Using AI for listing marketing without violating Fair Housing

AI writes listing descriptions quickly and they usually read well. That combination is exactly what makes this worth thinking about carefully, because the failure here is not a bad description. It is a good description that contains language the law does not allow, published by an agent who read it, liked it, and did not notice.

This is a practice post, not legal advice. Fair Housing obligations are federal, and many states and localities add protected characteristics beyond the federal list. Your broker and your counsel are the authority for your market.

The rule, briefly

Federal fair housing law prohibits discrimination in housing on the basis of protected characteristics, and — this is the part that matters for marketing — it prohibits making, printing or publishing any statement with respect to the sale or rental of a dwelling that indicates a preference, limitation or discrimination based on those characteristics.

Three things follow from how that is written, and all three surprise people.

Intent is not the test. You do not have to mean it. The question is what the statement indicates to an ordinary reader.

It applies to the description of the property, not just to your conduct. The copy itself is the violation.

Many jurisdictions protect more characteristics than the federal list. Copy that is clean in one market may not be in another, which is a specific hazard for any tool trained on listings from everywhere.

Why AI drifts toward this language specifically

This is not a random failure. There are three mechanisms, and understanding them tells you where to look.

It was trained on decades of listing copy, including the bad decades

These models learned from text that includes an enormous volume of real estate marketing written before enforcement tightened and before the industry's current training existed. Phrases that were once unremarkable in listing copy are in that corpus in quantity, and the model has no way to know which era a phrase belongs to. It only knows the phrase is common in this kind of writing.

It is built to describe who something is for

Ask any writing assistant to make a description more appealing and it will reach for the reader. That is ordinary copywriting advice and it works in every other industry. In housing, describing who a property is for is precisely the move that is prohibited — because a statement about the ideal occupant is a statement of preference about occupants.

This is the deepest version of the problem: the model is not malfunctioning. It is doing good copywriting in a domain where that particular technique is not allowed.

It will do what you ask

Prompts drift. An agent who asks for copy that will "appeal to the right sort of buyer for this street" will get exactly that, and the result may be a compliance problem written to order. The prompt is part of the output, and nobody reviews prompts.

Where the violations actually show up

Four categories cover nearly all of it.

Describing the occupant rather than the dwelling. Any phrasing that says who would suit the home, who would love it, who it is right for, or who it is not suitable for. The property has features; people have characteristics. Describe the first.

Neighbourhood characterisation. This is the most dangerous category and the one AI produces most enthusiastically, because it sounds like helpful local knowledge. Describing the people of an area — as a community, a demographic, a character, a type — indicates a preference about who belongs there, whatever the intent. Amenities and facts are fine. Population is not.

Proximity claims that function as proxies. Naming a specific institution that stands in for a protected characteristic carries the same meaning as naming the characteristic. Factual distance to publicly relevant amenities is generally acceptable; framing a location by the group associated with a nearby institution is not.

Accessibility described as a limitation on people. Describe the physical facts — the stairs, the level entry, the doorway widths. Do not describe who can or cannot manage them.

A useful test for any sentence: does it describe the building, or does it describe a person? If it describes a person, rewrite it. This one question catches the large majority of problems and requires no legal knowledge at all.

A review process that actually works

The reason this is worth writing about is that the obvious safeguard — "have a human read it" — does not work on its own. A human reads AI copy in editing mode, checking whether it is good, and compliance problems in fluent prose do not look like errors. You need a pass with a different question.

1. Generate with a constrained prompt. Ask for the physical features, the dimensions, the materials, the updates and the factual location. Do not ask for appeal, for a target buyer, or for a description of the area's character. Most of the risk is removed here, before any review happens.

2. Do a dedicated compliance pass, separate from the editing pass. Different sitting, different question. Not "is this good" but "does any sentence describe a person, a group, or who this home suits." Read it once with only that question in mind. It takes under a minute per listing.

3. Check the adjectives specifically. Compliance problems in listing copy are overwhelmingly carried by adjectives and by phrases about suitability. Nouns describing the property are rarely the issue.

4. Have a second person review anything unusual. New agent, unusual property, unusual phrasing — someone else reads it. Your broker has seen more of this than you have, and the whole point of a brokerage is that this expertise is shared.

5. Keep the published version. If a question is ever raised, you will want the record of what was actually published and when.

What your brokerage should put in place

A written standard. One page: what may be described, what may not, and the one-question test above. Give it to every agent, including experienced ones, because the AI era changed the volume of copy an office produces and volume is what turns a rare mistake into a regular one.

A rule that AI output is a draft. Never published as generated. This should be policy, not preference, and it should be stated the day you introduce any tool that writes.

Training on the mechanism, not just the rule. Agents who know why these tools produce this language will spot it. Agents who only memorised a list of banned phrases will miss the novel phrasing that means the same thing — and novel phrasing is what a language model is best at producing.

A clear escalation. An agent who is unsure should have somewhere to ask that does not feel like an admission. The alternative is that they guess, and they will guess in the direction of the copy that sounds better.

The honest summary

AI is genuinely useful for listing marketing. It removes the blank page, it is fast, and it is good at describing physical things accurately when you ask it to.

It is also, structurally, inclined to produce the one kind of sentence this industry is not allowed to write — not occasionally, and not because of a flaw someone will patch, but because describing the audience is what good copywriting does everywhere else.

That is a manageable risk. It is manageable through process: a constrained prompt, a separate compliance pass, and a written standard everyone has read. It is not manageable through care alone, because the copy reads well, and that is exactly the problem.