No copyright carve-out for AI training: Government draws a line
Australia has ruled out a copyright exemption that would let AI developers train models on creative works without permission. The message from the Attorney-General is blunt: there will be no text and data mining carve-out.
A government reference group on AI and copyright will meet over the next two days to assess whether current laws need a refresh. Any update will stop short of giving tech companies free rein over Australian content.
What sparked this decision
The idea of a training exemption surfaced in the Productivity Commission's interim report on data and digital technology, backed by claims that AI could add $116 billion to the economy over a decade. The creative sector pushed back hard, arguing such a move would strip income and control from artists and rights holders.
After weeks of mixed signals, the government has now closed the door. As the Attorney-General put it, "we will not be entertaining a text and data mining exception," framing the stance around fair pay and fair terms of use for creators.
Reactions across the sector
- Screenrights welcomed the clarity, calling it the "signal" needed to begin serious negotiations on licensing frameworks.
- ARIA said major companies have been using copyrighted works without permission or payment and can, in fact, approach rights holders to license content. The view: claims that licensing is too hard don't stack up.
- Opposition voices also rejected any carve-out, saying it's not acceptable for big tech to use Australian works without paying for them.
What this means for you
- Creators and publishers: Treat dataset use like any other exploitation of your catalog. Document ownership, set clear licensing terms for AI training, and prepare standard quotes for bulk access. Consider collective licensing options to reduce friction.
- AI companies: Assume consent is required. Secure licenses for training corpora, keep audit trails of sources, and build processes to handle takedown or exclusion requests. Budget for content costs in model training and updates.
- Agencies and the public sector: Update procurement and grant conditions to cover AI training rights, dataset provenance, and record-keeping. Include indemnities and escalation paths for copyright complaints.
Legal backdrop and open questions
Australia operates under fair dealing, not broad fair use. Without a specific exception for text and data mining, using copyright content for AI training generally requires permission or a licence. Expect debate to focus on practical licensing, transparency of datasets, and how collective schemes might work in practice.
- Will a standard industry licence emerge for training data? If so, who administers it and how are rates set?
- How will smaller developers access high-quality datasets on fair terms?
- Will transparency rules require developers to disclose sources used for training?
Immediate steps to reduce risk
- Audit past and current datasets for copyrighted materials and log provenance.
- License what you need; replace what you can with public domain or licensed alternatives.
- Contract for future-proofing: include training rights, update clauses, and dispute processes in deals with creators and vendors.
- Document policies for exclusion requests and maintain contact points for rights holders.
Where to learn more
- Productivity Commission - background on the interim report that raised the TDM exemption question.
- Attorney-General's Department: Copyright - overview of Australia's copyright framework and fair dealing.
- Complete AI Training: Courses by Job - practical courses for legal, policy, and creative teams working with AI.
Bottom line
The policy direction is set. No free pass for AI training on Australian works. If you build models, plan for licensing. If you create content, prepare to be asked - and paid - for access.
Your membership also unlocks:
