Licensing the Scale Problem

Once a generative AI use creates a meaningful copyright question, the practical issue is often not whether licensing exists in the abstract, but whether the right kind of license can cover the right works at the scale, price, and speed the system requires. This chapter compares direct and collective voluntary licensing with compulsory licensing, extended collective licensing, and opt-out systems. The aim is to match each tool to the rights landscape instead of treating either “licensing is impossible” or “one deal proves feasibility” as a universal answer.

The Copyright Office’s training report treats licensing as a market response to uses that may not qualify as fair use. It also recognizes that the market is young and uneven. A license can supply consent, define permitted uses, allocate compensation, and create controls around a model. It cannot automatically solve every technical problem, identify every rights holder, or guarantee that a model will never reproduce protected expression.

Voluntary licensing works where rights and value can be organized

A voluntary license is negotiated in the market. It can be direct, between a developer and a rights holder, or collective, through an organization that represents multiple owners and negotiates on their behalf. The agreement can be tailored to the work, the model, the training phase, the deployment, the territory, the duration, the treatment of outputs, and the compensation structure. That flexibility is the main advantage over a one-size-fits-all statutory rule.

Direct licensing is most practical when the rights are concentrated and the content is valuable, identifiable, and available in useful volume. A major stock-photo catalog, a national news archive, a large music catalog, or a film studio’s library may be easier to license than millions of isolated personal posts. The rights holder may already have metadata, an established licensing operation, and a reason to negotiate. The developer may also value the catalog for its quality, recency, curation, or exclusivity, not merely because it reduces litigation risk.

This is why licensing activity can be evidence of a functioning market without proving that the market covers all training needs. A deal for a high-quality catalog demonstrates that the parties can agree on terms for that catalog and use. It does not show that a general-purpose model can lawfully obtain every work needed for every language, domain, style, or capability through comparable deals. The denominator matters: “licensed data” may mean a large collection in one sector, not the universe of material that a state-of-the-art system might use.

A collective management organization, or CMO, can address a different problem. Instead of requiring a developer to locate and negotiate with every owner, a CMO aggregates rights and acts as a clearinghouse. It may reduce the number of transactions, standardize reporting, collect royalties, and distribute payments. Collective licensing is especially attractive when many owners have similar interests but the value of each individual work is too small to justify a separate negotiation.

Aggregation is not costless. A collective must know what rights it represents, maintain reliable ownership and metadata records, decide how to distribute payments, and give owners meaningful control over participation and terms. A sector may be organized enough for collective licensing in one country but not another.