AI for Real Estate Listings: Fair Housing Guide

Forty-six percent of REALTORS report using AI-generated content, with listing descriptions the headline example (NAR's 2025 Technology Survey). Several MLSs now ship a description generator inside the listing-entry screen. The tools work — and left alone they will hand you copy that is a federal advertising violation, because the sentence that tests best with a seller is very often the sentence that names the buyer.

This guide is about that failure, and it does what no vendor page does: it tells you which words HUD has actually ruled on, which it has not, and which are folklore.

The Challenge

The advertisement is itself the violation. Section 804(c) of the Fair Housing Act — 42 U.S.C. § 3604(c) — makes it unlawful to "make, print, or publish… any notice, statement, or advertisement, with respect to the sale or rental of a dwelling that indicates any preference, limitation, or discrimination" based on race, color, religion, sex, handicap, familial status, or national origin. You do not have to refuse anyone. You do not have to intend anything. The text on the screen is the act.

Four properties of that rule make AI-drafted copy unusually dangerous:

• Intent is irrelevant. The test is whether an ordinary reader would take the ad to suggest a preference. "The model wrote it" is not a defense.
• Liability runs to whoever publishes. HUD's guidance is explicit that publishers and advertisers alike are responsible for an ad that violates the Act on its face. Paste model output into the MLS and you are the advertiser.
• It reaches exempt properties. The "Mrs. Murphy" exemption at 42 U.S.C. § 3603(b) can exempt a dwelling from parts of the Act — but not from § 3604(c).
• The base rate is not zero. The National Fair Housing Alliance counted 32,321 housing discrimination complaints in 2024, among the highest totals in two decades. Disability drove 54.6% of them, race 15.6%, sex 7.1%, national origin 5.7%, familial status 5.5%.

The model is not being careless — it is doing what it was trained to do. Asked for a listing description, it reproduces what listing descriptions look like and what a human rater would call warm. The most common move in that corpus is to picture the occupant: imagine your family gathered around the fireplace, perfect for a young professional, a wonderful place to raise children. That sentence tests beautifully. It is also, verbatim, a statement of preference based on familial status.

The violation is not a bug. It is the objective function — which is why the prompt matters more here than the tool.

How AI Solves It

One rule turns a general-purpose model into a compliant listing writer, and it is short enough to memorize:

Describe the property. Never describe the buyer.

Every FHA advertising problem in listing copy breaks that sentence, and nearly every compliant rewrite applies it. "Perfect for a growing family" describes a buyer. "Four bedrooms, fenced yard, second-floor laundry" describes a property — and sells the same house to the same person without stating who that person should be.

Under that constraint the model is good at four jobs:

• Turning a fact sheet into prose. You supply verified facts; it supplies sentences. This is the only safe direction, because facts you did not supply are facts it invented.
• Producing length variants. Public remarks, a portal blurb, a social caption and a neighborhood page are four word counts from one set of facts.
• Reading photographs — with a large caveat below.
• Auditing copy against § 3604(c). The strongest use here and the most neglected. Reviewing is a classification task; writing persuasive copy is the objective that produces the violation. The model is a better reviewer than writer.

What it cannot do is know your rules — your MLS's policy, its character limit, your state's protected classes.

Recommended Tools & Models

Any frontier model writes competent copy; what differs is constraint-following and cost at volume. Claude Sonnet 5 and GPT-5.5 hold a long rubric without drifting. Claude Haiku 4.5 is cheap enough to audit every listing a brokerage publishes. Gemini 3.5 has the context window for a whole photo set.

The compliance surface is three layers

Agents get this wrong in both directions — panicking about words that are fine while shipping words that are not. Three rulebooks apply, and they disagree:

1. Federal law. The FHA's seven protected classes: race, color, religion, sex, national origin, disability, familial status. This is the layer that produces lawsuits.
2. State and local law. Almost always longer — many jurisdictions independently protect age, marital status, source of income, sexual orientation, gender identity or military status. (HUD's 2021 memo reading federal "sex" to cover sexual orientation and gender identity was withdrawn in 2025; state and local protections are untouched by that.) Look up your own state's list.
3. Your MLS's rules. The strictest layer, the one most likely to cost you money, and the one no AI tool knows. MLSs monitor public remarks for fair-housing language and fine violations — commonly a few hundred dollars apiece. They also cap length, typically between 1,000 and 2,000 characters, and ban contact details, URLs, showing instructions, lockbox codes and open-house dates from the field.

Copy can be lawful under the FHA and still get cited by your MLS. Plan for the union.

What HUD actually says — and what it does not

Most "forbidden words" lists circulating in real estate are not HUD's, and several entries contradict HUD outright. The paper trail:

• 24 CFR § 100.75 is the operative regulation. It states the prohibition; it lists no words.

• 24 CFR Part 109, the old Fair Housing Advertising Guidelines, did list words — and was removed from the CFR effective May 1, 1996 as part of HUD's regulatory-streamlining initiative (61 FR 14378). HUD still applies it as guidance.

• HUD's memorandum of January 9, 1995 (Roberta Achtenberg, then Assistant Secretary for Fair Housing and Equal Opportunity) supersedes parts of Part 109 and remains the most recent HUD guidance on advertising language. It exists because the guidelines "have been interpreted (usually by persons outside of HUD) to extend the liability for advertisements to circumstances which are unreasonable."

  Read it for what it is. The memo tells HUD's own intake staff when not to file a complaint. It is not a safe harbor: it does not bind a private plaintiff suing directly under § 3604(c), and it does not bind a state agency enforcing a stricter state statute. And HUD has been withdrawing FHEO guidance since 2025, so do not assume permanence. It is the best evidence of what HUD thinks, which is not the same as immunity.

Part 109's list flags these families:

The bolded entries are the ones an AI actually produces. It will never write "no Irish." It will absolutely write exclusive enclave, great for active adults, a traditional neighborhood, steps from St. Mary's — all on HUD's own list.

The words HUD says are fine (which the industry bans anyway)

The 1995 memorandum, in terms:

• Race, color, national origin: "complaints over use of phrases such as master bedroom, rare find, or desirable neighborhood should not be filed."
• Sex: "Use of the term master bedroom does not constitute a violation of either the sex discrimination provisions or the race discrimination provisions. Terms such as mother-in-law suite and bachelor apartment are commonly used as physical descriptions of housing units and do not violate the Act."
• Disability: "Advertisements containing descriptions of properties (great view, fourth-floor walk-up, walk-in closets), services or facilities (jogging trails), or neighborhoods (walk to bus-stop) do not violate the Act." Describing accessibility features (wheelchair ramp) is lawful.
• Familial status: "Advertisements describing the properties (two bedroom, cozy, family room), services and facilities (no bicycles allowed) or neighborhoods (quiet streets) are not facially discriminatory and do not violate the Act."

So "master bedroom" is not an FHA violation. The 2020 shift to "primary bedroom" — begun by the Houston Association of REALTORS, followed by the homebuilder groups — was a voluntary style change, and NAR said at the time that HUD had advised the term did not violate fair housing rules. Comply with your MLS if it requires "primary," but know which rule you are following: an agent who believes HUD bans "master bedroom" will also believe it bans "walk-in closet," and will spend their compliance attention in exactly the wrong place.

Where HUD has not ruled — say so plainly

• "Safe neighborhood." On no HUD list, and HUD has expressly blessed desirable neighborhood and quiet streets. But "safe" is a claim you cannot substantiate — a misrepresentation problem under NAR Code of Ethics Article 2 before it is a fair housing one — and an ordinary reader can hear it as a proxy for a neighborhood's demographics. Most brokerage policies ban it. Point buyers at published crime statistics; do not characterize.
• "Walking distance." HUD explicitly permits walk to bus-stop and fourth-floor walk-up. Many MLS word lists flag "walking distance" anyway, on the theory that it presumes an ambulatory buyer. HUD does not agree with those lists — but HUD is not the one fining you. Follow your MLS, and know it is a house rule.
• Schools. Naming the district is normal and is a field in most MLSs. Top-rated schools is a substantiation problem and, for many fair housing trainers, a familial-status and race proxy.

On AI specifically: no guidance, and no exception

HUD released two AI guidance documents on May 2, 2024 — one on tenant screening, one on the targeting and delivery of housing ads on online platforms, i.e. the algorithmic decision about who sees an ad. HUD then withdrew the advertising one effective September 17, 2025 (91 FR 17291), along with seven other FHEO guidance documents, stating they "should not be relied upon as authoritative." The tenant-screening guidance survives.

Neither ever addressed the wording of an ad written by a generative model — and the withdrawal removes guidance, not liability. There is less paper telling you what HUD thinks, and exactly as much § 3604(c).

NAR's position is that AI-generated content may be inaccurate, may create fair housing risk, and should be reviewed by a licensed professional before publication, and that Articles 2 and 12 of the Code of Ethics apply to it as to anything else you write.

And the gap most guides leave open: AI-altered photos. Virtual staging, decluttering, object removal, sky replacement — this is the most common AI use in listing marketing and it carries a disclosure obligation the description does not. NAR ties AI-enhanced listing images directly to Code of Ethics Article 2, which prohibits exaggerating, concealing or misrepresenting pertinent facts about a property. Many MLSs require virtually staged photos to be labeled as such. Two rules: label any image AI put furniture into, and never alter a defect out of a photo. Adding a sofa is styling. Removing a crack is misrepresentation.

No regulator has carved out generative copy, so ordinary § 3604(c) liability applies unchanged. The review step is not overhead. It is the product.

Step-by-Step Implementation

1. Build a fact sheet first

Every hallucination traces back to a fact the model was not given and invented to fill the gap. Collect: beds, baths and square footage with their source (tax record, appraisal, floor plan — they disagree, and which you used is a disclosure question); year built, lot size, HOA fee, taxes, parking; ages of roof, HVAC, water heater, windows; renovations with years and permits; verified specifics like countertop material and appliance brands; genuine accessibility features (a zero-step entry, a first-floor primary suite — describing these is lawful); and your MLS's character limit.

2. Draft with a constrained prompt

The constraint must be in the prompt, not in your head.

That last instruction — make it cite its own facts back to you — is the cheapest hallucination check in existence, and it costs the model ten seconds.

3. Extract from photos, carefully

Multimodal models will read a photo set and enumerate features. Useful, and the richest source of invented facts in this workflow.

Two failures that list prevents. Vision models confidently misidentify materials — quartz becomes granite, a den becomes a bedroom. A bedroom is a legal definition (egress, ceiling height, a closet in many jurisdictions), not a room with a bed in it, and calling a den a bedroom is a misrepresentation under NAR Article 2 and an MLS citation in most markets. And photos leak protected characteristics — a crucifix on a wall, a mezuzah on a doorframe, a family portrait. 24 CFR 100.75 covers photographs and illustrations, not only words.

4. Audit in a separate pass

Do not ask the model that wrote the copy whether the copy is good. Open a fresh chat — ideally a different model — and give it one job.

Run this on copy you wrote yourself, too. It catches the friendly-sounding line you never thought twice about — "stroll to Sunday farmers' market at Holy Cross" is a direction given by reference to a parish, which is the exact case 24 CFR 109.20 calls out.

5. Do the human review anyway

The audit is a filter, not a signoff — it is a language model, and it can miss things and invent rules that do not exist. Read the copy yourself against this checklist:

• Does every sentence describe the property, not the occupant?
• Is every number traceable to a document you hold?
• Are beds and baths counted as your MLS defines them?
• Any person, religious object or personal photo left in the images?
• Any catch word: exclusive, restricted, private, integrated, traditional?
• Any worship-adjacent landmark used as a direction?
• Any claim about safety, crime or school quality?
• Does it fit the character limit, with no contact info, URLs or showing instructions?
• Would a member of any protected class read this and conclude the home was not meant for them?

6. Local SEO: on your own site, not in the MLS

Public remarks do not rank — they are syndicated boilerplate duplicated across every portal. Search traffic comes from pages you control: a property landing page, a neighborhood guide, a "recently sold in X" post. Use a model to expand one fact sheet into those longer pages, with local specifics it cannot invent (the street, the school attendance zone, the transit line, the year the building went up).

The compliance trap moves with you. Neighborhood pages are where steering language creeps in, because the natural way to write one is to describe the people who live there. Run the same audit prompt over every neighborhood page. And check facts: a model asked for local landmarks will cheerfully invent a park.

7. Scale it — but only after the prompt is right

For volume: put fact sheets in a spreadsheet, run the drafting prompt then the audit prompt over each row with n8n or a script against an API, and route every flag to a human. A cheap model audits a listing for cents.

Bulk generation before you have a reviewed prompt does not save time; it manufactures the same violation across your whole inventory. One bad phrase in a template is one complaint. The same phrase across two hundred listings is a pattern — and a pattern is what turns a complaint into a case.

Real-World Examples

Example 1: The natural prompt, and what it produces

What an agent actually types:

"Write a warm, compelling MLS description for a 4-bed, 2.5-bath colonial in Maplewood. Great backyard, updated kitchen, finished basement, near the elementary school and St. Anne's."

What comes back (condensed, but this is the shape of it every time):

Welcome home! This charming 4-bedroom colonial is perfect for a growing family. The sun-drenched kitchen is where your family will gather for years to come, and the enormous fenced backyard is a safe, private haven for kids and pets alike. The finished basement makes an ideal playroom or man cave. Located in a quiet, safe neighborhood just steps from top-rated Maplewood Elementary and historic St. Anne's, this exclusive pocket of Maplewood rarely sees inventory. Don't miss your chance to raise your family here!

Count them: perfect for a growing family, your family will gather, safe haven for kids, man cave (a sex-coded occupancy label), safe neighborhood, top-rated schools, St. Anne's as a landmark, exclusive. Familial status, sex, religion and a catch word straight off HUD's own Part 109 list — in five sentences. It is also, by the conventions of listing copy, good. That is the entire problem.

The same house through the constrained prompt:

Four-bedroom, two-and-a-half-bath colonial on a 0.28-acre lot. The kitchen was renovated in 2023 with quartz counters, a 36-inch gas range and a south-facing window over the sink. Hardwood runs through the main level; the finished lower level adds 640 square feet with an egress window and full-height ceilings. The fenced rear yard has a bluestone patio and mature maples. Roof 2019, HVAC 2021. Two blocks to the Maplewood Elementary attendance zone, half a mile to the Route 27 bus.

What changed. Not the appeal — a buyer with three kids reads "four bedrooms, fenced yard, two blocks to the elementary school" and knows exactly what they are looking at. The copy now says it with facts about the house instead of a description of the household. Specificity went up, exposure went to zero, and the second version is checkable where the first was not.

Example 2: The photograph that wrote a lie

Before. An agent uploads 24 photos and asks a multimodal model to "write the listing." It returns: "the chef's kitchen features gleaming granite countertops and stainless Sub-Zero appliances, while the fifth bedroom on the lower level offers flexible space for guests."

The counters are quartz. The refrigerator is a Kenmore behind a panel-ready door. The "fifth bedroom" is an unpermitted basement room whose window well fails egress — advertising it as a bedroom is a misrepresentation, an MLS citation, and in some markets a licensing complaint. None of it was malice: the model saw a stone-look counter, a large built-in fridge and a room with a bed in it, and filled the gaps with the most probable words.

After. The extraction prompt puts countertop material, appliance brand and is this legally a bedroom into the UNCERTAIN column, where the agent checks them against the appraisal and the permit file before a word of copy is written.

What changed. The model stopped being asked to write and started being asked to observe — and to say what it did not know.

Industry-Specific Applications

Rentals and property management. Higher risk, not lower. Familial-status and disability complaints concentrate in rentals, and disability alone drove 54.6% of 2024 complaints. Occupancy-standard language ("2 people max") is a familial-status minefield: get it from counsel, not from a model.

Senior housing. The one place age-restricted marketing can be lawful, under the Housing for Older Persons Act exemption. A model does not know whether your community qualifies, and will cheerfully write "active adult community" for one that does not. If you are not certain you qualify, that phrase is a familial-status violation.

New construction. Models describe finishes from renderings. A rendering is not a fact — every claim must trace to the spec sheet.

Best Practices

• Put the constraint in the prompt, every time. The rule lives in the text or it does not exist.
• Describe the property, never the occupant. If you keep one line from this page, keep that one.
• Make the model cite its own facts. It catches most hallucinations for free.
• Audit in a separate chat, ideally a different model. A model reviewing its own draft is a model defending its own draft.
• Treat the photo pass as observation, not writing, and strip people and religious objects from the images — 24 CFR 100.75 covers pictures too.
• Know which rule you are following — HUD, your state, or your MLS. Only one of the three is federal law.
• Never let AI write the disclosures. Square footage, bedroom count, permits, defects and flood history come from documents.
• Save the prompt, not the output. Your fiftieth listing should take four minutes.

Common Pitfalls

The model reverts. Hand it a long fact sheet and a friendly "make it warmer," and the constraint quietly stops binding — "warm" and "describe the buyer" are nearly the same instruction in its training data. Re-paste the constraint block with every revision, and audit the revision, not just the first draft.

Invented specifics that sound like facts. Granite that is quartz. A 2021 roof that is a 2011 roof. The fluency is what makes them dangerous: nobody double-checks a sentence that reads well.

Bedroom and square-footage inflation. The most expensive AI error in real estate is not a fair housing violation — it is a bedroom that isn't one. Beds, baths and square footage are legally defined and separately actionable.

Copy that is compliant and empty. Strip out every buyer-directed sentence and a lazy model returns a spec sheet. The fix is more property facts, not more adjectives — a south-facing window, a 2023 range, mature maples. Concrete detail is what persuasive listing copy always was, before the industry mistook it for occupant fantasy.

Believing a word list. In both directions. Agents ban "master bedroom," which HUD explicitly permits, while shipping "perfect for families," which § 3604(c) prohibits. The list is not the rule. Property-versus-occupant is the rule.

Assuming the MLS's own generator is compliant. Several MLSs now auto-generate public remarks. The generator is a convenience; the content is still your responsibility, and the fine still arrives addressed to you. No vendor indemnifies you against a HUD complaint, and no regulator has suggested the FHA looks past the person who clicked publish.

Measuring Success

Compliance incidents: zero. The only metric that can go catastrophically wrong. Track MLS citations, broker-review rejections and — the one you hope never to count — agency complaints.

Audit hit rate. What share of AI drafts does the compliance prompt flag? Near zero means the audit prompt is too soft: feed it the bad example above and confirm it catches all eight problems. Near 100% means your drafting prompt has lost its constraint block.

Minutes per listing, end to end. Fact sheet, draft, audit, human review. If the rewrite loop eats the drafting savings, the prompt is underspecified.

Claim traceability. Pick a published listing at random. Can you point at the document behind every number in it?

Engagement, with a large asterisk. Listing performance is dominated by price, photos and inventory, and one agent's listings are far too small a sample to isolate a description's effect. Anyone selling you a percentage lift from AI descriptions is selling you a number they did not measure.

Cost Analysis

Tools and setup. The free tiers of ChatGPT, Claude and Google Gemini cover everything here; paid consumer tiers run about $20/month and buy speed and limits, not compliance. At brokerage scale, API access to a cheap model for the audit pass costs cents per listing. Budget three to four hours, once, to build the fact-sheet template and pressure-test both prompts.

What you get back. Honestly: drafting time, and a review step you probably did not have before. Anyone quoting a percentage increase in showings, offers or sale price from AI-written descriptions is making it up — no credible public study measures it, and the confounders dwarf the claimed effect.

What it costs when it goes wrong. An MLS fair-housing citation runs a few hundred dollars. A HUD or state civil-rights complaint costs legal fees, months, and your license's good standing, whatever the outcome. A thirty-second audit on every listing is the cheapest insurance in the business.

Related Use Cases

• AI for Airbnb Descriptions — the short-term-rental sibling: different platform and rules, same discipline of describing the property rather than the guest.

Related Tools

• ChatGPT — the default drafting tool for most agents
• Claude — long constraint blocks; strongest for the audit pass
• Google Gemini — photo batches and Workspace drafting
• Microsoft Copilot — drafting inside Word and Outlook
• Jasper AI — brand-voice templates across an inventory
• Perplexity AI — sourcing a neighborhood claim before publishing
• n8n — bulk draft-then-audit pipelines

Related Models

• Claude Sonnet 5 — reliable constraint-following on long rubrics
• GPT-5.5 — capable generalist for drafting and photo extraction
• Gemini 3.5 — large context window for whole photo sets
• Claude Haiku 4.5 — cheap enough to audit every listing you publish

New to writing prompts? Start with prompt engineering. Here the prompt is not a convenience — it is the only thing standing between a persuasive sentence and a federal advertising violation.