An Australian AI Lab Just Blew Up Big Tech’s Main Argument
At Australia’s Federal AI Inquiry this week, Google told senators that requiring AI companies to get consent before training on copyrighted material would make AI development “impossible.” Microsoft said legal uncertainty could drive investment out of the country. Anthropic’s Jeff Bleich argued “you can’t license the entire internet.”
Then a small Australian AI lab walked in and said: we’re already doing it.
Trellis Data, backed by Main Sequence Ventures, told the inquiry that Australian developers can build competitive large language models without a copyright carve-out. Not in theory. In practice. The company is profitable this year, including its R&D spend, and its own LLM is expected to launch within a month. No exemption from copyright law required.
Senator David Pocock has been pointing to exactly this kind of contradiction. Earlier this week, he said big tech was “dictating terms to every single Australian.” After Trellis testified, it’s hard to disagree with his read.
What “impossible” actually means
When Google says an opt-in copyright regime makes AI development “impossible,” it means impossible at the margins big tech currently operates at, not impossible in any meaningful engineering sense.
The difference matters. Building a frontier model at OpenAI scale requires vast quantities of training data scraped at speed. Licensing that data through any kind of permission framework would be slow and expensive. So “impossible” is accurate if what you mean is “cannot be done without significantly raising costs and slowing timelines.”
That is not what most people hear when a company says something is impossible.
Trellis Data’s chief executive Michael Gately put his company’s position plainly: copyright holders should be paid, artists should be paid, and the government should establish a licensing or brokering mechanism to make that happen. His company has built competitive AI working within existing copyright rules. The idea that it cannot be done is, according to the people who just did it, wrong.
This is not a minor quibble about wording. The entire legislative case for a copyright carve-out rests on necessity. If it is not necessary, the argument collapses.
The shape of the ask
What the big tech companies want is a broad training data exemption. Any content accessible online could be scraped and used for model training unless the creator explicitly opts out. The burden shifts from the company to the creator.
ARIA, the peak body for Australian music, called Australia a potential “Trojan horse” for global AI policy. Win the carve-out here, set the international precedent, watch other countries follow. It is a pointed thing to say, but the strategic logic holds. Australia is a smaller market with a government that has a history of being responsive to investment arguments. If you are going to test a regulatory ask, this is not a bad place to start.
The ABC took the opposite position from Google at the same inquiry. Existing copyright law is capable of handling AI, the public broadcaster argued. The problem is enforcement, not the statute. No reform needed.
Between the ABC, the Writers’ Guild, ARIA, and now Trellis Data, the counter-argument at the inquiry was not a single dissenting voice. It was a pattern.
Why this matters for small businesses
Australian small businesses are not spectators in this fight between creators and tech giants.
The content your business produces every day sits in this argument. Blog posts, product descriptions, case studies, testimonials, how-to videos. A broad training data exemption means any of that could be swept into a model’s training run without your knowledge or consent. You would not be asked. You would not be paid. The only way to prevent it would be to figure out each platform’s opt-out mechanism and hope it works.
The creative freelancers your business hires sit in it too. Photographers, copywriters, designers. Their ability to charge for their work depends on having some control over how it is used. Erode that control and you change the economics of the freelance market they operate in.
For more on what Australia’s AI copyright fight means in practice, see our earlier look at the inquiry and what it means for small business owners.
The Trellis moment
At the inquiry, Trellis Data did something that was straightforward but pointed. They showed up, described what they had built, and declined to use the “we need a carve-out” argument that every other AI company had deployed.
Their model is not yet public. Their LLM is weeks from launch. They are a small team against companies with market caps that would dwarf Australia’s GDP. And their position is: we can do this properly, the rules exist for a reason, let’s build a licensing mechanism instead of abolishing the requirement to license.
That is not a counterpoint to Google’s argument. A counterpoint suggests a similar level of abstraction, two arguments at odds with each other. What Trellis offered was empirical evidence against a factual claim. Google said it was impossible. Trellis said they are doing it.
The inquiry continues. Submissions close, recommendations get written, and the big tech companies have lobbyists in Canberra who will be working the process through to the end. But the “impossible” framing took real damage this week, from a company that had no reason to make that argument except that it happens to be true.
Senator Pocock’s read was right. What has been happening at the Federal AI Inquiry is not primarily a technical debate about model training pipelines. It is a negotiation, with real stakes for creators, for Australian AI companies, and for any business that produces content online.
Trellis Data just changed the terms of that negotiation.
Sources
- Australian AI lab challenges case for copyright carve-out (Capital Brief)
- Big Tech’s copyright demands at the AI Inquiry (SmallBizAI.au)
- Federal AI Inquiry: what it means for Australian small businesses (SmallBizAI.au)
More hot takes: Hot Takes →
📖 More deep dives: News Deep Dives →
