Skip to content
Sell My Data To AI

Buy game and simulation data for AI training

Last updated How we research

In short

Game and simulation data is gameplay video paired with player inputs, plus replays, level files and behavior logs, used to train world models and game-playing agents. Few prices are public. Microsoft says its Muse model learned from more than 7 years of continuous Bleeding Edge gameplay, and The Next Web, citing The Information, reported a $500 million OpenAI offer for a gaming clip platform.

Key takeaways

  • AI teams use game and simulation data for (1) world models that generate playable scenes, (2) game-playing and computer-use agents and (3) player behavior, matchmaking and churn models.
  • Usually priced per hour of gameplay or per title archive. Not published.
  • Common delivery formats: Video with frame-aligned controller or keyboard inputs, engine replay files and match logs in JSON and level, map and asset files from the engine.
  • Check before you license game and simulation data: (1) US copyright: studio ownership of game assets, with assignments from contractors, (2) player EULA and privacy policy cover recording and AI training and (3) US: COPPA for players under 13, including separate consent for AI training disclosures.
  • Every seller we shortlist passes 8 vetting checks before you see a sample.

Game and simulation data: buyer's spec

Used for
  • World models that generate playable scenes
  • Game-playing and computer-use agents
  • Player behavior, matchmaking and churn models
  • Simulation for robotics and driving
Delivery formats
  • Video with frame-aligned controller or keyboard inputs
  • Engine replay files and match logs in JSON
  • Level, map and asset files from the engine
  • Telemetry event tables in Parquet or CSV
Usually priced
Per hour of gameplay or per title archive. Not published
Check before you sign
  • US copyright: studio ownership of game assets, with assignments from contractors
  • Player EULA and privacy policy cover recording and AI training
  • US: COPPA for players under 13, including separate consent for AI training disclosures
  • US: CCPA for player IDs, device IDs and chat logs
  • US: BIPA if voice chat is used to create voiceprints
  • Engine, middleware and licensed IP terms
  • Inputs synchronized with frames
  • UK and EU: UK GDPR, EU GDPR and the ICO Children's code

Why AI teams work with us

  1. 1

    Every seller of game and simulation data we shortlist has passed eight vetting checks, from authority to sign through to a re-identification test.

  2. 2

    Each shortlist comes with a vetting file per seller: rights evidence, provenance record, de-identification method, data manifest and proposed license terms.

  3. 3

    Samples are drawn at random across the archive and labeled with the sampling method, so you can judge the whole package rather than its best rows.

  4. 4

    You contract directly with the seller, and we help agree acceptance tests, staged payments and remedies before any data is delivered.

Book a consultation

What is game and simulation data?

Game and simulation data is the record a game produces as people play it: screen video, the inputs that drove it, match replays, telemetry events, chat and the level and asset files that define the world. Simulation data is the same record from engines built for training, such as driving or robot simulators.

Its value for AI is that every frame comes with a known cause, the input the player gave. That pairing is hard to find in ordinary video.

What do AI teams use game data for?

AI teams use game data mainly for world models and agents. A world model trained on gameplay learns how a scene changes after each input, and an agent learns which inputs reach a goal.

Published examples of game data used for AI training
ProjectData usedSource
Microsoft MuseMore than 1 billion images and controller actions, over 7 years of continuous Bleeding Edge gameplayMicrosoft Research, February 2025
OpenAI Video PreTraining70,000 hours of online Minecraft video, labeled by a model trained on contractor dataOpenAI, June 2022
Reported OpenAI offerA gaming clip platform producing about 2 billion clips a year, for $500 millionThe Next Web citing The Information, June 2026

Microsoft says Bleeding Edge matches were recorded only if the player agreed to the end user license agreement, and that its researchers worked with Ninja Theory and Microsoft compliance teams on collection. That is the standard to ask any seller to match. See the Microsoft and OpenAI profiles.

How much does game data cost?

Game data has no published rate card from any buyer or marketplace we checked. The $500 million figure was a reported offer for a whole company, not a price per hour of gameplay, and it was an offer rather than a completed sale.

Expect a per-hour or per-title license, priced on how much gameplay has synchronized inputs, how many titles and genres it spans and whether level and 3D state data come with it. Our guide to AI training data pricing models explains these structures.

What rights does a game studio need to license player data?

A game studio needs rights in two layers to license player data: the game itself and the players' activity. Code and art made by employees within their jobs are works made for hire owned by the studio, according to the US Copyright Office's Circular 30, but art, music and voice work by contractors need a written assignment, and licensed brands or engines can carry their own limits.

Player data depends on the end user license agreement and privacy policy in force when it was recorded. The FTC has warned that quietly changing terms of service to allow AI training could be unfair or deceptive, so check that the terms covered training and not just game improvement.

How does COPPA affect player data?

COPPA affects player data whenever a game collects personal information from children under 13. In December 2022 Epic Games agreed to pay a $275 million penalty after the FTC alleged it collected Fortnite players' data without notifying parents or getting verifiable parental consent.

The FTC's amended COPPA Rule took effect on June 23, 2025, with compliance due by April 22, 2026. In it the FTC states that disclosing a child's personal information to train or otherwise develop artificial intelligence technologies is not integral to a service and would require separate consent. Ask whether any under-13 accounts are in scope and how that consent was obtained.

Chat and voice add more. The CCPA covers unique identifiers such as player and device IDs, and Illinois' BIPA lists voiceprints as biometric identifiers.

What should you check in a game data sample?

A game data sample should show that inputs, frames and metadata line up and that players cannot be identified. Check these points on a random draw across titles and dates.

  • Inputs aligned with frames at the stated rate, with no dropped segments
  • Chat, usernames and voice removed or de-identified
  • Bot, cheat and test-account sessions flagged
  • Titles, modes and skill levels spread as the manifest says
  • Engine version and replay format documented so files can be re-rendered

Our guide on how to evaluate a dataset sample covers the tests, and the agent trajectories guide explains how game environments relate to reinforcement learning.

If you hold UK or EU data

If player data comes from the UK or EU, UK GDPR or EU GDPR applies to identifiable players. The ICO's Age appropriate design code, known as the Children's code, applies to online services likely to be accessed by children and names online games as an example.

How do you source vetted game data?

You source vetted game data by naming the genres, input types and hours you need, then checking the player terms behind each archive. Every seller we shortlist passes our eight-point vetting standard. Book a free consultation to start.

Frequently asked questions

Can a game studio sell gameplay data for AI training?
A game studio can license gameplay data for AI training when it owns the game assets and its player terms allowed recording and that use. Microsoft says the Bleeding Edge matches used for its Muse model were recorded only if players accepted the end user license agreement. Contractor art, licensed IP and data from players under 13 need separate checks.
Why do world models need controller inputs?
World models need controller inputs because they learn how a scene responds to an action. Video alone shows what happened but not what the player pressed. Microsoft's Muse was trained on images paired with controller actions, while OpenAI's Video PreTraining project had to train a separate model to infer the missing actions in 70,000 hours of online Minecraft video.
Does COPPA apply to game data sold for AI training?
COPPA applies when the game collected personal information from children under 13. The FTC's amended rule, with compliance due by April 22, 2026, treats disclosure of a child's data to train artificial intelligence as not integral to the service, so it needs separate verifiable parental consent. Ask sellers to exclude under-13 accounts unless that consent exists.
Is scraped gameplay video from streaming sites safe to train on?
Scraped gameplay video carries the rights of the game publisher, the streamer and any music or voices in the stream, and none of them has necessarily agreed to training. A licensed archive with player consent and input logs gives a buyer better data and a clearer chain of title. Take independent legal advice before training on scraped footage.

Sources

  1. Introducing Muse: Our first generative AI model designed for gameplay ideation, Microsoft Research, accessed November 2026
  2. Learning to play Minecraft with Video PreTraining, OpenAI, accessed November 2026
  3. General Intuition is raising $300 million to train AI agents on the video game data OpenAI tried to buy, The Next Web, accessed November 2026
  4. Fortnite Video Game Maker Epic Games to Pay More Than Half a Billion Dollars over FTC Allegations of Privacy Violations and Unwanted Charges, Federal Trade Commission, accessed November 2026
  5. Children's Online Privacy Protection Rule (final rule amendments, 90 FR 16918), Federal Register, accessed November 2026
  6. AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive, Federal Trade Commission, accessed November 2026
  7. Works Made for Hire (Circular 30), US Copyright Office, accessed November 2026
  8. California Civil Code section 1798.140 (CCPA definitions), California Legislative Information, accessed November 2026
  9. Biometric Information Privacy Act, 740 ILCS 14, Illinois General Assembly, accessed November 2026
  10. Children's code guidance and resources, Information Commissioner's Office, accessed November 2026

For AI teams

Ready to source?

Book a consultation to brief us. We shortlist vetted sellers, each with a vetting file covering rights evidence, provenance and a representative sample.

Related guides