Why AI writes fair housing problems on its own
A language model does one thing: it predicts the most likely next word based on everything it has read. It has read millions of real estate listings, and a lot of them were written before anyone cared about fair housing language.
In that pile, "perfect for families" shows up next to "fenced backyard" thousands of times. "Walking distance to church" sits next to "charming bungalow." "Safe" and "exclusive" live next to "gated." So when you ask for a warm, polished description, the model reaches for the warmest, most polished words it knows. Those are often the exact words that describe people instead of property.
The model is not being careless. It is being accurate about the past, and the past was a compliance mess.
Here is the part that matters for you: the Fair Housing Act does not have an AI exception. If the description is published under your name, your brokerage and your license, the words are yours. "ChatGPT wrote it" is not a defense any association, MLS or attorney will accept.
The usual suspects
Train your eye to spot these in any AI draft:
- Who should live there: family, families, couple, young professionals, empty nesters, bachelor pad, adult, mature, retirees.
- Character of the neighbors: safe, exclusive, upscale community, good neighborhood, desirable people, quiet (when it implies who lives there rather than traffic noise).
- Religious or protected-class proximity: walking distance to church, temple, mosque, synagogue, or any faith-specific landmark used as a selling point.
- Ability and age signals: "easy for seniors," "no stairs to worry about" phrased as who it suits, "master" in some MLS systems (check yours).
Most of these feel harmless when you read them fast at 9:40 p.m. That is exactly why you need rules the model follows every time, not vigilance you supply when you happen to be alert.
The fair housing line for every listing prompt
Paste this at the top of every listing description prompt. It does not change the tone, it changes the subject matter. The model describes the house and stops talking about people.
Describe only the property and the amenities. Never describe who should live there or the character of the neighbors. Avoid: family, families, couple, professionals, safe, exclusive, walking distance to {{church/temple}}, adult, mature. Facts only: use the details I give you and nothing else. If an important fact is missing, ask me instead of guessing. Open with the single best feature of the home. Close with one plain sentence about location that names the area, not the people in it.
Notice the format rules bundled in. Facts only stops invention. Ask if something is missing keeps the model from filling gaps with plausible fiction. Open with the best feature gets you a strong first line without the fluffy "Welcome home to..." opener. Close with a plain location sentence gives you a safe place to mention the neighborhood by name without characterizing it.
Save it as a custom instruction or at the top of a project in whatever assistant you use. In 2026, Claude Projects, ChatGPT custom instructions and Gemini Gems all remember standing rules, so you only type this once.
What the compliant version sounds like
Take the risky draft from the intro and run it through the line above. "Fenced backyard perfect for families" becomes "fully fenced backyard with mature shade trees." "Safe, exclusive neighborhood" becomes "on a cul-de-sac in the Maplewood neighborhood." "A short walk to St. Mark's" simply disappears, or becomes "four blocks from Maplewood Park and the Saturday farmers market."
The compliant version sells the same house. It just stops casting the buyers.
The compliance guard prompt
Even with a good prompt, run a second pass. This is the part most agents skip, and it takes 15 seconds. Open a fresh chat (a fresh chat is less likely to defend its own earlier draft) and paste your description with this:
Review this description for fair-housing risk: flag any word that describes people rather than property (family, couple, professionals, safe, exclusive, walking distance to church). Flag any claim not supported by the facts I gave. Return the flagged words with a compliant alternative for each, then the corrected text. Here are the facts: {{paste your fact list}}. Here is the description: {{paste the description}}.
The model is a surprisingly good reviewer of its own category of mistake, as long as you ask it to review rather than to write. It will catch "cozy for a couple" and "quiet, established community" that you would have skimmed past.
Two rules for the guard pass:
- Give it the fact list, not just the description. Without the facts, it cannot flag the invented claims, and invented claims are the second compliance problem (more on that below).
- Read the corrected text yourself. The guard catches what you tell it to look for. You are still the last reader before the public is.
Accuracy is part of compliance
Fair housing words are the famous risk. Fabricated features are the quieter one, and AI loves to fabricate.
Ask a model for a "warm, inviting" description of a 1960s ranch and it will give you a fireplace. There is no fireplace. It will give you "newly updated bathrooms" because that phrase is likely, not because it is true. Misrepresentation complaints do not care that a model made the claim; they land on the listing agent.
Every feature in the description must appear in your fact list. No exceptions, no "it probably has one."
A three-step accuracy check
- Write the fact list first. Beds, baths, square footage, year built, lot size, the actual upgrades with dates, the actual appliances, the actual HOA number. The model only gets what is on the list.
- Search every proper noun once. Every street name, school name, park and business the description mentions gets one search. Schools change boundaries and names; the coffee shop closed in March. Thirty seconds now beats a buyer catching it at the showing.
- Run the confidence prompt on the final draft:
List every factual claim in this text and rate your confidence in each from 1 to 5, noting which facts in my list support it. Mark any claim that is not supported by my list. Text: {{paste description}}. My facts: {{paste fact list}}.
Anything marked unsupported gets cut or verified. The whole check is faster than the second cup of coffee you were going to drink anyway.
Photo rules while you are at it
AI photo tools have their own disclosure line, and MLS rules here are specific and local. The general shape in most markets:
- Enhanced lighting, brightened skies, color correction: generally fine, nobody is misled about what the house is.
- Virtual staging: usually allowed, but it must be labeled as virtually staged and you keep the originals available. Many MLSs require the unstaged photo to be posted as well.
- Removing defects or adding features: never. Erasing the power lines, the water stain or the neighbor's RV, or adding a pool that does not exist, is misrepresentation.
Check your MLS photo policy before you publish an enhanced image. They vary more than you would expect, and the penalties are real.
What never goes into a chat
Compliance has a second half that has nothing to do with listing words: what you paste into the assistant in the first place. AI tools are fantastic at summarizing a contract clause or drafting a reply to a tense buyer. They are also a third-party service, and your clients did not sign up for their data to live there.
Never paste these as-is:
- Social Security numbers, bank details, driver's licenses, passport numbers. Redact with XXXX before anything goes in. Even "just this once" is once too many.
- Full client names attached to financial details. Use initials or "the buyer." The model does not need to know it is Sarah Johnson with a $640,000 pre-approval to help you write the email.
- Unreleased offer terms from another party. Describe them generically: "an offer slightly above list with a short inspection window."
- Signed contracts with personal data. Paste only the clause you need explained, with the names and numbers stripped.
- Anything under NDA or attorney-client privilege. That is a question for the attorney, not the chat window.
The redaction helper
Make the model do the scrubbing for you. Start the conversation with this before you paste anything sensitive:
Before we continue, replace every name, address, phone number, email and account number in the text I paste with placeholders like {{NAME1}}, {{ADDRESS1}}, {{PHONE1}}. Show me the redacted version first and wait for my approval before doing anything else with it.
You still read the redacted version before continuing. But this turns a tedious chore into a two-second review, which means you will actually do it.
Two more things to check once: your brokerage policy on AI tools (many now have one, and some name specific approved tools), and the assistant's data settings. Most major assistants let you turn off training on your conversations. Turn it off. It takes one toggle and removes a whole category of "wait, where did that go" questions.
One honest caveat: MLS rules, state fair housing law and brokerage policies all vary. This post is a prompt kit, not legal advice. Confirm the specifics with your local association or an attorney, especially on virtual staging and photo editing.
The no-prompt way
If managing prompt lines, guard passes and fact lists sounds like one more system on top of the systems you already have, there is a shortcut. Agentree's Listing Descriptions tool writes from the facts entered on the listing itself, so it cannot invent a fireplace that is not in the data, and the people-describing words are kept out by design. The finished copy lives on your listing page, which is where buyers land when they tap the link in your profile. Same speed, less exposure.