Australia is considering two copyright frameworks that would give AI developers a legal path to train on publicly available material that rights holders haven’t protected through an opt-out. One would route payment through a central body or collecting societies. The other would require a developer to complete a minimum number of licensing agreements before receiving legal protection to train on additional material without paying every remaining rights holder. Neither framework has been adopted, and the government is still consulting on the options.
The quota model would change the economics of a catalog license. A deal with a major studio, label, publisher, image library, or other large rights holder could cover the licensed material and also help the AI company qualify for access to a much larger pool of unprotected work. That gives catalog scale value beyond the works inside the agreement and puts more bargaining power in the hands of organizations capable of clearing large bodies of content at once.
A Quota Would Turn Catalog Licenses Into Market Access
Under the quota proposal, an AI developer would need to complete a minimum number of data-access agreements for a required period. Once it met that threshold, it could receive legal protection to train on other online material whose owners hadn’t opted out. The consultation documents state that no additional payment would be required for those remaining creators and rights holders.
The government hasn’t disclosed the number of agreements required, the amount of material each agreement would need to cover, or whether several large catalog licenses could satisfy the quota. Those terms determine the commercial value of the model.
A license that helps an AI company reach the quota carries two forms of value. It authorizes use of the catalog inside the agreement, and it moves the developer closer to legal access to material owned by people outside the deal. A rights holder able to clear millions of works can therefore sell something an individual creator can’t: progress toward broader legal certainty.
That increases the negotiating value of catalog aggregation. Studios, labels, publishers, collecting societies, stock-media companies, and other organizations that can document rights across large inventories become more useful counterparties because each transaction resolves more of the developer’s clearance problem.
Independent creators remain part of the training supply while contributing far less to the developer’s path through the licensing threshold. Unless they’re represented inside a qualifying catalog, their practical protection would depend heavily on the opt-out system.
Collective Licensing Would Centralize Payment and Administration
The alternative framework would let AI companies pay a central body for access to unprotected material or obtain licenses through collecting societies that represent categories of rights holders. The government has been examining licensing arrangements through its Copyright and Artificial Intelligence Reference Group, while existing Australian collecting societies already aggregate permissions and distribute licensing revenue across music, screen, writing, and other copyright markets.
Collective licensing solves a transaction-cost problem that direct negotiation can’t handle efficiently at model-training scale. An AI developer working with billions of files needs a clearance mechanism that doesn’t require a separate contract for every article, photograph, recording, script, or video.
The commercial question moves from whether payment exists to how that payment gets attributed. A central body still needs rules for registration, repertoire matching, usage evidence, unclaimed money, category valuation, audit rights, and disputes over ownership. Large rights holders can maintain rights databases and staff the claims process. Smaller creators may own valid rights while lacking the administrative infrastructure needed to identify use and collect a share, which gives organizations that can aggregate rights and administer claims at scale another source of leverage even if the payment pool reaches more rights holders than the quota model.
Opt-Out Makes Technical Control Part of Copyright Value
Both proposals would allow rights holders to protect material through an opt-out and would impose penalties on AI companies that evade recognized protections. The leaked consultation materials also contemplate transparency, audit requirements, protections for Indigenous cultural and intellectual property, and best-efforts obligations around pirated material.
Owning copyright doesn’t guarantee control of the technical environment where the work appears. A publisher can manage crawler instructions on its own site, while an independent writer, performer, filmmaker, photographer, or musician may distribute through services operated by somebody else. The creator can own the underlying work without controlling the robots.txt file, metadata field, account setting, API rule, or other technical signal an AI crawler is expected to honor.
The final rules would need a recognized technical standard that survives syndication, reposting, platform migration, and other forms of distribution. They would also need to address material collected before a rights holder applies the signal. Without reliable records of when content was gathered and which protections were present, enforcement becomes difficult to verify.
Technical rights management therefore becomes part of the economic value of representation. A large catalog owner can negotiate licenses, maintain machine-readable rights data, monitor crawler behavior, and pursue violations. Smaller rights holders face the same legal question with higher administrative cost per work.
Film, TV, Music, and Publishing Don’t Collapse Into One Rights Layer
AI training can involve finished films and episodes, scripts, subtitles, recordings, production images, artwork, photographs, reviews, interviews, promotional assets, and other materials that sit under different ownership and contractual structures. A studio may control a finished program while separate rights, participation interests, reuse terms, or contractual restrictions apply to writers, performers, composers, photographers, or other contributors. Music catalogs can divide rights between recordings and compositions, while publishers can control text and license accompanying images from another owner. A broad training permission therefore has to operate across those layers without assuming that one transaction clears every interest attached to every asset.
That gives rights organizations with clean ownership data another advantage. AI companies need counterparties that can state what they control, which territories are covered, what uses are authorized, and which claims remain outside the agreement. The value of the license increasingly includes the quality of the rights information attached to the content.
The music business is already testing that commercial structure. Warner Music’s agreement with Suno turned a copyright fight into a licensed supply relationship for newer Suno models, with artist participation and compensation still dependent on the terms governing who opts in and how revenue gets allocated.
Training Investment Carries a Different Copyright Requirement Than Inference
AI companies have linked Australian copyright certainty to decisions about where they perform model training. OpenAI has said the current rules affect whether it can build training infrastructure in Australia, while earlier government briefing materials described Anthropic’s interest in greater copyright clarity around training data.
Anthropic’s September 16 agreement for its first Australian data center sharpens that issue. The Queensland facility is planned for inference rather than model training, showing that AI infrastructure investment can proceed without resolving every training-data question. Training carries the direct copyright exposure because it requires the developer to assemble the datasets used to build or update a model. Inference runs an already-trained model to answer user requests.
Australia can therefore attract some AI compute while the copyright dispute remains unresolved. The policy decision has greater influence over whether companies perform data-intensive model training inside the country and which licensing costs attach to that activity.
The Missing Rules Decide Whether Payment Tracks Use or Administrative Scale
The government hasn’t chosen a framework or introduced legislation. The economics remain undefined until it specifies the quota size, minimum deal coverage, recognized opt-out signals, treatment of previously collected material, audit requirements, payment formulas, claims procedures, and the role of collecting societies or any new central body. Those details determine whether compensation follows actual use of protected work or flows primarily through the organizations best equipped to negotiate, register catalogs, and administer claims.
A low quota satisfied by a few large agreements would concentrate leverage around major catalogs. A broader collective system could spread payment across more rights holders, but distribution still depends on accurate repertoire data and a claims process creators can actually use. An opt-out can preserve control only when the rights holder can apply it, the developer must recognize it, and an audit can prove whether it was honored.
The Streaming Wars Take
If Australia adopts a licensing quota, major catalog agreements would carry extra value for the AI buyer: each deal could help unlock legal access to work owned by people who weren’t part of the negotiation. That gives large rights holders a stronger reason to price access around the clearance value they provide, not simply the volume of material they license.
A collective model shifts that value toward collecting societies and whatever organization administers the payment pool. Their databases, membership systems, ownership records, distribution formulas, and audit processes become part of the infrastructure AI developers need to train legally at scale.
Independent creators face the opposite economics. Their work can remain useful to a training dataset while their individual license contributes little to solving the developer’s clearance problem. Their leverage therefore depends on representation, enforceable opt-outs, reliable rights metadata, and a payment system that can connect model inputs back to the people who own them.
Australia’s decision will set the price of legal certainty and determine who gets paid for providing it. The central commercial question is whether an AI company pays for the work it actually uses or pays a smaller number of organizations whose scale gives it permission to reach everything else.
The Streaming Wars is intentionally ad-free
We don’t run display ads. Not because we can’t, but because we don’t believe in them.
They interrupt the reading experience. They cheapen the work. And they burn advertisers’ money on impressions nobody actually wants.
So we chose a different model.
We say the things people in this industry are already thinking but don’t say out loud. We connect the dots beyond the headline and focus on explaining why things matter to the people working in this business.
If you believe industry coverage can exist without clutter and interruption, you can support it here → SUPPORT TSW.
Support is optional. But it directly funds research and continued coverage — and helps prove this model can work.
Support TSW →
