A language model reads a text and sees where a line of reasoning ends up. Almost everything that came before escapes it, because text, images, code and video preserve the outcome of an action and very little of the process behind it. For systems expected to act, whether agents working on a computer or robots moving through physical space, that process matters a great deal.
Video games keep track of it. Every play session brings together the state of the environment, the available information, inputs, decisions, errors and adaptations in a single digital space, where they can be observed and synchronized.
For years the relationship between games and AI ran in one direction. Chess, Go, StarCraft and Dota gave machines environments in which to learn to compete and to decide. A second direction is now opening, with the player at the center and the machine watching. Some platforms already offer AI developers large volumes of human decisions produced by playing (1), and others promise to translate strategies and adaptations into data a machine can read (2). In March 2026 Microsoft was granted a patent for systems that measure player engagement, including through audio and video when the player consents, and adapt the game experience in response (3).
We are at a point where gaming plays an increasingly clear role in producing data on human behavior for artificial intelligence. This data is deeper than what the web preserves, because it records the process that leads to a decision as well as its result. The signals are visible, although the evidence is partial. Most of what is known comes from the companies trying to build this market, while little is known about the buyers beyond a few names, and almost nothing about the terms of the deals. Still, the questions are already strategic. Who owns the behavior produced by playing? How far does the consent given to play extend? Who controls the points where that behavior is captured? What value do large archives of actions and behavioral profiles take on?
Recording a Decision Process
For an AI that has to act, the most valuable material is the process that leads to a decision, and the web rarely records it in structured form.
More compute and more data. For years AI's growth followed that formula, and a large share of the data came from the web. But the web has a limit too. Epoch AI estimates the effective stock of public human text usable for training at about 300 trillion tokens and projects that, if current trends hold, it will be fully exploited between 2026 and 2032 (4). Volume matters, but the quality and kind of data count for even more.
Robotics shows how much it costs to obtain the right material. Open X-Embodiment, one of the main open datasets of real robot experience, had to pool 60 datasets contributed by 34 labs to pass one million trajectories across 22 types of robot (5). By the standards of the field this is a large scale, and it took physical, costly collection spread over more than thirty labs.
Faced with that cost, labs try several routes, among them teleoperation, simulation, synthetic data and the recording of human activity in digital environments. This is where the video game comes in, which has already left its mark on the history of AI on four fronts. There is computing, because GPUs built for game graphics, widespread and cheap, were repurposed to train models thanks to CUDA, the software with which Nvidia opened them to general-purpose computing (6). There are environments, controlled worlds in which the machine learns from its own experience (7). There is conversation, the language players produce in and around games. And there is the player's behavior, the front that concerns people directly and the one this paper is about.
The Player Enters the Dataset
In 2022 OpenAI had thousands of hours of Minecraft footage available. The videos showed what players did, while the commands that produced it escaped the recording. With Video PreTraining OpenAI hired players to record about 2,000 hours of play with synchronized inputs, and trained a model on that material to infer the commands from the images. That model then labeled about 70,000 hours of public video (8).
The method shows that the process hidden behind a video can be reconstructed, at least in part, and that observation and action can be linked again. It also has an economic side. A small amount of precisely recorded behavior makes an enormous public archive usable, and for the same reason the value of labeled data has a ceiling.
Google DeepMind chose another route with SIMA, working with game developers on seven commercial games. In some sessions one person played while another told the player what to do. In others people played freely and the instructions were matched to their actions afterward (9). Human behavior is the bridge between a goal put into words and the action that carries it out, and access to the games runs through agreements with their developers.
In 2025 Microsoft Research and Ninja Theory took this logic to a different scale with WHAM, the World and Human Action Model. The model was trained on about 500,000 anonymized sessions of Bleeding Edge, more than seven years of play, with every frame aligned to the controller inputs (10). Players who opened Bleeding Edge did so to play, and their choices during a match became the material from which the model learns how an environment evolves and how human actions produce consequences.
What changes between VPT and WHAM is where the data comes from. In the first case players were recruited and paid to play. In the second, the behavior was that of people playing for themselves, collected by the company that owns the game. I consider this a preliminary result, though a sufficient one to settle a single point. What people do while playing can be collected and turned into training material. To talk about a market, though, takes more evidence.
Who collects this material, and what actually reaches the market?
Who Collects, What Is Sold
So the technical possibility is becoming a commercial offer. In 2026 several operators are offering AI developers the same raw material in different forms, the human behavior produced by playing.
About 150 billion human decisions. That is the figure Arkadium attributes to GameLab, launched in 2026 on the strength of more than one hundred titles and 1.7 billion plays a year. The company says it has agreements with some frontier labs, and its site lists Google DeepMind and Meta AI among the partners, without saying what the collaboration is or on what terms (1).
Where GameLab counts decisions, Framestime measures their density. It comes out of BUFF, an app that rewards people for playing, and says it synchronizes video and inputs with the player's gaze, the three-dimensional state of the environment, strategic intent and communication between teammates. Commercial access to the datasets goes through a license agreement separate from the game's terms of service (2).
Pleiada reverses the order. First comes the behavior the client needs, then a virtual world built to obtain it and populated by real people (2). The game becomes an instrument designed to produce what someone wants to observe, and to me this is the most instructive case in the paper.
Skillprint moves up a level and sells interpretation. It says it ties decisions, adaptation, cognition and mood to the actions, and promises studios a share of licensing revenue (2). That exclusivity is short-lived, because other operators, Framestime among them, claim to synchronize the player's strategic intent with the actions. Cognition and mood are in any case estimates produced by a model, and their validity outside the commercial context in which they are claimed remains to be demonstrated.
Between the player and the AI lab, then, runs a chain of four steps (capture, structuring, interpretation and training) that starts with gameplay. Each operator occupies a different stretch of it. GameLab works mainly on capture at scale, Framestime on structuring signals, Pleiada on designing the session upstream, Skillprint on interpretation. Gameplay has always produced data. But now part of it may be worth something outside the video game industry.
At this stage supply is more visible than demand. In a rising market, suppliers put their numbers in the shop window, prized goods on display. The buyers stay in the half-light. A few names appear among the partners, while the subject and terms of the deals remain confidential. Capital meanwhile seems to reward the companies that turn data into models. General Intuition, for example, comes out of the Medal ecosystem, a platform where players upload and share gameplay clips, and uses gameplay labeled with actions to build its own models. It says it has raised more than $650 million across 2025 and 2026 (11).
One consideration follows from these steps. If the raw material tends to become common, power shifts toward whoever controls the points where behavior is captured.
Where Power Sits
As long as gameplay data serves to balance a game or measure its retention, it lives largely inside the product that generates it. That telemetry mostly records events and purchases. When the data becomes training material for AI systems, its purpose changes, and with it the question of who may use it.
A model that has to learn to act needs continuous sequences in which every frame is aligned with the command that produced it, and a record like that exists only if someone captures it while the match is under way. The first advantage therefore goes to the company that controls the capture point. That can be the engine the game runs on, the software that overlays the experience, the anti-cheat system, which watches the state of the machine during play, or the platform that distributes the title. A company sitting at one of these points can decide which sessions become data and on what terms. Two of the cases seen earlier arise in this very layer. General Intuition comes from Medal and Framestime from BUFF.
Four links, then. Capture, structuring, interpretation, training. Whoever holds one of them decides how behavior becomes data.
In the China paper I used a similar image, a loop with four links (identity, payment, presence, recognition), to read a market. I counted how many links sat under a single owner, and that number indicated the depth of monetization the market could sustain (12). Here the same count applies to the data supply chain, and in my view an actor's advantage depends on how many links it holds. The diagram below shows three fictional cases.
Take a session that has just ended. Who decides what becomes of it? To become data it passes through the links one after another, and at each step it can change hands. The party that receives it sets the conditions for the next step.
With Bleeding Edge, as far as can be established, three links out of four had the same owner. The studio that created the game, Ninja Theory, joined Microsoft Studios in 2018 (13), and Microsoft Research then used the sessions to train its own model (10). A company holding a single link is in a different position. It has to hand the session to someone or receive it from someone, and at each hand-off the other party can set a price or a condition.
The same mechanism applies to capital and to states. When a consortium including the Saudi sovereign wealth fund completed the acquisition of EA (14), a link changed owner, and with it the behavior of millions of players came within the perimeter of new shareholders. In China the entry link, the identity of the player, is set by the regulator. Since 2021 no gaming service can be offered to anyone who logs in without real-name registration (12), and the large publishers operate inside this framework with a degree of autonomy from political power that is hard to gauge from the outside. In both cases, the holder of a link decides on what terms the behavior of players becomes visible to others.
On the question of rights, the first critical point is consent. Under the GDPR it must be freely given, specific, informed and unambiguous (15), and permission to record a session is hard to stretch to the training of someone else's models. In the United States the FTC warned in 2024 that quietly changing terms of service for this purpose may amount to a deceptive practice (16). The market already has a partial answer. Framestime's separate license and Pleiada's pre-cleared rights make the legal provenance of behavior a commercial attribute of the data.
Equally delicate is what can be inferred from these records. Under the GDPR, behavioral characteristics become biometric data when they are technically processed to identify a person (15), and research on mouse dynamics shows that the way someone interacts with an interface can tell one user from another (17). The Microsoft patent also provides for audio and video to assess the player's reactions, with the player's consent (3). An inferred state is thus added to the observed action, and the data becomes more sensitive.
On the demand side, by contrast, the AI Act can help. It requires providers of general-purpose models to publish a summary of the content used for training (18), and it could bring part of the sources to light, although not every dataset provider would be named.
The final question concerns the key piece of the whole puzzle, the player. Who can collect and transfer this material, and what rights does the player keep when their behavior leaves the game to become raw material for another system?
What Comes Next?
The answers depend on how far the phenomenon consolidates, and it is still early. Human gameplay could remain a niche source, useful mainly for evaluating or supplementing agents and models, while simulation, self-play, synthetic data and teleoperation keep its value in check. Or it could become a stable resource for systems that act in the world, and open a new revenue stream for publishers and platforms that control access to players.
It is too early to speak of a consolidated AI infrastructure, yet late enough to stop calling the phenomenon a gaming curiosity. To tell which of the two paths is opening, we can only watch, and the next signal will come from the lab. The year-long partnership between Mila and Worldmodeldata, which began in September 2026, will test whether virtual experience carries over into real-world performance (19). Meanwhile the first buyers could emerge, through the transparency obligations of the AI Act (18) or through litigation. Contracts will offer another indication, since explicit clauses on model training in the terms of service of the large publishers would show that monetization is already under way. Capital, finally, will show where the market places value. The acquisitions to watch are those of studios, engines, platforms or capture software motivated in part by access to behavioral data, and above all those in which an actor that already holds a link in the chain adds another.
Whatever the scenario, the player's position has changed. They enter a match to have fun or to compete, and the sequence of their actions can feed other systems. I believe that understanding who can collect that sequence, and on what terms, now matters to strategic analysis as much as to game design.