In this article
Kooky
Builder of Shaka, the payment router that pays every agent their commission on closing date.
About Kooky and Shaka →A property description used to take an agent the better part of an hour: a walk through the house, a page of notes, three attempts at a first sentence. A text generator now produces four paragraphs in the time it takes to paste in the bedroom count. The copy reads well. It has rhythm, it flatters the kitchen, it finishes with the lifestyle. Some of it may also be untrue, and nothing in the way the software works will flag which part.
That is the question this guide follows. It explains how generative tools assemble a description, the kinds of error that turn up in property copy and why, what Queensland and national law say about a false statement in an advertisement, who is answerable when the words came from software, and how an agency can check a draft before it is published. No product is named or compared here: the mechanics and the legal position are the same whichever tool sits on the desk.
Section 212 maximum as set out in an REIQ article of 27 June 2023; Australian Consumer Law maximums as published by the Queensland Office of Fair Trading on its property advertising page.
How a text generator produces a description
The tools that write listing copy are built on large language models. The name sounds grand, and the principle is simple to state. In an explainer published by TechRadar on 21 March 2025, a software engineer described these models as next-word predictors: they produce text by choosing the word that is statistically most likely to come next, then the next one, then the next. The same article calls them sophisticated pattern predictors, as opposed to systems that reason about the world.
For a property description, the pattern is easy to picture. The model has absorbed an enormous quantity of writing, real estate advertising included. Given "four-bedroom family home in a quiet street", it continues in the way such sentences usually continue. Family homes in listing copy are usually near schools. Quiet streets are usually leafy. Kitchens usually have stone benchtops, and outdoor areas are usually perfect for entertaining. The model is not looking at the house. It is completing a genre.
Related readGenerative AI on a conveyancing file: what Queensland's rules sayThis is why the output is so fluent, and also why fluency tells the reader nothing about accuracy. A sentence that is common in advertisements is, to the model, a good sentence, whether or not it describes this particular property.
Why the tool does not know when it is wrong
The TechRadar explainer sets out the causes of what the industry calls hallucination, meaning confident statements that are false. Four of them matter for anyone drafting an advertisement.
The first is that the models have no internal fact-checking mechanism. In the article's words they favour "lucid-sounding" sequences over correct ones. There is no step at which the software compares its sentence with the property.
The second is gap-filling. When the information available is incomplete, conflicting or insufficient, the model generates plausible but incorrect material to fill the gap. A prompt that gives the suburb but not the distance to the station invites the model to supply a distance.
The third is misapplication: the model may hold the right information and still fail to retrieve or apply it properly, particularly when the context is misread.
The fourth is the habit of answering anyway. A model defaults to producing a response and does not stop to say that it has no idea how far the beach is.
The REIQ makes the same point in terms an agent will recognise. An article on artificial intelligence published by the institute on 2 February 2026, written by a special counsel at Carter Newell Lawyers, describes generative AI as convincingly filling informational gaps with plausible-sounding details that may be inaccurate or fictitious.
Related readA rent figure from software: what Queensland's Act asks before it runsEvery description begins with what the agent types or uploads: the address or suburb, the number of bedrooms and bathrooms, a list of features, sometimes the notes from the appraisal, sometimes photographs. Whatever is missing from that input is exactly where the gap-filling described above takes place.
Three habits make the gap wider. A short prompt ("write a listing for a three-bedroom unit in a coastal suburb") leaves almost everything to the model. A request for a certain length forces it to find more to say than the facts supplied can carry. And a request for a style ("luxury", "family", "investor") pulls in the stock claims of that style: the resort pool, the school run, the rental return.
The opposite habits narrow it. A draft built from a complete, checked fact sheet has less room to invent. An instruction to use only the facts supplied reduces the padding, though it does not remove the risk, because the model, as the sources above explain, cannot verify its own compliance. The practical conclusion is modest: a careful prompt lowers the number of errors, and only a person comparing the text with the property finds the ones that remain.
Where the errors turn up in property copy
The errors are not random. They cluster where listing copy is most formulaic and where the facts are most specific to one address. The table groups them by type. It is an illustration of the mechanism described by the sources above, not a count of real cases.
| Error type | How it arises | What settles it |
|---|---|---|
| Invented feature | The genre expects it: ducted air, stone benchtops, a second living area. | The inspection notes and the seller's own list. |
| Wrong distance | No distance was supplied, so a typical one is produced. | A measured route, stated as a distance or a time. |
| School zone | Family copy usually names a school; the model picks a likely or imaginary one. | The official catchment information for the address. |
| Development potential | Investor copy usually promises it. | The planning scheme, covenants and easements. |
| Views and aspect | Suburb names suggest water, hills or city lights. | What can be seen from the property itself. |
| Condition and age | "Renovated" and "as new" are high-frequency words. | Dates and documents for the work done. |
Two rows deserve a word more. Distances fail in a particular way: "five minutes to the station" is not false in the way an invented pool is false, yet it creates a belief about walking or driving time that the buyer will test on the first visit. And development potential is the row with the most money attached. The REIQ's article of 27 June 2023 on misleading conduct in advertising says that statements about how land might be used must take account of legal constraints, and it names planning requirements, restrictive covenants and easements. It suggests wording such as "Subject to Planning Consent" together with encouragement to make independent enquiries. A model asked for investor copy knows none of this about the lot in question.
Related readVirtual staging and AI-edited photos: what the law allowsThe invented schools: a documented case
The school row of the table is not hypothetical. The REIQ's February 2026 article refers to a real incident in which an AI-generated listing fabricated schools that do not exist. The article does not need to say more for the mechanism to be clear. A family home was being described; family homes are near schools; the model wrote the sentence and supplied names to make it complete.
School claims are worth singling out for a second reason. The REIQ's 2023 article lists school catchment areas, along with the suburb and local amenities, among the location details an agent must describe accurately. A catchment is a line on a map, and a home can sit a street outside it. A buyer who chooses a house for a school has relied on the advertisement in the most direct way possible. A real school named for the wrong address is as much a problem as an imaginary one.
What Queensland law says about a false description
Two laws apply to the wording of a Queensland property advertisement. The first is the state's own licensing statute, the Property Occupations Act 2014.
According to the REIQ's 2023 article, section 212 of that Act prohibits licensees and real estate salespeople from making false or misleading representations about the sale, exchange or letting of property, with a maximum penalty of 540 penalty units. Three further parts of the section, as the article sets them out, explain why machine-drafted copy is exposed.
Section 212(3) defines the test. A representation is misleading if it would reasonably tend to lead to a belief in a state of affairs that does not in fact exist, whether or not the representation says outright that it exists. Suggestion is enough: "stroll to the village" need not give a distance to create a belief about one.
Related readAgency chatbots and tenants: what an automated reply cannot doSection 212(4) deals with grounds. A representation made without reasonable grounds is taken to be misleading. A sentence produced by a text predictor, and not compared with anything, is hard to describe as resting on grounds of any kind.
Section 212(5) places the burden on the person who made the representation to show that reasonable grounds existed. The practical meaning is that an agent asked about a line in an advertisement needs to be able to point to where the fact came from.
The national law that sits on top
The second law is the Australian Consumer Law, which applies across the country. The Queensland Office of Fair Trading's page on property advertising states that licensed agents must not engage in misleading or deceptive conduct, and names both the consumer law and the Property Occupations Act as the legislation behind that rule.
Section 18 of the consumer law prohibits conduct in trade or commerce that is misleading or deceptive, or likely to mislead or deceive. The REIQ's 2023 article describes this as a broad concept that covers words, actions and pictures, and it stresses one feature above the others: intention is irrelevant. An advertisement can mislead although nobody meant it to. Section 30 of the same law deals specifically with false representations about land.
The maximum penalties are set out on the Office of Fair Trading's page. For a corporation it is the greatest of three amounts: $100 million, three times the benefit received, or 30 per cent of annual turnover during the period of the breach. For an individual it is $2.5 million per breach. These are ceilings set for the whole economy, not typical outcomes for a listing error, but they show the weight the law gives to accuracy in advertising.
Related readAI-drafted condition reports: what Queensland tenancy law still asksThe ACCC, which administers the consumer law nationally, states the principle in two short sentences on its advertising and promotions page: a business must be able to prove any claim it advertises, and claims should be true, accurate and based on reasonable grounds. The same page notes that the ACCC does not resolve individual disputes between consumers and businesses. In Queensland, complaints about an agent's advertising go to the Office of Fair Trading.
The Office of Fair Trading's property advertising page concentrates on price: it explains bait advertising and the meaning of an "offers over" figure. It does not address artificial intelligence or description wording specifically. The general prohibitions do that work, because they apply to any representation in any advertisement.
Who is responsible when software wrote the words
On the sources read for this guide, the answer is the agent. The law described above attaches to the person who makes the representation, and the advertisement is published in the agency's name. How the sentence was drafted does not enter into either test.
The REIQ's February 2026 article puts it without qualification.
The agent answers for the advertisement, whoever or whatever drafted it
The REIQ's article of 2 February 2026 says agents remain fully accountable for the accuracy of all representations whether or not the content is AI-generated, and that an agent can be liable for misrepresentation without knowing the tool had introduced an error. Intention is irrelevant under section 18 of the Australian Consumer Law.
Two consequences follow. The first is that "the software wrote it" explains how an error happened without changing who made the representation. The reasonable-grounds test in section 212 asks what the agent relied on, and a tool that predicts likely words is not a source of facts about a house.
The second concerns the seller. Much of what goes into a description comes from the owner: the year of the renovation, the brand of the appliances, the approval for the deck. The REIQ's 2023 article puts the burden on agents to verify statements, and recommends obtaining expert advice about a property's characteristics where needed. A seller's claim passed through a text generator arrives in the advertisement twice removed from any check, which is a reason to check it before it goes in, not after.
Related readAI suburb forecasts and growth scores: what an agent may pass onThe same February 2026 article draws a boundary that sits beside the advertising question. Using AI to produce legal advice, it says, risks breaching Queensland's Legal Profession Act, and incorrect or incomplete legal information can expose an agent to disciplinary action or civil liability. A description tool asked to explain a contract term or a tenancy right is being used outside what an agent may offer, however confident the paragraph it returns.
What the industry bodies have said
The REIQ's position, as published in February 2026, is summed up in one sentence from its article: AI should be seen as a drafting and support tool only, not a substitute for professional judgment or due diligence. The article says agents are fully responsible for verifying all content before publishing. It states that obligation firmly and leaves the method to the agency: it does not set out a checking procedure.
The institute's 2023 article supplies some of the method by another route. It asks agents to make sure every description and every word used to promote a property is accurate, to encourage buyers and tenants actively to make their own investigations, to include appropriately worded disclaimers in marketing material, and to be able to establish the grounds for what they say.
No guidance on artificial intelligence from the Real Estate Institute of Australia, the national body, was found in the research for this guide. That is a statement about what could be read, not a finding that none exists.
How an agency checks an AI draft
A checking routine does not need to be elaborate. It needs to exist, to be done by a person who knows the property, and to leave a trace. The five steps below are one way to organise what the law and the REIQ's articles require; they are a sketch of a process, not an official procedure.
- Build the fact sheet firstList rooms, features, inclusions, land size and location details from the inspection and the seller, each with its source.
- Draft from the fact sheet onlyGive the tool the checked facts and nothing vague. Ask for nothing the sheet cannot support.
- Mark every factual claimGo through the draft and underline each feature, number, distance, name and date.
- Match each claim to a sourceTick what the fact sheet supports. Delete or verify everything else, starting with schools, distances and planning.
- Sign off and keep the recordThe listing agent approves the final text. The fact sheet and the marked draft go on the file.
The third step is the one that catches machine errors, because it changes how the text is read. Read for style, an AI draft is persuasive and the eye slides over "moments from sought-after schools". Read claim by claim, the same phrase becomes three questions: which schools, how far, and is the address inside the catchment.
The fourth step has an order of priority that follows from the table earlier in this guide. Features can be checked against the inspection notes in minutes. Distances need a measured route. School catchments need the official information for the address. Statements about subdivision, a granny flat or a home business need the planning scheme and the title, and where the answer is uncertain, the wording the REIQ suggests: subject to planning consent, with encouragement to make independent enquiries.
The fifth step answers section 212(5). If the burden is on the agent to show reasonable grounds, a file that holds the fact sheet and the marked draft is how that is shown.
Agencies that use these tools widely tend to write the routine down, so that it does not depend on one careful person. A short written policy also settles questions that otherwise get decided at 5 pm on a Friday: who may publish without a second reader, whether edited copy is rechecked, and what information about clients and tenants may be typed into an outside service at all.
What a disclaimer does and does not do
Most property advertisements end with a line inviting buyers to rely on their own enquiries. The REIQ's 2023 article recommends this: marketing material should carry an appropriately worded disclaimer, and agents should actively encourage buyers and tenants to investigate independently.
The same article places the burden of verifying statements on the agent, and the ACCC asks that a business be able to prove any claim it advertises. Read together, the sources treat the disclaimer as a companion to accurate copy, not a replacement for it. A line of small print does not turn an invented school into a reasonable representation, since the test in section 212(3) is the belief the advertisement tends to create.
The same reasoning applies to a label saying that a description was written with AI. None of the sources read for this guide treats such a label as changing who is responsible, and the REIQ's February 2026 article says in terms that accountability stays with the agent regardless of how the content was generated.
What buyers and sellers can take from it
For a buyer, the practical point is that a description is an advertisement, and its specific claims are the ones worth testing. Distances can be walked or driven. A school claim can be checked against the official catchment for the address. Statements about what could be built can be put to the council or a town planner before a contract is signed. The agent can be asked where a particular statement came from; under the reasonable-grounds rule described above, there should be an answer.
For a seller, the description is published on their behalf and built largely from what they tell the agent. Reading the draft before it goes live is the simplest safeguard available, because the owner is the person most likely to notice a feature the house does not have. An owner who is asked to confirm a fact sheet is seeing an agency that checks its copy.
For agents, the sources agree on a division of labour. The software can produce a first draft quickly and well. The facts in it, and the answer for them, remain with the person whose name is on the listing.
Kooky, from Shaka
Kooky edits Queensland Estate and builds Shaka, the payment router he made for Queensland property professionals. One payment comes in, and every agent, agency and party in the deal receives their signed share on closing date.
