On April 3 - Donald Trump's so-called "Liberation Day" - the penguins of Heard Island awoke to find themselves unexpectedly subject to US trade tariffs. Indeed, all of Australia's external territories, inhabited or not, were swept up in Trump's latest outlash at the world. Norfolk Island (population 2188), despite having no known exports to the US, received a bewildering 29 per cent tariff.
Subscribe now for unlimited access.
or signup to continue reading
Prime Minister Anthony Albanese said, "I'm not quite sure that Norfolk Island, with respect to it, is a trade competitor with the giant economy of the United States, but that just shows and exemplifies the fact that nowhere on Earth is safe from this." Except Russia, notably absent from Trump's expansive list.
Global leaders, economists and business owners scrambled to make sense of Trump's school-poster-style presentation of his latest economic plan. France, under a blanket EU tariff of 20 per cent, discovered its overseas region Reunion Island bizarrely penalised with a 73 per cent tariff. Most Asian nations faced sky-high tariffs, whereas virtually all of Latin America is at the base 10 per cent rate. Beyond the obvious flaw - that American consumers bear most tariff costs - what could possibly explain these wildly inconsistent numbers?
Finance writer James Surowiecki quickly solved the mystery, using some straightforward mathematical reverse engineering.
In short, the White House formula is: (Export to US) / (Imports from US) = Tariff rate. Additionally, only goods were included in the calculations not services.
Flexport (a global commerce company) confirmed this by posting a chart plot of the formula to the tariffs showing an exact match.
Never one to admit blame, the Trump Administration (via Kush Desai) hits back at Surowiecki saying that he was wrong, then posted the actual formula they used (also available on the official Presidential website).
It looks impressively "mathy", except, it's actually the same thing that Surowiecki and Flexport already said. Parameter values were required for "price elasticity of import demand" and "elasticity of import prices": these were assigned 4 and 0.25 respectively. For some schoolyard revision 4 x 0.25 = 1, which means they just cancel each other out.
The problem? This ratio ignores the full trade balance, excludes services (a major part of the US economy), and applies a mechanical logic to a deeply political and contextual issue. It's like deciding household expenses based only on grocery bills while ignoring rent, income or childcare - technically tidy, but wildly misleading.
This kind of thinking reflects a broader failure in model design: abstraction without accountability. When a model simplifies complex systems into elegant ratios, it risks erasing the very dynamics that matter most. And when those abstractions are elevated to policy without scrutiny or context, we move from economic modeling to economic myth-making
Online speculation soon pointed to generative AI. When tested, leading Large Language Models (LLMs) including ChatGPT, Grok, Gemini, and Claude responded similarly to this prompt prompt posted by tech writer Krishna Rohit:
"What would be an easy way to calculate the tariffs that should be imposed on other countries so that the US is on even-playing fields when it comes to trade deficit? Set minimum at 10 per cent."
I independently verified that these models still produce similar outputs. Most include caveats - acknowledging the model's limitations or the potential oversimplification - but the pattern is unmistakable. If four separate models quack the same economic policy, it's not just noise - it's time to ask who put the duck behind the presidential podium.
While not definitive proof that the administration relied on generative AI for economic policy, the fingerprints are hard to ignore. The uncanny resemblance between the White House tariff formula and a ChatGPT output is both compelling and deeply concerning.
If it is indeed true that the current White House administration is using ChatGPT and the like to help set economic policy - the impact of which wiped out US$4.5 trillion on the Wall Street stock exchange in one day - we are facing a new AI safety issue: Vibe Governing.
The term "Vibe-Governing" riffs off "Vibe-Coding," coined by OpenAI co-founder Andrej Karpathy, describing inexperienced or lazy programmers who let generative AI handle their coding - a practice "Where you fully give into the vibes" (Karpathy).
Similarly, "Vibe-Governing" suggests policymakers letting AI-generated content guide major economic or political decisions, leading us into surreal, potentially dangerous territory.
So how did Heard Island's penguins become caught up in this economic absurdity? It appears the White House simply used the ISO country code list from the International Organization for Standardization, which includes territories regardless of economic activity or even human habitation.
This fiasco underscores a deeper issue: poorly constructed mathematical models, whether AI-generated or human-designed, can amplify flawed assumptions into devastating real-world consequences. Policymakers and AI practitioners alike must remember that even simple errors in foundational formulas can have trillion-dollar impacts.
At the core of this isn't just AI misuse, it's a model design failure. An oversimplified equation, stripped of context, was treated as objective truth. Whether made by humans or machines, models are never neutral; they reflect assumptions, omissions and values. In this case, a blunt trade ratio became policy; algorithmic myopia dressed as economic logic.
This is the danger of black-box rationality: when outputs are trusted more than the framing behind them. As generative tools creep into policy and business, the real responsibility lies not in asking what the model says, but what it leaves out.
The issue here isn't just about whether Trump used AI, it's about how fragile our systems of governance become when decisions are offloaded to tools that prioritize pattern over prudence. In this new era of "vibe-governing," the challenge for AI ethics is not only preventing misinformation or bias but ensuring that technical authority does not mask weak reasoning, or worse, become a substitute for it.
As Dennis Denuto famously said in the classic 1997 Australian film The Castle: "It's the vibe of the thing, your Honour." But with global markets at stake, perhaps it's time to take "the vibe" more seriously.
- Rebecca Johnson is an expert in the ethics of generative AI (GenAI). She is in the final months of her PhD at the University of Sydney.
