If you use AI to write listing descriptions — and most agents now do — you have a compliance exposure that didn't exist three years ago, and it is almost certainly not on your broker's checklist yet.
The models are good at real estate copy. That's the problem. They're trained on decades of listing descriptions, and decades of listing descriptions contain phrasing that has been illegal since 1968.
Try this now
Ask any AI assistant to write a description for a three-bedroom house near a good elementary school. Then read what comes back for these:
- "perfect for families" / "great for families"
- "family-friendly neighborhood"
- "safe neighborhood"
- "good schools" / "great school district"
- "quiet neighborhood, ideal for young professionals"
- "an up-and-coming area"
- "walking distance to churches"
- "master bedroom"
Most of the time you'll get at least two without asking for them.
Every one of those is a Fair Housing problem.
Why these specific phrases
The Fair Housing Act prohibits statements that indicate a preference or limitation based on race, color, religion, sex, familial status, national origin, disability, or (in California, under FEHA) several additional categories including source of income and sexual orientation.
The phrasing rarely looks like discrimination. That's what makes it dangerous:
"Perfect for families" — familial status is a protected class. Stating a preferred household type indicates a limitation, even framed as a compliment. "Great for empty nesters" is the same violation pointed the other way.
"Safe neighborhood" / "good schools" — long treated by HUD and fair housing groups as proxies that correlate with racial composition. You are also making a claim about crime or school quality that you cannot substantiate and are not qualified to make.
"Up-and-coming" — conventionally read as signalling demographic change. It describes the residents, not the house.
"Walking distance to St. Mary's" — religion. Naming a nearby place of worship as an amenity indicates a religious preference. A grocery store is fine.
"Master bedroom" — not a protected-class issue, but NAR and most MLSs moved to "primary bedroom" years ago and some MLSs now reject the term outright.
The rule underneath all of it is simple: describe the property, not the buyer.
Why AI makes this worse, not better
Three reasons.
It writes what it was trained on. Its training data is full of listing copy from decades when this language was normal. It reproduces the pattern because the pattern is what it learned.
It's fluent, so it reads as safe. A junior agent writing "perfect for families" sounds like a junior agent. An AI writing it sounds like polished professional copy. Fluency suppresses the instinct to check.
It's fast, so nobody reads carefully. The whole reason you're using it is to stop writing descriptions by hand. Volume goes up, scrutiny goes down. That's the actual mechanism by which this reaches a live listing.
Who is liable
You are. The agent whose license is on the listing.
"The AI wrote it" is not a defense, and no vendor's terms of service transfers this to them — go and read yours, the indemnity runs the other direction. DRE and HUD both treat published copy as the licensee's statement regardless of the tool used to produce it.
What to actually do
Screen it in software, not in your head. A checklist works until the day you're tired. A keyword filter that runs on every description before it publishes works every day. Build it into whatever tool you write in.
Ban the phrase list explicitly in your prompt — and then still screen the output, because prompts are advisory and models drift.
Rewrite toward the property. Not "perfect for families" but "four bedrooms, fenced yard, 0.3 miles to Roosevelt Elementary." The second is more useful to the buyer and it's compliant. Facts do more work than adjectives.
Never characterise the people. Not the current residents, not the neighbors, not the buyer you imagine. The house, the lot, the measurable facts about location.
How we handle it
We built the screen into the publishing path rather than the prompt. Any AI-written page on our platform is scanned before it can go live, and a hit blocks publication rather than quietly rewriting it — because a silent fix hides the fact that the generator is producing non-compliant copy, and you'd want to know that.
We'd encourage you to do the equivalent wherever you write, with whatever tool you use. The specific implementation matters far less than the principle: the check has to be a gate, not a reminder. A rule you have to remember is a rule you will eventually forget, on a Friday, on a listing that goes live over the weekend.