The Machine Is Friendly. That's Why It Wins.
How Australia kept its third places alive, what it hid inside them, why Britain’s fix made the problem worse, and what happened when I tried to beat the machine.
I don’t gamble. It’s not that I have a principled objection against it, but I’m just incapable of finding it appealing. I once turned down a bet at near-certain odds and I keep my savings in a low-interest account rather than an index fund or stock market (because the variance due to true social posts bothers me). If there were a machine that dispensed guaranteed small amounts of money, I would use it very enthusiastically though.
So perhaps it is a little strange that I am sitting in the Cronulla RSL, running gambling models while surfers outside pick off the hardest waves and, thirty metres behind me, the actual machines hum.
I am not here to tell you gambling is bad. I want to write about something more specific: that in a small outback town called Balranald, gambling machines extracted AUS$3,961 last year from every man, woman, and child in a town of 2,209 people. Once my models failed to explain that number, I realised that I had misunderstood something important about Australia (despite being an Aussie national and having come here since I was three months old).
The ‘pokie’ room at Cronulla RSL, thirty metres from the bistro and the sea view.
The rooms that closed
Sociologists have a name for the spaces that are neither home nor work, but where community actually forms: the third place. The pub, the library, the bowling club, even the post office. The place you go because it is there, other people are there, and that is enough. For many people, that is exactly what the Cronulla RSL is.
Britain lost these at scale. Austerity closed the libraries, Amazon closed the post offices, property prices and private equity closed the pubs: 14,000 of them between 2001 and 2021, one every fourteen hours for two decades. I wrote about this a few months ago: “Britain Lost 14,000 Third Places. They Were Called Pubs. Is Your Local Next?” The statistic that stayed with me most was what replaced them. For most communities that was nothing: the room closed and stayed closed.
This is the baseline to keep in mind, because Australia made a different choice.
The deal
In the 1950s, Australian state governments cut a deal with registered clubs such as RSLs (Returned Services League, the veterans’ organisation, the equivalent of the British Legion), football clubs, bowling clubs. The deal was simple: you get the right to operate gambling machines, (referred to often as pokies) and in exchange, you fund community services and keep the doors open.
It worked: the rooms stayed lit. The RSL stayed open and the third place which Britain was losing, Australia kept. Nobody looked too hard at the extraction rate but the clubs were drawing 40 to 60 percent of their income from the machines, and had become so embedded in community life that any threat to the pokies read, politically, was a threat to the community itself. The machines and the rooms had become the same object: you could not remove one without destroying the other.
The Cronulla RSL has a bistro, sports bar, sea views, ANZAC photographs on the wall and a Sunday afternoon crowd so relaxed that you cannot for a second accidently think that you are in London. It is also, behind me, running a variable ratio reinforcement schedule which is the same psychological mechanism as the pull-to-refresh on your phone, resulting in people who came in for a quiet afternoon ending up behind a screen. B.F. Skinner already documented this in rats in the 1950s.
The machines are legally required to return at least 87% of turnover, but turnover is not the same thing as the amount you first put in. Every small win is usually fed back into the next spin, so over the course of a session a $50 note can be churned through the machine several times. The machine is therefore returning 87% or more of that larger recycled total, not 87% of your original $50.
Now, Australia loses more money per capita to gambling than any other country on earth: roughly AUD $1,300 per adult per year, nearly double the next highest. Over 620,000 Australians are classified as problem gamblers and a further 2.9 million are at-risk gamblers: together about one in six adults. You could see it as the bill for keeping the third spaces open.
So I opened my laptop and tried to measure the bill.
Where the gambling takes place
The first question I wanted to answer was simple: where does the money actually go?
I downloaded the NSW Independent Liquor and Gaming Authority’s full venue register: 2,184 licensed venues, machine counts, and six months of net profit figures. ‘Net profit’ is the industry’s preferred term for ‘money extracted from the public,’ which is excellent branding if you are the industry. I joined this to Australian Bureau of Statistics deprivation and rurality scores by Local Government Area.
I first let the data sort places into a few broad types. It is good at telling you what kinds of places exist but less good at telling you why a particular place sits where it does, or whether something unusual is happening inside an otherwise familiar pattern. To answer that question, I needed a different tool.
I then trained a gradient boosting model on deprivation, rurality, density, and income to see how much of the pattern the obvious structural factors could explain. It explained about 71% of the variance in extraction rates. The more interesting part is the 29% it can’t explain though. I turned that unexplained remainder: the gap between what the model predicts and what actually happens, into a vulnerability score. It measures how much a place is being extracted above and beyond what its observable circumstances would predict. A high score means that something the model cannot see is making that place unusually exposed.
Balranald scored 98 out of 100. Every structural variable said it should be close to average, instead the extraction rate was $3,961 per resident in six months, and no feature in my model explained it.
At that point the model had done its job, and I had to do mine. The explanation brought me straight back to the idea of the third place. In a town of 2,209 people, the club is the only room where the lights are on on a Friday night. The machine has no competition because the club has no competition. It is not merely somewhere to gamble, it is the only place left to be. In other words, the machine does not only go where people are poor, it goes where people are captive.
Once I knew where the money was going, the next question was: who was taking it? The LGA (Australia’s version of a local authority area) picture above tells you where the losses are concentrated while the venue picture tells you which institutions are collecting them.
To look into that, I ran a sentence-transformer across all 2,184 venue names, sorting them into five types: independent hotels, large commercial groups, RSL and ex-services clubs, general community clubs, and ethnic and faith community clubs. Sentence transformers are language models that turn text into numerical vectors capturing semantic meaning, so “Slovenian Cultural Club” and “Polish Community Centre” end up close together in the mathematical space even if they share no words.
I found that small independent clubs, community anchors, RSL and ex-services clubs have near-identical medians, clustered very tightly. Then the ethnic and faith community category: median $73,000 profit per machine, nearly double every other type.
Dooleys Lidcombe Catholic Club, Cabra-Vale Ex-Active Servicemen’s Club, Triglav Mounties Group in St Johns Park, a Slovenian community club in Fairfield, are places where people speak their first language, eat their heritage food, maintain connections the broader city does not provide. Which is why the machines inside them extract so efficiently: the social cost of leaving mid-session is self-exclusion from the community itself. Again, I find that people are walking away from a gambling machine, would be actually walking away from the community space.
By this point, three things were becoming clear. The first model had shown the shape of the gambling landscape. The gradient boosting model had shown what that landscape could not explain. The venue classifier had shown which kinds of institutions were most effective at turning social belonging into revenue.
There was one last question left: which individual venues looked strange even by those standards?
For that, I used the bluntest of the three models: an Isolation Forest. It doesn’t try to predict anything or explain anything, it rather learns what normal looks like across the full dataset: normal machine counts, normal revenue per machine, normal venue type and flags the observations that are hardest to fit into that picture. The logic is that you repeatedly cut the data with random splits until each point is isolated in its own partition. Points that are genuinely unusual get isolated quickly, in very few cuts, because they’re far from everything else.
It kept pointing at Markets Hotel in Homebush West which has thirty machines which earned $228,000 profit per machine in six months. This is nearly four times the rate of Mounties Group, which has 603 machines and I could not explain it with structural features.
By then, all three approaches were converging on the same conclusion from different directions: the machine doesn't go where people are poor, it goes where people are captive. Captivity, it turns out, looks exactly like a third place, the room that stayed open because the machine was inside it. Before I tell you what Britain already learned, I want you to see the pattern for yourself. Here is the interactive gambling map I developed: the NSW gambling map.
Toggle to revenue per machine and watch the map reorganise: the Western Corridor fades, the bright spots shift west and north: smaller towns, thinner dots, places that don’t appear in the headlines. Then drag the vulnerability slider to 80 and above: Sydney nearly empties and the outback shows. Find the dot on the flat plain west of the Riverina: twelve machines with a population 2,209 and $3,961 extracted per head. That’s Balranald.
What Britain already learned
Now that I hopefully convinced you to see the machine as an captivity optimisation system, the next question is: what happens when you regulate it? Britain already ran that experiment, and the result is very useful for Canberra.
Britain’s version of the machine was the Fixed Odds Betting Terminal (FOBT): a touchscreen in a high-street bookmaker letting you stake £100 every twenty seconds on virtual roulette. At their peak, 35,000 FOBTs generated £1.7 billion annually. A single machine in a Ladbrokes on Bethnal Green Road could take £1,000 in an hour. One study found machines in the most deprived 10% of English neighbourhoods generating twice the revenue of those in the least deprived, mainly because the sessions were longer and the losses higher.
Where Australia’s machines sat inside community institutions funded by nostalgia and social obligation, Britain’s sat in shops with blacked-out windows funded by footfall and proximity to deprived neighbourhoods. After years of campaigning and parliamentary inquiries, the government cut the maximum FOBT stake from £100 to £2. Bookmakers closed shops, and to many people who had worked very hard for a very long time, it felt like a victory.
Betting shops fell 42%. Yet, online gambling yield nearly quadrupled: from £1.8 billion in 2013 to £6.6 billion in 2023. The lines cross almost exactly at April 2019, the month the reform came into force.
Due to this reform, the industry did not lose revenue though, instead it migrated the customer and kept the money. It even added a welcome bonus and became available at 3am in bed while the shops had closing times. Basically, the reform didn’t fail strictly speaking, it actually worked but in working, made the underlying problem structurally worse. It constrained supply in one channel and demand found another.
The extra worrying thing? Online sports betting in Australia is already legal, growing, and marketed during every NRL broadcast I’ve seen including those sponsored by the clubs whose pokies produced the numbers above. The NSW cashless gambling card pilot is a solid reform, worth fighting for, but you should take into account that if it reduces physical machine revenue, demand does not disappear.
Side quest: can I beat it?
Britain’s reform pushed gambling online, and Australia’s is about to. Which raises an another curious question for someone with a laptop, a VS Code window, and too much time on a Sunday: if the machine has moved to an app, can it be treated less like gambling and more like a pricing problem?
Online bookmakers are, in one narrow but important sense, a market. Each bookmaker publishes odds that imply a probability. If Bet365 offers 2.10 on an NRL team, it is effectively saying that team has a 47.6% chance of winning. Add up the implied probabilities across all outcomes and you get a number above 100%, usually something like 103 to 108%. That excess is the overround, which is the bookmaker’s built-in margin. The house does not need to be right every time, it just needs to be consistently less wrong than you.
The arbitrage scanner I am building does the following. It pulls live odds from six bookmakers at once: TAB, Bet365, Sportsbet, Ladbrokes Australia, Ladbrokes UK, and William Hill, then converts them into implied probabilities, and looks for moments when the combined overround across different books on different sides of the same event drops below 100%. When that happens, you can bet both sides and lock in a profit regardless of the result. The window usually closes within minutes as the books adjust, so the scanner runs continuously and flags opportunities in real time.
I found, using several weeks of data on Australian and British markets, that the edges are there, consistent, and small. Margins between TAB and Bet365 on NRL markets typically ran from 0.5% to 2.1%. On Premier League lines between Ladbrokes UK and Sportsbet, I even found gaps as high as 3.4%. I would say that is more than just noise.
Arbitrage doesn’t per se tell you whether you are actually good at reading the market though. For that, the useful metric is closing line value (CLV). The closing line is the final price a bookmaker offers before kick-off. It is the closest thing these markets have to a consensus estimate of true probability, because by that point the maximum amount of information, including sharp money, has been absorbed. If you consistently beat the closing line, you were ahead of the market when you placed the bet, whether or not the individual bet won.
So I also built a second model. It estimates true probabilities from a weighted combination of closing lines across markets, then flags any live line that sits outside a confidence interval around that estimate. In plain English, it looks for prices that are better than they should be and checks whether I got there before the rest of the market did.
My model works and that, it turns out, is the problem.
There is one category of bettor the industry cannot tolerate: the one who is actually right. Sharp bettors, that is, people who identify mispriced lines and consistently bet into them, are restricted or banned as soon as they are recognised as profitable. A former industry employee told me recently about a friend who made serious money this way before the accounts dried up, one by one, as each bookmaker noticed the pattern. The system is perfectly happy for you to gamble but is much less enthusiastic if you treat it as a pricing problem.
That is what makes bookmakers different from prediction markets. In a prediction market, informed money is the point. If someone with a better model bets into a mispriced line, the price improves and the market becomes more useful. The sharp bettor profits, and everyone else gets a better estimate. The bookmaker model works the other way around: it depends on most customers being wrong in roughly predictable ways.
(As a side, with caution Polymarket can be used as a good proxy for several things since people with money on the line arguably have a stronger incentive to be right than pollsters, pundits and anyone responding to a survey. The sharp money mechanism means informed bettors push the price toward truth and profit as they do so. We could see this in the 2024 US elections and even the timing of Israel’s attack on Iran. But my own work shows that it didn’t work for predicting the next pope though;)
This is the online version of the same logic that governs the machines. The physical system extracts most effectively from the people who cannot leave. The digital system removes the people who can calculate. In other words, it is a different channel with the same principle: the market is open only to the extent that you lose in the approved way.
What to do about it?
If the machine is funding the room, then the solution cannot just be to attack the machine and hope the room survives. Australia needs a transition plan that pays for third places directly. That could mean a public Third Place Transition Fund for clubs that reduce machine dependence, a buyback scheme for machine entitlements in the most captive towns, and a captivity index that targets reform where the club is functionally the only remaining room.
My personal favourite option though would be to create a second room. Think of late-opening cafes, youth spaces, outdoor cinema nights, subsidised sports halls, cheap transport to nearby towns, all of these are more motivating than gambling reform, but they go directly at captivity. They would create competition (an economist’s favourite remedy) for the club that currently has none.
My side quest also suggests another lesson. Online betting is not some cleaner, more rational successor to the pokies. Just like the pokies, it behaves like a market until someone is actually good at it, who then gets banned. So, please policymakers, tackle online betting markets simultaneously and don’t treat this reform in isolation. Britain has shown why. A simple start would be to prohibit advertising and sponsoring (apologies to my favourite cycling team, the Unibet Rose Rockets, but you are not helping) and introduce affordability checks and identity-linked loss limits across the various operators.
The machine is friendly
Balranald is a long way from Canterbury-Bankstown and a long way from my beloved Islington. It is a small town whose population has been declining for thirty years. On a Friday night, the club is the only room where the lights are on. In six months, its machines extracted nearly $4,000 from every man, woman, and child in the town.
Driving through small towns between Sydney and Melbourne, I kept seeing the same pattern: empty main streets, shops closing early, and one club still lit up and full. Australia did not solve the problem of disappearing third places, it financed them. The people who run the club are not villains, and the people who use the machines are not fools. They are operating inside a system now so load-bearing for community life that nobody in politics can touch it without being seen to close the room.
Britain lost its rooms and got nothing. Australia kept its rooms and hid the cost inside them. The system expels the people who find the edge and keeps the people who cannot leave. The machines are still humming behind me. Outside, the surfers are still taking the difficult waves.
The machine is friendly. That is the point.
If you found this interesting, please share it. And if you want to support my Substack project while I’m travelling through Australia and turning somewhat questionable side quests into posts, you can buy me a coffee or become a paid subscriber.
I just think you are awesome. I want people like you in a world government. And when they select those that get to escape Earth from the meteorite in a rocket, I vote for you.
Interesting piece!