions of gameplay clips every year (checked 2026-11-16).

Microsoft states that its Muse model was trained on more than 7 years of continuous human gameplay, over 1 billion images and controller actions from Ninja Theory's Bleeding Edge, recorded if the player agreed to the EULA (blog post of 2025-02-19).

Environment builders use the game as the task, rules and scoring that an agent trains against, and Vmax states that its GlyphBench, released 2026-09-28, holds 360+ game tasks (checked 2026-11-09). The page on RL environments and who builds them covers that side.

Who Acquires Gameplay Data, and Who Only Helps License It

Gameplay reaches AI developers through four routes: data vendors that license it from studios and supply labs, environment builders that license released games, a company-data program that takes a studio's records with its gameplay, and labs that train on gameplay from their own games or platforms.

Good Start Labs helps publishers license their gameplay data to labs and does not buy it. General Intuition and Microsoft train on data from their own platform or game, and no purchase program of either is on record as of 2026-11-16.

The table lists each company's role for gameplay as of 2026-11-16, grouped by role and alphabetical within a role.

Company Role for gameplay What it acquires or handles Status
Claru (Reka AI Inc. d/b/a Claru AI) data vendor licensing for labs synthetic and game-engine data from partners with engineering teams live profile
Origin Lab (Origin Lab Holdings, Inc.) data vendor licensing from studios to labs gameplay capture, 3D assets, code and commit history, shelved projects profile not yet published
Sieve, Inc. data vendor licensing for labs gameplay, virtual environments, 3D assets profile not yet published
Patronus AI, Inc. RL-environment builder released games live profile
Mercor (Mercor.io Corporation) company-data program, Game Studio listing company records, game and engine code, gameplay recordings and input logs, analytics live profile
Good Start Labs (Good Start, Inc.) RL-environment builder; helps publishers license, does not buy publishers' gameplay data, licensed to labs through consented pipelines live profile
General Intuition lab training on gameplay from Medal, the platform it was spun out of Medal players' gameplay clips no profile
Microsoft Corporation lab training on its own studio's game Bleeding Edge gameplay recorded under the EULA profile not yet published

Gameplay Licensing Terms by Program: Pay, Scope and Deletion

Gameplay programs differ on three terms that matter most to a studio: what unit they pay for, whether the license is limited to training, and what happens to copies when a deal falls through.

The table compares six programs as of 2026-11-16, each value as the program states it. Programs for other data types sit in the program terms compared, dated across the site.

Program Pay unit and timing (stated by the program) License scope and exclusivity Deletion or retention Source
Claru Partnerships archive licenses paid on signed license, net-15 non-exclusive training rights; partner keeps ownership; no republishing term and deletion not stated; delivered content cannot be recalled from client copies claru.ai/partnerships; claru.ai/privacy
Good Start Labs, services for game studios not published not published; "player consent designed in from the start" not published goodstartlabs.com/for-game-studios
Mercor Game Studio listing USD 100,000 to 1,000,000 per company; "Larger studios with full company data have earned more" records, code and gameplay the studio owns not stated in the listing Mercor listing feed
Origin Lab $1M Indie Game Program USD 100 to 100,000 per game; advance against future earnings, "as quickly as 24 hours after signing"; revenue share, percentage not published; per player-hour pricing for core data AI training only, no reselling or republishing; no ownership transfer; exclusivity named as a pricing factor working copies deleted, confirmed in writing, if either side passes originlab.ai/indie; originlab.ai/ip-holders; originlab.ai/faq
Patronus AI no published price; written offer after a sample; paid on acceptance after review of up to 30 days non-exclusive; does not resell records material deleted with written confirmation if no deal data.patronus.ai
Sieve partners "Checks them, then pays"; no amount, rate or timing "You keep ownership"; scope in Sieve's agreement not published sieve.ai/partners

How Gameplay Data Is Priced: Per Game, Per Player-Hour and Per Company

No public rate card exists for gameplay: the figures on record are one vendor's per-game range and one program's per-company range, both stated by the acquirer.

Three pay units appear on acquirers' own pages:

  • Per game: Origin Lab states USD 100 to 100,000, paid as an advance against future earnings, with live games verified on Steam starting at the top of the range (checked 2026-11-09, re-read 2026-11-16).
  • Per player-hour: Origin Lab's FAQ states "Core is priced per player-hour" and publishes no rate (re-read 2026-11-16).
  • Per company: Mercor states USD 100,000 to 1,000,000 for a studio with at least one shipped title that earns revenue (re-read 2026-11-16).

The two ranges use different units and come from different acquirers, so they are neither averaged nor converted to an hourly rate.

No gameplay licensing deal is on record in the deal tracker as of 2026-11-16, and no primary source gives a per-hour price. Price points for every other data type are collected on the page of AI training data prices per work, hour, image.

Player Data in Gameplay Recordings: Privacy Before Licensing

Recordings of players carry personal data in three places, which are voices and faces, children's accounts and reused telemetry, and each needs a basis before a license. This section is information, not legal advice.

Voices, Faces and Accounts in Recordings

A gameplay recording becomes personal data when it carries a player's voice chat, webcam image, username or account ID, and acquirers handle that part separately.

Origin Lab offers player webcam and mic capture "with written player consent" (originlab.ai/faq, checked 2026-11-16), and Good Start Labs states "player consent designed in from the start" (checked 2026-11-09). Claru states it auto-scrubs faces, cards and plates.

A studio removes or masks the identifying parts before delivery, a step covered under de-identification before licensing.

Children Among the Players

Where a game collects personal information from children under 13, the US COPPA Rule applies to a license, and the FTC treats disclosure to train AI as needing its own parental consent.

The Rule is 16 CFR Part 312 under 15 U.S.C. 6501 to 6506, as amended at 90 FR 16918, effective 2025-06-23 with compliance from 2026-04-22 for most provisions.

The FTC states at 90 FR 16950 that disclosing a child's personal information to third parties to train or develop AI is not integral to the service and needs separate consent under 312.5(a)(2).

The page on COPPA and children's data in AI training sets out the Rule.

Reusing Telemetry Under the GDPR and the CCPA

Telemetry collected to run or balance a game was collected for that purpose, so licensing it for training is a new use that EU and California law test separately.

Under Art. 6(4) of Regulation (EU) 2016/679 (applies from 2018-05-25), further processing for another purpose needs a compatibility assessment unless it rests on consent or law.

The factors are the link between purposes, the context of collection, the nature of the data, the consequences and safeguards such as pseudonymisation. The test is set out under purpose compatibility when reusing data for AI training.

Under the CCPA (Cal. Civ. Code 1798.100 and following; in force 2020-01-01, amended 2023-01-01), licensing personal information for money is a sale, and 1798.120 gives consumers the right to opt out of it.

Regulators also watch platform defaults: the Dutch DPA recorded on 2026-08-20 that Twitch's "Training for Generative AI" setting is on by default.

EULAs and Ownership: What a Studio Can License

A studio can license only gameplay and assets it owns or controls: recordings of its own titles made under a player agreement that allows it, its own code and art, not footage of other studios' games or assets bought from a store.

Mercor lists "Footage of other studios' games" and "Asset-store or template code" as not a fit, and requires company records the studio owns (re-read 2026-11-16). Microsoft states that Bleeding Edge gameplay was recorded if the player agreed to the EULA, so the player agreement is the instrument that decides whether recordings can be used.

A studio reads three documents before it licenses: the player EULA and privacy policy, the engine license and any publisher contract. Each can restrict recordings, assets or training use, and this page states no clause of any engine license. Ownership is checked in a rights audit before licensing that covers the code, art and recordings.

Where gameplay sits on a platform the studio does not run, that platform's terms apply: Discord's terms s.9 prohibit "Selling, licensing, or otherwise commercializing content or data obtained from the services" (checked 2026-11-09), as covered under when services train on their users.

Under 17 U.S.C. 201(b) the employer owns a work made for hire unless a signed written instrument says otherwise. Programs license rather than buy: Origin Lab states no ownership transfer and "You retain ownership of the underlying IP" (checked 2026-11-09).

Licensing Gameplay Data: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

Licensing gameplay can earn from recordings and builds a studio already holds, but a model trained on them cannot be untrained and player-privacy duties stay with the studio.

Benefits:

  • An advance can arrive early: Origin Lab states "as quickly as 24 hours after signing".
  • Non-exclusive terms leave the studio free to license elsewhere, as Patronus AI states.
  • A training-only scope keeps resale out, as Origin Lab states.

Risks:

  • Children's data needs separate parental consent under COPPA before it goes to a training license.
  • Reused telemetry faces the GDPR Art. 6(4) test and the CCPA opt-out of sale.
  • Webcam and mic capture needs written player consent, as Origin Lab states.

Downsides:

  • An advance is set against later revenue share, so the first payment is not additional income.
  • No deal benchmark exists, so a price has no public comparison.

Alternatives:

  • License a sample under a non-exclusive license first.
  • License assets and builds but not player recordings.
  • Use the data in-house to train the studio's own agents.
  • Not license at all, as the decision page on whether to sell or license your data sets out.

The rights audit, player agreements and terms offered decide which fits.

Neighboring Game Data Types and the Seller Steps for Gameplay

This page covers gameplay, game worlds and builds as training data; footage without inputs, a studio's code and its company records each have their own page.

Neighboring pages:

Seller steps for gameplay:

  1. Read how to sell or license company data, the complete seller guide.
  2. Clear the recordings, builds and player data, as above.
  3. Agree the license under AI data licensing agreements, clause by clause.
  4. Check when AI data buyers pay: advances, milestones and revenue share.

Player data differs by country, as the AI training data law by jurisdiction shows.

Frequently Asked Questions

Can I sell gameplay footage of games I play?

Not as a license: acquirers want footage of games the seller owns. Mercor lists "Footage of other studios' games" as not a fit (re-read 2026-11-16). Paid capture work is contributor work, covered under AI training jobs and platforms.

How much do AI companies pay for game data?

No public rate card exists, and no deal is on record. Origin Lab states USD 100 to 100,000 per game (checked 2026-11-09) and Mercor USD 100,000 to 1,000,000 per studio (re-read 2026-11-16). Valuation is covered in what company data is worth to AI companies.

It depends on the data and the law. Origin Lab asks written player consent for webcam and mic capture, COPPA needs separate parental consent for under-13 data and GDPR tests reuse. See customer consent before licensing.

Do I keep ownership of a game I license for AI training?

Under the programs checked, yes. Origin Lab states no ownership transfer, Patronus AI and Claru state non-exclusive terms and Sieve states "You keep ownership" (checked 2026-11-09). The trade-offs are set out in exclusive vs non-exclusive data licenses.

Advisers and agents can use the introducer form.