Playbook
ChatGPT listing descriptions without a fair housing problem
The short answer
Use ChatGPT for listing descriptions, but edit every one. HUD's 2024 guidance applies the Fair Housing Act to AI-generated marketing copy exactly as it does to human writing, and the licensed agent carries the liability. The risk is describing who lives somewhere rather than what is there.
The rule, stated plainly
Writing listing descriptions is the single biggest use of AI in real estate: 68.5% of Realtors who use AI use it for exactly this. So the compliance question is not niche, it applies to most of the profession.
HUD's 2024 guidance is unambiguous. AI-generated listing descriptions and marketing copy are held to the same Fair Housing Act standard as anything written by a person, and the licensed agent is liable. The AI wrote it is not a defense, and neither is the vendor's terms of service.
This is general guidance rather than legal advice, and your brokerage almost certainly has its own policy that is stricter.
What actually goes wrong
The failure mode is not obvious slurs. It is a model reaching for warm, aspirational language and accidentally describing the people rather than the property.
Trained on a corpus of existing listings, it will happily produce phrases that have been generating complaints for decades: describing a neighbourhood as safe, a home as perfect for a young family, a location as walking distance to churches. Each of those describes who should live there.
The tell is simple. If a sentence would still be true if the property were empty and nobody had moved in, it is describing the property. If it needs a hypothetical occupant to make sense, look harder.
If a sentence needs a hypothetical occupant to make sense, look harder.
The phrases to watch
Not an exhaustive list, and not a substitute for your brokerage's own. These are simply the ones an assistant produces most readily because they are common in the training data.
Common triggers
- Family, family-friendly, perfect for a growing family: describes occupants, not rooms
- Safe neighbourhood, good area, quiet street: a judgement about who is around
- Walking distance to churches, synagogues or any place of worship: names a religion
- Master bedroom, in some markets and brokerages, though this varies
- Ideal for professionals, retirees, students: all describe occupants
- Exclusive, private, integrated: carry history in a housing context
- Any reference to schools by quality rather than by name and distance
A prompt pattern that behaves
The reliable fix is to constrain the model to physical facts before it starts, rather than editing warmth out afterwards.
Something like: Write a listing description for a property with these features. Describe only the physical property and its measurable location. Do not describe who might live here, do not characterise the neighbourhood or its residents, do not mention schools, places of worship, safety or family suitability. Plain, specific, under 150 words. Features: and then your list.
That produces flatter copy than the default, which is the point. Flat and compliant beats warm and complained about, and the specificity of real features does more selling than adjectives anyway.
The edit checklist
Thirty seconds per listing, and it catches almost everything.
Before you publish
- Does any sentence describe a person or a type of person
- Is any place of worship, school quality or demographic mentioned
- Does it use safe, quiet or good about the area rather than about a measurable fact
- Is every factual claim about the property actually true, including square footage and year built
- If photos are virtually staged, is that disclosed on the image itself
- Would you be comfortable reading this sentence aloud in a complaint hearing
The wider point about accuracy
Fair housing is the compliance risk. Plain invention is the practical one. 63% of Realtors name accuracy of outputs as their top concern with AI, and listing copy is where it bites: a model asked to make a description appealing will add a fireplace.
The habit that solves both is the same. Give it only facts you have verified, constrain what it may talk about, and read every word before it publishes. That takes a minute and it is the entire difference between a useful tool and a liability.
Common questions
- Can I use ChatGPT to write listing descriptions?
- Yes, and most agents do: 68.5% of Realtors using AI use it for this. The requirement is that you edit every one, because HUD applies the Fair Housing Act to AI-written copy exactly as it does to yours.
- Is the AI wrote it a defense?
- No. HUD's 2024 guidance is explicit that AI-generated marketing is held to the same standard, and the licensed agent carries the liability. The vendor's terms do not transfer it.
- What words should I avoid in a listing description?
- Anything describing occupants rather than the property: family-friendly, ideal for professionals, safe neighbourhood, walking distance to churches. If a phrase needs a hypothetical occupant to make sense, rewrite it.
- How do I stop it writing that way in the first place?
- Constrain the prompt before it starts. Tell it to describe only the physical property and measurable location, and to avoid neighbourhood character, schools, places of worship, safety and family suitability. The copy comes out flatter and safer.
- What else does AI get wrong on listings?
- It invents features. 63% of Realtors name output accuracy as their top AI concern, and a model asked to make copy appealing will add a fireplace that is not there. Give it verified facts only, and read before publishing.