On 13 July, the Australian Financial Review reported that Anthropic told Treasurer Jim Chalmers its US$15 billion ($21.6 billion) Australian data-centre investment is contingent on the government clarifying local copyright law. The company is not asking for the text-and-data-mining exemption Canberra already rejected; it wants certainty over its liability to rights holders, particularly the long tail of small creators it says it cannot practically license. In the same week, the Prime Minister signalled a pivot: AI must "earn its social licence."
Strip away the dollar figure and this is a supplier conditioning a capital commitment on the state removing a legal risk it does not want to carry. That is a legitimate negotiating position, and Australian boards should read it closely, because the risk Anthropic is seeking certainty on does not evaporate when the model is sold. It moves downstream. If you deploy AI trained on contested material, the copyright exposure, and the social-licence exposure, land partly on your organisation, not just the vendor's. There are 81 AI-copyright suits running in the US as of February, and Anthropic itself paid $1.5 billion to settle a pirated-books class action. This is not abstract.
For a director, the piece is a signal on two fronts at once. First, capital allocation: watch how a sophisticated counterparty ties investment to regulatory certainty, because your own AI business case rests on assumptions about a policy settlement that has not happened. Second, social licence: when a government makes earning it an explicit precondition, AI stops being purely an efficiency decision and becomes a licence-to-operate question, which is squarely a board matter.
- When we license an AI tool, what does the contract say about copyright indemnity? If the vendor is fighting for certainty on its own liability to rights holders, what exactly have they left sitting with us?
- Our AI business case assumes a regulatory settlement that does not yet exist. If Australia lands on statutory licensing with per-use payments, does the economics of what we have deployed still work, or have we underwritten a cost that has not arrived yet?
- The Prime Minister says AI must earn its social licence. For our most visible AI use, could we defend it to a customer, a regulator, or a journalist tomorrow, or are we relying on no one asking?
- Do we actually know what our AI tools were trained on? If a rights holder alleged our output derives from their unlicensed work, could we answer the question, or would we be discovering our own exposure in real time?
- Anthropic is racing to secure 1.4 gigawatts of local capacity while its needs, and the policy, are still moving. Where in our own AI plans are we committing capital ahead of a settled position, and what is the off-ramp if that position moves against us?
- An AI procurement decision made purely on capability and price, with no one in the room having read the indemnity and liability clauses. The interesting risk in these contracts is who carries the copyright exposure, and it is rarely the vendor by default.
- A single-vendor AI dependency where that vendor's own continuity is tied to an unresolved legal and policy fight. Anthropic has explicitly made its build-out conditional; a board relying on it should treat that condition as its own continuity risk.
- Treating social licence as a communications problem rather than a governance one. When the state names it as a precondition, it is a licence-to-operate variable, and those are owned at board level, not by the marketing team.
- Assuming US fair use logic protects an Australian deployment. Treasury explicitly disputed that framing and noted the law is unsettled. Comfort borrowed from another jurisdiction's contested case law is not a control.
The upside is that this is early, and early is cheap. A board that asks for the indemnity position, a register of what its AI tools were trained on where knowable, and an honest "could we defend this publicly" test on each material deployment, is doing unglamorous work that costs almost nothing now and positions the organisation to keep using AI confidently whatever policy settlement arrives. The firms that get this right will still be deploying while more exposed competitors are pausing to untangle liability they never mapped.
The downside is asymmetric and quiet. Copyright and social-licence exposure does not show up in a management report; it arrives as a rights holder's letter, a regulator's inquiry, or a headline that makes a useful tool suddenly indefensible. A board that waved AI through on efficiency grounds, without reading who carries the legal risk or asking whether the use would survive daylight, has accepted a contingent liability it never priced, at precisely the moment the government has signalled it intends to make that liability real.
Read the condition, not just the cheque. When a vendor as sophisticated as Anthropic ties billions to getting certainty on its liability to rights holders, the disciplined board response is to ask what liability is left with the buyer, and to answer three things concretely: what our AI contracts say about copyright indemnity, what we can establish about how our tools were trained, and whether each material deployment would survive a public "earn its social licence" test. Put AI on the agenda as a licence-to-operate item, not only an efficiency one, insist on a multi-vendor position rather than a single dependency whose own future is conditional, and stop borrowing comfort from US fair-use cases the Australian Treasury has openly disputed. None of this requires resolving the copyright debate; it requires the board to know where the organisation sits inside it before someone else decides for us.
Researched and drafted by Brad's agentic AI team. Edited and published by Brad Ferris.