When AI Invents a Feature That Is Not There: Listing Copy and the Verification Step

AI writes confident copy. That confidence is the problem.

When you ask a language model to write a listing description, it does not look at the property. It works from whatever you give it in the prompt, and it fills the gaps with what sounds plausible. The result reads like normal marketing text. The invented detail sits next to the real detail, formatted the same way, at the same reading level, with no signal that anything is wrong.

This is not a glitch. It is how these models work. They predict the next plausible word or phrase. For a three-bedroom property near a school, the model knows that "sought-after school zone" is the kind of phrase that appears in listing copy. It writes it. Whether the property is actually in that zone is a separate question the model cannot answer.

Your buyer does not know it is a separate question. They read the description, trust the zone claim, choose the property in part because of it, and then discover the error after the sale is progressing.

This article covers why the risk is higher with AI copy than with human copy, which categories of detail are most often invented, why the errors survive a casual read, and how to run a verification pass in under two minutes per listing.

Why AI Fills Gaps Rather Than Leaving Them

A human copywriter working from a sparse brief will sometimes write around the gaps. They hedge. They leave a line blank. They call the agent.

A language model does not call anyone. It is trained to produce complete, fluent output. Gaps in the prompt are not treated as missing information. They are treated as space to fill with plausible content.

This behaviour is well documented and has a name in the research literature: hallucination. But in listing copy it does not look like a hallucination. It looks like a marketing claim. There is no warning label, no flagging, no asterisk.

The density of the problem scales with the sparseness of your prompt. If you give the model a full fact sheet, with confirmed measurements, zoning details, and consent records, the model has less to invent. If you give it a brief, a photo description, and a street address, it has a great deal to invent and will do so smoothly.

Most listing prompts are closer to the second scenario.

The Five Categories Most Often Invented

Across listing copy, five categories show up repeatedly as sources of generated errors.

School zoning. Zone boundaries are precise and change periodically. A property can be a short walk from a school and still be outside the zone. The model does not know the boundary. It knows that proximity to a school and zone references co-occur in listing copy, so it writes both.

Land area and room dimensions. Measurements are specific. Without a survey or floor plan in the prompt, the model estimates. It picks numbers that fall within a plausible range for the property type, which means the invented figure is close enough to survive a skim read and wrong enough to matter in a negotiation or a building inspection.

Consents and permits. Renovation history is a common listing feature. The model describes a renovated kitchen or a new deck as if these are established facts. Whether those works had the required consents is not something the model considers, because it does not consider anything. It writes what fits. A buyer relying on the implied consent status of past work may later face costs you did not intend to represent.

Renovation history. Related to consents but distinct. The model may describe a renovation as recent, extensive, or high-specification based on nothing more than a photo description of a modern-looking interior. If the works were done years ago, or were cosmetic rather than structural, the description misrepresents the asset.

Distances and travel times. "Ten minutes to the CBD" or "walking distance to the beach" are stock phrases in listing copy. The model uses them because they fit the pattern. Actual travel time depends on traffic, route, and mode. The beach may be 900 metres on foot but accessible only via a road with no footpath. The model does not know. It writes the phrase that fits.

Why the Errors Survive a Casual Read

Invented claims in AI copy do not look different from real claims. They use the same structure, the same vocabulary, and the same confident tone as every other line in the description.

A human reader, including an experienced agent reading their own copy back, is not primed to be suspicious. The text reads fluently. The brain reads for sense and skips over the question of verification. This is a well-understood feature of how people process familiar text formats.

Listing descriptions are a familiar format. Agents read hundreds of them. The familiarity makes them easier to approve uncritically.

There is also a timing problem. AI copy is often produced quickly, approved quickly, and uploaded quickly. The speed is the point. The verification step, if there is one, is informal. If it does not surface the invented claims, they go live.

The Two-Minute Verification Pass

The fix is not to stop using AI. The fix is to treat AI output as a first draft that requires a structured check before it becomes a listing.

Build one standard into your workflow: no AI copy is published without a factual pass against source documents.

Here is a pass that takes under two minutes for a typical listing.

Step one. Read the draft with a marker or cursor. Underline every claim that states or implies a specific, verifiable fact: measurements, zones, times, distances, consent history, renovation scope, and year.

Step two. Match each underlined claim against a source document. Acceptable sources are the certificate of title, the council property file, the floor plan or survey, the listing authority or vendor disclosure, or a direct written confirmation from the vendor. If there is no source document for the claim, delete the claim or mark it for vendor confirmation before publishing.

Step three. Check travel and distance claims separately. Use a mapping tool at the time of listing, not from memory, not from a prior visit. If the time or distance varies significantly by mode or time of day, either remove the claim or specify the condition: "approximately 15 minutes by car at off-peak times."

This pass does not require legal training. It requires only that someone asks, for each specific claim, whether there is a document that supports it.

If you are building this into an agency workflow, a structured AI usage policy is the clearest way to make the step mandatory rather than optional. An example of what that looks like for a real estate context is at getplainstart.com/ai-policy-for-real-estate-agencies.

Worked Example

Here is a generated paragraph, typical of what a language model produces from a sparse prompt about a three-bedroom home near the coast.

Generated version:

Situated in the coveted Maplewood school zone, this beautifully renovated three-bedroom home sits on an impressive 650 square metres with an open-plan living area that captures sea views. The recently updated kitchen features high-end appliances and flows to a large north-facing deck, consented and completed in 2021. A short five-minute walk brings you to the beach, making this a genuinely lifestyle-driven address.

Three claims in this paragraph cannot be verified from a sparse prompt.

First, "Maplewood school zone." Zone boundaries are set by the relevant authority and change. This claim requires a current zone check against the property address, not an assumption from proximity.

Second, "650 square metres." Without a survey, title, or council record confirming this figure, it is an estimate. Estimated land area in listing copy is a known source of disputes.

Third, "consented and completed in 2021." The model wrote this because deck and consent claims co-occur in listing copy. Whether the works were consented, and when, is a matter of council record. The year 2021 is invented.

Corrected version, after verification pass:

This three-bedroom home is within 800 metres of Maplewood Primary. Buyers should confirm current zone boundaries directly with the school or the relevant authority, as these change. The property sits on 618 square metres, per the certificate of title. The kitchen was renovated prior to the current vendor's ownership. The rear deck was added in 2021. Buyers should confirm consent status with the council. Walking distance to the beach is approximately 950 metres on foot via Coastal Road.

The corrected version is less glossy. It is also defensible. Every claim is sourced. The uncertainty about the zone and the consent is flagged rather than papered over with confident language.

The first version would read well on a portal. The second version is less likely to become the basis of a dispute.

The Underlying Point

AI listing copy is not unreliable because the technology is poor. It is unreliable in a specific way: it produces confident, well-formatted text that includes claims it cannot possibly know to be true.

The two-minute verification pass addresses that specific failure. It does not slow your workflow in any meaningful way. It does create a clear point in the process where a human confirms what the machine cannot.

That is the step. Build it in, make it standard, and the risk drops significantly.

This article is general guidance for workflow and process. It is not professional, legal, or compliance advice. What rules apply to your listings, your agency, and your jurisdiction is a matter for your own adviser.