Did Sora Train on YouTube Videos Without Permission

The public record does not answer whether Sora used YouTube videos. It does show that such use, if it happened without permission, would collide directly with the rules creators accepted when they uploaded their work.

OpenAI has not confirmed whether YouTube videos helped train Sora, its text-to-video generator. YouTube CEO Neal Mohan has not claimed that the company found proof either. His position is narrower and clear: if OpenAI downloaded video clips or transcripts from YouTube to train Sora without authorization, OpenAI would have violated YouTube’s terms of service.

That distinction matters. The available reporting establishes a serious dispute over permission, but it does not establish the contents of Sora’s training dataset. OpenAI has described its sources as publicly available and licensed data. The company has avoided identifying the specific sources behind the model, including whether YouTube supplied any material. That silence leaves the central question open while making the policy question easy to understand.

In an interview with Bloomberg, Mohan framed the issue around the promises made to YouTube creators. People upload work to the platform under specific rules, and those rules govern how other parties may access and use the material.

Mohan said creators expect YouTube’s terms of service to be followed. He specifically pointed to restrictions on downloading transcripts and video portions. His warning was direct: using those materials to train Sora would be a “clear violation” of the platform’s terms.

The important point is that Mohan spoke conditionally. He said, in substance, that such conduct would violate the rules if OpenAI had done it. He also said he had no personal information confirming whether Sora used YouTube data and directed the question back to OpenAI.

That wording avoids an accusation while still drawing a firm boundary. YouTube does not need to prove that Sora used its videos before explaining what the platform permits. The rules already provide the boundary, and unauthorized downloading would cross it.

The public record does not show whether OpenAI trained Sora on YouTube videos. It does show that YouTube CEO Neal Mohan views unlicensed downloading of platform videos or transcripts for Sora training as a clear breach of YouTube’s terms. OpenAI has described Sora’s data only in broad categories—publicly available and licensed—while declining to identify specific video sources. That leaves the factual question unresolved but underscores a larger principle: public access is not the same as permission for bulk AI training, and transparent licensing is the clearest way to protect creators and model developers alike.

The uncertainty began after clips from a Wall Street Journal interview with OpenAI Chief Technology Officer Mira Murati circulated online. Asked what data Sora used and whether YouTube videos were included, Murati responded that she was unsure.

Murati later described Sora’s training data in broad terms. She said OpenAI used publicly available data and licensed data. That description gives a general category, not a source list. It does not identify which video platforms supplied material, whether specific types of YouTube content appeared in the data, or whether OpenAI obtained permission from the relevant rights holders.

OpenAI also declined to confirm whether Sora used Shutterstock’s video library. The company had entered into a multiyear data agreement with Shutterstock that covered access to images, videos, and music for training. That deal demonstrates that licensed video data was available to OpenAI. It also makes the refusal to identify particular sources more noticeable.

A licensing deal can answer one part of the training-data problem: it can give a company permission to use material within an agreed scope. It does not answer whether the company also collected material from sources outside that agreement. The public statements about Sora leave that question unresolved.

YouTube’s terms of service set practical limits on how content can be accessed and reused. Creators grant YouTube rights to host and distribute their work through the service. That arrangement does not automatically give every outside company permission to scrape, download, transcribe, or repurpose the videos for a separate commercial system.

A video on a public website can be viewed by a person without being free for bulk collection by an AI company. Public visibility and permission for machine-learning use are separate questions. The distinction has become central to disputes involving large language models, image generators, and video generators.

For creators, the issue reaches beyond a legal technicality. A video may represent hours of filming, editing, research, narration, animation, or performance. A training system can absorb patterns from thousands or millions of examples, while the individual creator may have no practical way to see whether a particular video was included or to negotiate compensation.

YouTube’s stated position gives creators at least one recognizable rule: access through the platform does not grant unlimited permission for extraction. If a company wants to use videos for model training, a license or another valid authorization provides a clearer path.

“Publicly available” is not a permission slip for a commercial AI company to ingest creators’ work at industrial scale.

If a developer cannot explain where its training video came from and what rights cover it, creators are being asked to bear the risk of the developer’s opacity.

Mohan’s statement concerns YouTube’s contract and platform rules. Copyright law raises a separate set of questions. A violation of YouTube’s terms could support a dispute between YouTube and OpenAI even if a court later reached a different conclusion about copyright infringement.

Copyright holders could also bring their own claims. Those claims would require analysis of ownership, copying, the nature of the use, the amount and importance of the material taken, and the effect on the market for the original work. Courts would also consider any defenses OpenAI raises, including fair use in the United States.

OpenAI’s broader legal strategy has included arguments that training models on copyrighted works can qualify as fair use. That argument remains contested, and ongoing lawsuits will help determine how courts treat large-scale copying for AI training. A company cannot settle every platform or copyright question simply by calling its data publicly available.

The legal outcome may depend on details that remain hidden: which videos were collected, how they were obtained, what rights the relevant creators held, how the material affected Sora, and whether any licenses covered the use. The public statements so far provide almost none of that detail.

Mohan acknowledged that YouTube content is not categorically unavailable for AI training. Individual creators can make licensing agreements, and those agreements can authorize uses that the standard platform terms would otherwise restrict.

This is an important qualification. YouTube hosts content from many different rights holders, and the rights attached to a video can vary. A creator may own all the relevant rights, share them under a contract, or have permission from a studio, broadcaster, music label, or other participant. A negotiated license can define what a model developer may copy, how long the permission lasts, and whether the creator receives payment.

Mohan also indicated that Google may use some portion of YouTube’s content to train its own Gemini models. He said Google follows YouTube’s terms and existing contracts when doing so. That statement does not establish that Google has blanket permission to use every video. It describes a company-specific process for working within the platform’s rules and agreements.

The difference between Google’s internal use and OpenAI’s alleged use is therefore about authorization and compliance, not simply about who owns the platform. If Google has a contractual basis for particular content and OpenAI does not, the two activities can carry very different legal and business consequences.

Shutterstock’s data licensing arrangements offer a useful comparison. The company has described datasets containing visual material such as images, videos, 3D models, and music. Its contributor materials say data deals began generating payments for contributors in late 2022.

The licensing model includes limits. Shutterstock says customers receive a license for machine-learning or computer-vision training within the agreed scope. The licensed material cannot simply be reused for public marketing or other commercial applications outside that purpose. The company also describes security requirements intended to prevent unauthorized distribution.

That kind of agreement creates a paper trail. It can identify the supplier, the categories of content, the permitted purpose, and the payment structure. It gives contributors a mechanism for compensation and gives the AI company a defined set of rights.

OpenAI’s public description of Sora does not provide the same level of detail. The company has disclosed that licensed data was part of its approach, but it has not published a complete list of datasets or explained how YouTube material would fit into that framework. Without those details, outside observers cannot distinguish licensed footage from material gathered through ordinary platform access.

Training data is central to a generative video system. Sora’s ability to produce scenes, camera movements, editing patterns, visual styles, and other video characteristics depends on the examples used to build the model. The source of those examples affects both the system’s quality and the rights of the people whose work contributed to it.

A clear account of the data would also help creators understand their position. They could determine whether a license exists, whether an opt-out is available, whether compensation is possible, or whether a dispute should be directed to YouTube, OpenAI, a rights organization, or a court.

OpenAI has reportedly used an opt-out approach for some Sora-related content concerns. Under that model, rights holders may ask for specific characters or videos to be blocked, but a blanket request covering all uses may be unavailable. That approach places a burden on rights holders to identify material after the system has been built or released.

For video creators, that process can be especially difficult. A model may learn broad visual patterns from many works without producing a recognizable copy of any single one. Detecting the use of an individual video can require technical analysis, access to information that OpenAI has not provided, or a prompt that exposes a distinctive result.

Transparency would reduce that burden. A dataset summary, licensing explanation, and clear removal process would not resolve every copyright dispute, but they would give creators facts instead of speculation.

The available evidence does not prove that Sora trained on YouTube videos without permission. Murati’s uncertainty and Mohan’s lack of confirmation leave that factual question unanswered.

The available evidence does support a firm judgment about the hypothetical conduct. If OpenAI downloaded YouTube transcripts or video clips without authorization, it would have ignored a basic platform rule. Calling the material publicly viewable would not erase the restrictions attached to access. A large technology company building a commercial model should know the difference between seeing a video and receiving permission to ingest it at scale.

That matters because creator platforms depend on predictable rules. Creators accept platform terms in exchange for distribution, audience access, and the services YouTube provides. They may tolerate certain uses, negotiate others, or leave the platform when the arrangement no longer works for them. Quietly treating their uploads as raw material for a competing product would weaken that bargain.

OpenAI has a straightforward way to settle much of the dispute: explain which video datasets trained Sora, identify the licenses that cover them, and describe how the company handled material without a license. Until then, the public cannot determine whether YouTube videos were included or whether any inclusion was authorized.

The current record therefore supports two conclusions. Sora’s use of YouTube videos remains unconfirmed. Unauthorized use, if it occurred, would represent a direct breach of the rules creators rely on and a serious failure to respect the permissions behind the platform’s content.