Why Trusting ChatGPT to Buy Your House Could Cost You $200,000

I run a team that uses AI every single day. Five of our forty staff are effectively AI-powered roles — property sourcing, HR screening, tenant communication. We're probably the most aggressive adopters of artificial intelligence in the entire Australian property sector.

So when I tell you that ChatGPT cannot buy you a house, I'm not speaking from ignorance. I'm speaking from the scars of knowing exactly where the technology breaks.

There's a seductive logic floating around social media: "AI is here, why would I pay a buyer's agent? I'll just ask ChatGPT." I've seen this take in Reddit threads, TikTok comments, even from people who should know better. And every time I read it, the actuary in me runs the same calculation and arrives at the same conclusion.

The maths doesn't work. Let me show you why.

The compounding error problem that nobody explains

Even if ChatGPT gets every individual step right 95% of the time — and that's generous — buying a property involves at least ten sequential decisions. Shortlisting suburbs. Evaluating school catchments. Comparing recent sales. Running valuations. Developing a bidding strategy. Reviewing contract terms. Checking section 32 disclosures. Assessing building condition. Negotiating the price. Managing settlement.

0.95 raised to the power of 10 equals 0.5987. Your entire chain of decisions has a 59.87% chance of being correct. That's worse than a coin flip.

AI engineers have a formal name for this: the compounding error effect. A 1% error rate per token, accumulated across 200 tokens, pushes the overall error rate to 87% 1. This isn't theoretical hand-wringing. It's been observed, measured, and published.

In practical terms, a quantity surveying firm in Melbourne ran a proper test. They fed ChatGPT a complete data set and a $1 million budget, asking it to recommend investment suburbs. More than half the recommendations were wrong. The model showed a bias toward apartments over houses, invented metrics like fictional "days on market" figures, and in several cases recommended suburbs that literally do not exist 2.

The punchline? When they switched on the deep research mode, expecting better results, the model became more confident in its wrong answers. Not less wrong. More confidently wrong.

The median price of a detached house in Melbourne sits around $900,000 3. We're not talking about picking the wrong restaurant on a Friday night. We're talking about the single largest financial commitment most people will ever make. And the tool you're relying on performs worse than chance across a full decision chain.

What a buyer's agent sees that AI literally cannot

There are three things that separate a human buyer's agent from any AI tool currently in existence. I don't say this to protect my job. I say it because the gap is structural, not temporary.

The first is room temperature.

I don't mean thermostat readings. I mean the feeling of walking into an open inspection and counting heads. How many active bidders are registering? Are they investors or owner-occupiers? Are they nervous first-timers or seasoned operators?

Here's a real example from Melbourne's west. Same suburb, same weekend, four auctions. The first property had a reserve of $850,000 and sold for $1,065,000 — a $215,000 premium — because seven to nine active bidders showed up. The other three properties in the same corridor on the same day all passed in. Nobody came 4.

If you'd asked ChatGPT what a house in that suburb was worth, it would have given you the median. One number. But on that Saturday afternoon, the actual range was $215,000 above or below that median depending entirely on who walked through the door. A buyer's agent running 40 to 60 auctions per week has an intuitive read on crowd energy that no data set can replicate.

The second is motivation.

Same house, two different sellers. One is divorcing and needs the cash out before Christmas. The other is upgrading and will wait for the right price. The negotiation strategy for these two situations is completely different. A divorce settlement sale might have $80,000 to $100,000 of negotiation room. An upgrade seller might have zero.

I've seen premium properties sit on the market for a year and then sell privately at massive discounts once the seller's circumstances changed. One property in a prestigious Melbourne suburb was listed at $9.5 million, failed to sell, and transacted privately twelve months later at $8.5 million — a million-dollar haircut 5. A buyer's agent plugged into the local network would have spotted that opportunity. ChatGPT wouldn't even know the property existed.

Which brings me to the third advantage: off-market access.

Independent industry data suggests that roughly 20% of Australian property sales happen off-market, and in certain Melbourne corridors where listing volumes have thinned, that figure may approach 25% to 30% 6. These properties never appear on realestate.com.au. ChatGPT can't analyse what it can't see.

Research from a Melbourne agency showed that off-market sellers accept an average of $50,000 less than they would through a public campaign, because the absence of competitive bidding removes emotional price escalation 7. That $50,000 discount goes straight into the buyer's pocket. But only if the buyer has access to the network — and that access comes from relationships, not algorithms.

The data lag problem nobody mentions

Most people assume that when CoreLogic publishes Melbourne's median house price, they're looking at current data. They're not.

The Victorian Department of Treasury and Finance published an analysis showing that initial monthly data releases from major property analytics firms capture only about 60% of actual transactions. The remaining 40% trickle in over subsequent months through delayed settlement reporting 8.

It gets worse. Expensive properties settle and report faster than affordable ones. So the first-release median is not just incomplete — it's biased upward. You're looking at a number that over-represents top-end sales and under-represents the suburbs where most investors actually buy.

During a busy spring auction season, median prices in active Melbourne suburbs can swing $50,000 to $100,000 within weeks. If you're using February data to bid at an April auction, you're not just working with old information. You're working with wrong information.

A buyer's agent who inspects 40 to 60 properties per week has a real-time feel for pricing that no quarterly data release can match. We know what sold last Saturday, what passed in, and what the underbidder would have paid. That's not data science. That's field intelligence. And it matters enormously when you're bidding on a $900,000 asset.

Where AI actually earns its keep

I'm not anti-AI. That would be absurd given how heavily we use it in our own operations.

AI is brilliant at the grunt work. Initial suburb screening — scanning population growth, rental yields, infrastructure pipelines, school rankings across fifty suburbs simultaneously — that takes a human analyst two weeks. AI does it in minutes.

Contract term comparison. Translating legal clauses. Pulling together comparable sales from multiple databases. Generating cash flow projections across different interest rate scenarios. For all of this, AI is ten times faster than any human.

Where AI falls apart is the last 20% of the decision chain. The bits that require physical presence, social intelligence, and access to information that isn't publicly available. The judgment call about whether this particular seller will accept $30,000 below asking. The gut read on whether that building report is hiding something the inspector didn't want to put in writing.

Smart buyers use AI for the 80% that is preparation and research. Then they hand the 20% — the decisions that actually determine whether you overpay by $50,000 or underpay by $50,000 — to someone who knows the terrain.

That's not AI versus buyer's agent. That's AI plus buyer's agent. And the combination is devastatingly effective.

The overconfidence trap you don't see coming

Researchers at a major European university ran an experiment with 500 participants. Half used ChatGPT to solve logic problems. The AI-assisted group scored higher — no surprise there. But every single participant who used AI dramatically overestimated their own ability. The more experienced they were with the tool, the worse the overestimation became 9.

The researchers called it cognitive offloading. You outsource the thinking, keep the confidence, and lose the ability to spot errors. Most people ask ChatGPT one question, accept the first answer, and never verify it.

In property, this plays out like this. Someone asks ChatGPT whether a particular suburb is a good investment. The model returns a plausible-sounding paragraph citing median prices, population growth, and infrastructure projects. The person feels informed. They bid at auction. They overpay by $40,000 because the "data" they relied on was six months stale and the "growth drivers" were pulled from a press release that was two years old.

They never realise the mistake because the model sounded so confident. That's the trap.

Our team has looked at over 350 transactions across Melbourne's southeast corridor. The properties that performed best — annual growth exceeding 15% even in a flat market — shared specific characteristics around land-to-building value ratios, proximity to infrastructure, and seller motivation that no AI model can currently assess. We found that weatherboard houses consistently outperformed brick equivalents in appreciation because the land value ratio approaches 100% and buyer competition is lower 10. That's the kind of insight that comes from reviewing 200 real cases over three hours, not from asking ChatGPT to summarise a suburb.

A detached house at $900,000 is three to five years of after-tax income for an average household. It determines your children's school, your commute, your retirement timeline. This isn't a double-eleven sale on shoes. This is the financial foundation of your family's next twenty years.

I use AI every day. Precisely because I understand its boundaries, I know what it can't do. Respecting the limits of your tools isn't a weakness. It's the first step to saving money.

I'm Yan, an actuary turned buyer's agent. And if there's one thing actuaries know, it's that probability compounds — in both directions.