Hundreds of millions of vibes.

Whenever some new technology rolls around, there's never a shortage of people willing to claim it's the end times; that the future of work and employment as we know it is over. The latest such prophet is The Australia Institute's Richard Denniss, who went viral for telling Australia's Joint Select Committee on Artificial Intelligence that "the business model of AI is to generate mass unemployment":

"As an economist, I understand productivity. People lose jobs when technology changes. But the valuations of these AI companies are literally based on the assumption that hundreds of millions of people will lose their jobs and employers will buy AI services instead of labour services. Where is the money coming from? It comes from not paying workers. The only way the Department of Health can spend a billion dollars on AI is if it spends a billion dollars less on humans, or if you guys increase their budget by a billion dollars."

Right. Let's start with what Denniss got right: that "the money" comes from "not paying workers". That's basic arithmetic; if a health department spends a billion dollars on AI, it must either spend a billion dollars less on workers (its main expense), or have its budget raised by a billion dollars. I'm of course assuming Denniss means a state health department; the Commonwealth one is mostly a funding body that employs few nurses.

Now, I don't for a minute think a health department would be able to replace nurses and doctors with AI, like-for-like. But suppose some advance in both AI and robotics makes it possible. A hospital that swaps a billion dollars of labour for a billion dollars of AI spending frees up a scarce resource – labour – for other uses. That's a tale as old as time, and it's why so many of us can earn a decent wage typing on keyboards in air conditioned offices instead of toiling in the fields.

What Denniss doesn't mention is that his conclusion of mass unemployment only holds if wants are finite: that once people have enough, no further work is worth paying for. That's known as the lump of labour fallacy, and it's a fallacy for a good reason: nobody has ever found a society that ran out of things to want. The hospital using more AI doesn't delete the demand for labour, it relocates it, and Denniss' arithmetic stops before that's even allowed to happen.

Let's give Denniss the benefit of the doubt here. If AI destroyed income on the scale implied by his claim, who buys all of that new output? It's plausible that there's a nominal shortfall during an extremely fast transition, even if monetary policy responds appropriately. In that situation you'd likely get a jump in unemployment as existing patterns of commerce and human skills fall out of sync with market realities, and it takes time for new patterns to emerge. But in such a world you also have growth going gangbusters, which is "a good problem to have".

Now, there is an internally coherent argument that Denniss never mentions: output rises enormously, and whoever controls the bottleneck captures nearly all of it, so demand for labour collapses even as the economy grows. Justin Wolfers has suggested such a scenario, and it's worth spelling out because it's where the catastrophe would actually have to come from. Basically, if one model is sufficiently better than its rivals, its owner can charge the employer almost exactly what the displaced worker used to earn. The employer saves a dollar, the worker loses a week's wages, and the surplus lands with the model's owner, with nothing creating demand for the person who was replaced.

The difference between Denniss and Wolfers is that the latter calls the outcome a policy choice – "it's either you or your boss" – not a technological destiny.

Back when I wrote about it last year, I thought Wolfers got the mechanism wrong. He seemed to assume the models would be the bottleneck forever, when the bottleneck was already moving from models to chips, to energy, and to land, so no single firm was likely to end up owning close to all of GDP. Since then, open weight models out of China and from Meta have reaffirmed my point.

The problem with Denniss is that he goes further, taking the same machinery, stopping at the worst distributional outcome, and presenting it as a forecast, with competitive forces and adaptation in both markets and politics assumed away.

My own view is that predictions of even modest disruption are overstated. Adoption is not the same as displacement, and putting a tool to work inside an existing organisation means rewriting processes, retraining staff, renegotiating contracts and absorbing liability, and firms do all of that slowly and unevenly. Regulation, procurement rules and plain organisational inertia each take a cut. If Denniss had ever seen how a government department works from the inside, he wouldn't be worried about a rapid internal shift from people to AI.

Institutions change slowly, and it won't be any different with AI. People like Denniss should attach numbers and dates to their apocalyptic forecasts, so there's an actual point at which they can be declared wrong.

I suspect he never will, because a prediction that can't be tested never has to be retired, and can be used to support almost any policy The Australia Institute wants to hang on it.