Zuckerberg Questioned Creators’ AI Leverage. Copyright Claims Survived

Mark Zuckerberg’s September 2024 argument was about bargaining power, not proof that creative work has no value. The Verge’s account of the interview records his view that individual creators and publishers tend to overestimate the importance of their particular material to a large training process, and that Meta could omit content rather than pay every rights holder.

The dispute has since become more concrete without producing a universal answer on AI training. The federal court’s March 2026 order records that the named authors had lost their training claim after failing to counter Meta’s fair-use defense with meaningful market-harm evidence, while direct distribution allegations had not reached summary judgment and a contributory-infringement claim tied to alleged BitTorrent uploading was allowed to proceed.

What Zuckerberg’s argument did—and did not—establish

Zuckerberg’s reasoning rests on the scale of modern model development. If a system is trained on an enormous collection, removing one author, photographer or publisher may produce no measurable change in its output, leaving that rights holder with little individual negotiating power.

That is an economic and technical claim, not a rule of copyright law. Whether one work is essential to a model is distinct from whether the developer had permission to copy it, whether an exception applies or whether the resulting system affects the market for the original.

His position also leaves room for selective licensing. Content that is scarce, authoritative, current or otherwise difficult to replace may be valuable enough to support a commercial agreement even when most individual works are treated as interchangeable inputs. The practical dividing line is therefore not whether creative work has value, but whether a particular owner controls material the developer considers costly to lose.

Meta broadened its use of public adult content in Europe

Meta later applied this scale-based approach to material on its own services. Meta’s EU training notice, published in April 2025 and updated in March 2026, states that public posts and comments shared by adults on Meta products, together with interactions people have with Meta AI, may be used to train and improve its models.

The notice gives people based in the EU a form for objecting to the use of their public data. It excludes public information from EU accounts belonging to people under 18 and distinguishes private messages with friends or family from messages that a participant chooses to share with an AI feature.

These distinctions narrow what the phrase “your data” can responsibly mean. A public Instagram post, a private conversation, a prompt deliberately submitted to Meta AI and a copyrighted book obtained outside Meta’s services involve different collection routes, product rules and potential legal claims. A regional objection mechanism for public platform data is not a universal opt-out from every AI-training dataset.

Meta’s court victory was narrower than a general fair-use ruling

The authors’ loss on the training claim did not establish that copying protected works for model development is always fair use. It reflected the evidence presented by the named plaintiffs in that lawsuit, particularly the absence of a meaningful showing that Meta’s copying had caused or was likely to cause harm in the market for their works.

That distinction matters because US fair use is assessed in context. The purpose and character of a use, the nature and amount of the copied material, and the effect on the potential market can lead to different outcomes when the works, evidence or conduct change.

The remaining allegations also concern a separate act. Training asks whether copies may lawfully be used as model inputs; distribution asks whether protected files were uploaded to other users during their alleged acquisition through BitTorrent. Contributory infringement adds another question: whether that conduct facilitated infringement by third parties.

Allowing an amended claim to proceed is not a finding that Meta committed infringement. It means the allegation can remain in the case and face later scrutiny, including summary judgment. The result is a divided procedural record rather than a simple victory for either AI developers or creators.

Replaceability does not erase aggregate value

Zuckerberg’s argument measures one work by the difference its removal would make to a completed model. Creators and publishers generally encounter value through other mechanisms: licensing access, controlling reproduction, preserving exclusivity and protecting the market for the original work.

A work can therefore be replaceable inside a dataset while remaining commercially valuable and legally protected. The two propositions are compatible because model developers and rights holders are measuring different things: marginal contribution to system performance on one side, and control over copying and exploitation on the other.

Aggregation creates the deeper conflict. A developer may require a vast supply of human-produced text, images, audio or video while maintaining that no single contributor matters enough to demand payment. Each work can appear dispensable even when the collection built from those works is central to the resulting product.

This imbalance explains the importance of publisher agreements, collective licensing proposals and class litigation. They combine claims that may be weak in isolation and shift the negotiation from the technical importance of one item to the economic value of access to a larger body of protected material.

The unresolved issue is leverage, not whether creators matter

The developments after the 2024 interview have not disproved the narrow prediction that withholding one creator’s material will rarely alter a large model’s capabilities. They have shown why that prediction cannot settle the broader dispute: Meta expanded the categories of platform content used for European training, introduced a regional objection route and continued defending distinct claims over training and alleged file distribution.

For the creator economy, the central question is whether technical replaceability should determine who receives permission, control or compensation. The surviving claims do not answer that question, but they prevent Meta’s success on one evidentiary record from becoming a blanket legal conclusion about all copyrighted training data.

Also read:

• X Is Killing Communities: The Controversial Shutdown That Has Creators Up in Arms
• What Forbes’ 2026 Top Creators Celebration Says About Sustainable Growth
• How to Leverage Big Data to Effectively Market Hotels