Epic will allow AI training on Unreal Engine content

Epic Games is preparing an amendment to its user agreement that will lift the ban on training neural networks on output produced in Unreal Engine and MetaHuman. At present the engine's EULA does not permit such use of these tools' output. After the change the restriction will be removed, and developers themselves will decide whether AI may be trained on a particular game — through the terms under which they release the project.

The plans were disclosed by Epic Games head Tim Sweeney — in passing, in a long post on X. "Does Epic restrict AI training on Unreal Engine and MetaHuman output? Right now yes, but we're working to remove that restriction from the EULA terms, with each game built on these tools being released on terms of the developer's choosing, which may allow or prohibit AI training," he wrote. Timelines and the exact wording of the new version of the agreement have not been disclosed publicly, and the company has so far given no comment on the details of the amendment.

The prohibitive line disappears, responsibility moves to the studios

Until now the ban in the EULA worked as protection by default: the output of the engine and MetaHuman formally could not be used to train models. The restriction applied at once to everyone using Epic's tools and required no action from a studio.

After the amendment there will be no blanket ban left, and the only barrier will be whatever a studio writes into the license of its own game. For large publishers with legal departments this is routine work. For small teams, indie developers and freelancers who publish projects on boilerplate terms, it is a task they previously did not have to think about. Artists and animators who hand assets over to a client will probably have to find out separately what a particular project's license says.

The change also affects MetaHuman — Epic's program for 3D animation of digital humans. Its output falls under the same rule: once the ban is lifted, it becomes available for training models within the limits left by the terms of specific projects.

The news surfaced in a dispute about double standards

Sweeney brought up the EULA not at a presentation and not on the company blog. The occasion was a Kotaku piece published on October 26: the outlet pointed to a possible contradiction in the stance of companies like Microsoft and Epic, which are betting on AI development yet at the same time shut down amateur recreations of their games built with AI.

In reply, Sweeney broke the accusation down into parts and tried to pull the story out of the frame of hypocrisy: in his words, Epic does not allow Fortnite intellectual property to be used without permission and compensation, and the company's actions against third-party ports of the game are tied to copyright and licenses. The message about the future EULA amendment turned up inside that same reply.

The context of the statement is notable: Epic is moving to a development model it itself calls "AI-first." It was also reported that the company's project with Disney is being shifted to that principle. Against this backdrop, the disappearance of the ban from the engine's license looks like bringing the terms into line with the new strategy.

Criticism aimed at Valve

Sweeney spent a considerable part of the past week on X answering questions about AI in development: by Game Developer's count, 11 posts in a day and more than 50 in a week. In one of them he criticized Valve for requiring that the use of AI be disclosed on games' Steam pages, calling it unacceptable to force developers to write a political statement that will later be turned against them. Valve itself does not characterize its disclosure requirements as a political statement.

The difference in the two platforms' approaches is now formalized in public: Steam requires disclosure of AI use, Epic is removing restrictions on training models on the output of its own tools.

Where this leads the tools market

If the trend continues, the license agreements of engines and DCC packages will become one of the main fields on which the fate of training data for generative models is decided. The ban in the EULA was for a long time the simplest form of protection for creators: it applied automatically and required neither legal expertise nor negotiations from an artist. Lifting it moves protection to the level of individual projects — which means boilerplate wording, license templates and disputes over what counts as a tool's output will most likely appear.

Epic's decision is probably connected to the fact that a blanket ban sits poorly with the company's own AI development plans: training models on material created in its own tools is obstructed by a line in its own agreement. If other tool vendors follow Epic, the choice of "allow or not" will become part of the production contract — alongside deadlines, cost and rights to assets. In such a scheme the least protected are those with the fewest resources for legal work: solo creators and small teams, whose release terms are usually written from a template.

Community reaction

On X, Tim Sweeney's statement about lifting the ban on training models on Unreal Engine and MetaHuman output was met mostly with anger and distrust. Under his post and in quote tweets, artists, developers and players see a double standard: Epic protects Fortnite intellectual property, yet considers other people's labor acceptable training material. The loudest argument comes down to the claim that, in their view, it is fine to steal from everyone except Epic itself. Alongside this runs a calmer dispute about the logic. A creator of projects built on Epic's tools and a former Pixar employee asks directly why developer consent should be made mandatory if training is in the same breath called legally permissible — as long as the result is not too similar to the original. A few object that this is not about seizing other people's games, but about lifting the ban on training models on your own work made in Unreal Engine, while the terms of a specific game are still set by its author. Such voices are few: support and enthusiasm are barely audible, and accusations of hypocrisy set the tone.

"Stealing from creators is fine as long as nobody steals from Epic? Do you really not see the problem? Artificial intelligence has blinded you."

— @Ruan3D_ (X)

"In short: I benefit from generative models, and you don't."

— @janetacarr (X)

"You're changing the license so that creators on Unreal Engine themselves permit training on their games, meaning consent matters. But in the same breath you write that per the courts you can train on anything if the result is different. Then why make it optional, if training is permitted anyway?"

— @tomjank (X)

What lies behind such a concession from Epic: pressure from the AI industry, requests from large studios, the impossibility of tracking violations?