A broker in Windsor pasted property notes into a general purpose chatbot and asked for listing copy. What came back described a sun-filled breakfast nook in a unit whose only kitchen window faces a parking structure, and a newly renovated primary bath that was last renovated in 2011. The listing ran for 6 days before a buyer’s agent flagged it. Fixing it cost most of a day, a seller who now checks everything, and one showing that never got rebooked.
That is not an argument against the tool. The same broker still drafts listing copy in about 4 minutes instead of 20. It is an argument for knowing where the money goes when this fails, because it only goes to three places.
1. Copy that describes a property you don’t have
Language models fill gaps. Give one 6 facts about a house and ask for 200 words of warm description, and you will get 200 words. The difference between the 6 facts and the 200 words is invented, and it reads well, which is the whole problem. Nobody proofreads a sentence that sounds right.
The cost is rarely a lawsuit. It’s a correction, a credibility hit with a seller, and occasionally a buyer who drove down from Cheyenne because of a nook.
2. Language a fair housing complaint can be built on
Ask for copy that sounds welcoming and you will get words about who the neighborhood is for. Perfect for a young family. Quiet building, ideal for retirees. Walking distance to the Catholic church. A model produces these because they appear constantly in the writing it learned from.
The Fair Housing Act has prohibited statements like these in housing advertising for decades, and it does not care that a machine wrote the sentence. This is the expensive one, and the exposure sits with the licensee and the brokerage, not the software vendor.
3. Client information typed into a tool that keeps it
Consumer chatbots on free plans commonly use what you type to improve the product unless you turn that off or move to a business account. Agents paste in seller financials, pre-approval letters, and whole contracts to get a summary out.
The cost here is not a headline. It’s a broker of record finding out during a review that confidential client material has been sitting in someone’s personal account under default settings, with no way to get it back.
What to do about it this week
Three habits, none of which require hiring anyone.
Read for facts, not for flow. Before an AI-drafted listing goes live, check every specific claim against the MLS sheet or the seller disclosure. Square footage, year built, appliances, HOA, schools. If your source doesn’t say it, delete it.
Describe the property, never the buyer. Add one line to your prompt: describe only the physical property and its features, and never who might live there. Then read the output for it anyway, because that instruction does not always hold.
Check the account, not the tool. Open the settings on whatever you’re using and find the data controls. If you can’t tell whether your inputs train the model, assume they do, and stop pasting anything you wouldn’t want read aloud at a hearing.
The first two habits cost about 90 seconds per listing. The third costs 10 minutes, once. That is the entire price of not being in one of these three stories.