Australia’s debate about sovereign artificial intelligence has largely been argued in the language of infrastructure: gigawatts, data centres, chips and the balance sheets of the hyperscalers building them. Clive Dickens wants to shift the conversation somewhere less concrete but arguably more consequential, which is what actually comes out the other end when Australians ask a machine a question.
The media and technology veteran, now attached to the venture Meliora, has been making the case that a country which outsources its entire AI stack to a handful of global models also outsources something of its character. In an interview with the advertising and media title B&T, Dickens argued that an over-reliance on offshore systems risks serving up what he calls “beige answers”, homogenised responses trained on the wider internet that flatten local nuance, context and voice. You can read his full argument in B&T’s write-up.
The pitch: ‘trained fairly in Australia’
The centrepiece of Dickens’ argument is a proposed standard he frames as “trained fairly in Australia”. The phrase does two jobs at once. It signals provenance, that a model has been built and tuned on Australian data and for Australian users, and it signals consent, that the material feeding the system was licensed or used with permission rather than scraped wholesale without payment or acknowledgement.
That second element lands squarely in the middle of the most contested question in the local sector right now. Australian publishers, musicians, artists and news organisations have spent the past two years watching large language models ingest their work, and the fight over whether copyright law should bend to accommodate AI training has already split the federal government’s own thinking. A “trained fairly” label, on Dickens’ telling, would let creators and consumers distinguish between systems that respected the rights of the people whose work they learned from and those that simply took what they could reach.
Dickens has a track record that gives the pitch some weight. He spent years in senior digital roles across Australian media, and his current focus at Meliora sits at the intersection of content, technology and the commercial machinery that connects them. He is, in other words, arguing from inside the industry that has the most to lose if training data becomes a free-for-all.
Playfully, the B&T piece notes that while Dickens is happy to champion a “trained in Australia” ethos for AI, he declined to weigh in on the Socceroos’ own “trained in Australia” question, a reminder that the sovereignty language now travels well beyond the technology pages.
Two ways to read the argument
The optimistic reading is that sovereignty is about quality and relevance as much as security. An AI system that understands the difference between a Melbourne laneway and a Brisbane arcade, that knows how Australians actually speak and what they care about, is simply more useful than one whose centre of gravity sits in California. On this view, “beige” is not a slur so much as a description of what happens when everything is averaged out across billions of documents written mostly by and for other markets. Local training, done fairly, becomes a product advantage rather than a compliance exercise.
The sceptical reading is harder to dismiss. Building and training frontier models is extraordinarily expensive, and no amount of national pride changes the economics of competing with companies spending tens of billions of dollars a year. Critics of the sovereign AI push have argued for months that Australia risks confusing hosting foreign models on local soil with genuinely building its own, a distinction FluentSea has covered as the country wrestles with whether it is truly making AI or merely renting it. A “trained fairly in Australia” badge only means something if there are Australian-trained models substantial enough to carry it, and that remains an open question.
There is also a definitional problem. Fairness in training data is easy to assert and fiendishly hard to certify. Who audits the claim? What counts as fair payment to a musician or a news publisher whose catalogue helped shape a model? Without a credible verification mechanism, a label risks becoming marketing rather than assurance, the AI equivalent of a vague “sustainably sourced” sticker.
What it means for Australia
Dickens’ intervention arrives at a moment when the policy scaffolding is finally being built. The federal government has stood up a national Office of AI and is edging towards a broader framework, while the states court offshore investment in the enormous data-centre projects that any sovereign ambition would depend on. The “beige answers” framing is useful because it reconnects those big-ticket infrastructure decisions to a question ordinary Australians can grasp: when a student, a small-business owner or a nurse asks an AI for help, whose values and whose knowledge are baked into the reply?
For the creative and media sectors the stakes are immediate. If a “trained fairly” standard gained traction with regulators or major buyers, it could hand Australian rights holders leverage they currently lack, effectively turning consent into a competitive credential. For the technology sector the challenge is the reverse, because meeting such a standard would require local models good enough that buyers would actually choose them, and that in turn requires capital, compute and talent the country is still assembling. The two agendas can complement each other, but only if the money and the policy line up.
What’s next
None of this is settled. There is no legislated definition of a “trained fairly in Australia” standard, no independent body positioned to certify it and no consensus on how the underlying copyright question should be resolved. What Dickens has done is give a slogan to an instinct many in the local industry already share, that sovereignty should be measured not only in racks and megawatts but in whether the answers Australians receive sound like they came from here.
Expect the phrase to resurface as the Office of AI fleshes out its framework and as publishers and artists keep pressing their case on training data. Whether “beige” becomes a rallying cry or a footnote will depend less on rhetoric than on whether anyone can build an Australian model distinctive enough to prove the point.
Sources: B&T.
