AI Training Consent Creator Rights: It Is Not a Checkbox

There is a word being used a lot this week: consent. The Prime Minister said no company should use Australian creative works for AI training without the creator’s control. The Australian Writers’ Guild called for a framework built on consent, credit and compensation. ARIA said artists must control the price, the value and the terms. Everyone agrees consent matters. Almost nobody is talking about what consent actually looks like as a record.

ScriptShield has been working on exactly this problem. Because consent is not a toggle. It is not a pop-up. It is not a checkbox buried in a platform’s terms of service that you clicked through at two in the morning. Real AI training consent is specific, informed, documented and verifiable. And right now, the infrastructure to record that kind of consent barely exists.

This article is about what needs to change — and what you can do while it does.

---

What did the Prime Minister actually say about consent?

The speech was unusually direct. At the University of Sydney on 15 July 2026, Anthony Albanese said Australian writers, musicians, artists and journalists must retain ownership and control of their work. He said no company should use Australian books, music, art or news to build or train AI without the artist’s control — including the artist’s control of the price and value of their work.

That last part is critical. The PM did not say creators should be able to opt out. He said creators should control the price. That is a fundamentally different proposition. An opt-out says “no.” Price control says “yes, on my terms.”

Dean Ormston, CEO of APRA AMCOS, pushed the point further. He called for the Office of AI to turn policy certainty into “serious industry-to-industry licensing negotiations between platforms and rights holders, not further rounds of tech sector avoidance.” He said Australian creative IP must be treated as “an integral, appreciating national asset in its own right, not a line item to be settled quickly and cheaply.”

The direction is clear. The infrastructure is not.

---

Why is a simple opt-out not enough?

Because a binary choice — yes or no — does not reflect how creative work is actually used.

Consider a novelist. She might be comfortable with her published synopsis appearing in search results. She might accept a research institution using her backlist for non-commercial literary analysis. She might refuse to let a commercial AI company train a generative model on her unpublished manuscripts. And she might welcome a paid licence for a streaming platform’s recommendation engine to summarise her catalogue.

That is four different answers for four different uses of the same body of work. A single checkbox cannot hold any of them.

The EU’s approach under the Copyright Directive allows creators to reserve rights through machine-readable signals — TDMRep, robots.txt extensions, C2PA assertions. These are genuine steps forward. But they are binary mechanisms. They say “reserved” or “not reserved.” They do not say “permitted for research at this price with attribution for twelve months.”

Australia rejected the text-and-data-mining exemption that the EU adopted. The Attorney-General’s Department is exploring a paid licensing model instead. That is more ambitious — and it requires more sophisticated consent records than any current standard provides.

---

What should a real AI training consent record contain?

This is where ScriptShield’s architecture matters. A meaningful permission record should answer every question that would arise in a licensing dispute, an audit or a court proceeding.

---

What tools exist today for recording AI training consent?

The honest answer is that the landscape is fragmented. No single standard handles everything. But several are worth knowing about because they represent the foundation of what will eventually become a coherent system.

C2PA content credentials can now embed training and data-mining assertions directly into media files. A creator can signal whether an asset may be used for data mining, AI inference, model training or generative training. C2PA 2.4 supports text, images, audio and video. The limitation is that these signals are advisory — an AI company can technically strip or ignore them.

TDMRep provides machine-readable rights reservations for web content. It works well for publishers who control their own websites. It does not help a creator whose work has been reposted to a third-party platform.

Really Simple Licensing goes further than any other standard by supporting specific use cases and payment types — per-crawl, per-inference, subscription. It is early-stage but moving in the right direction.

Spawning’s Do Not Train registry aggregates opt-out preferences and is respected by several AI companies. It is practical, accessible and operational — but it remains an opt-out system, not a licensing platform.

ScriptShield connects to these standards rather than replacing them. The Rights Graph adds what none of them provide alone: a persistent, creator-controlled record linking the work’s provenance to its permissions, declarations, version history and downstream use.

---

What should creators do while the consent infrastructure catches up?

Do not wait for a perfect system. Build your record now with the tools that exist.

---

The Bottom Line

The PM drew the line. Creators control the work, the use and the price. That is the right policy. But policy without infrastructure is a promise without a mechanism. The gap between “you have the right to consent” and “you have a verified, auditable, machine-readable consent record” is the gap ScriptShield is built to close.

Start building your record now. Because when the licensing framework arrives, the creators with documentation will negotiate. The creators without it will hope.