Compliance: Fair Housing Language and AI

You've spent four days building speed. Today you build defensibility. Speed without compliance is a liability — and AI's failure modes in real estate are unusually expensive. A single discriminatory phrase in an MLS description can trigger a HUD complaint, a state real-estate commission action, and an Article 10 ethics charge from your local board, all from the same posting.

The good news is that the rules are knowable, the violations are pattern-shaped, and AI itself is the best compliance tool you have — once you wrap your prompts correctly.

The Statutes, In Plain Language

Fair Housing Act, 42 U.S.C. §3604(c)

The operative text: it is unlawful "to make, print, or publish, or cause to be made, printed, or published 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, or an intention to make any such preference, limitation, or discrimination."

Three things to internalize. First, "publish" includes online listings, social posts, emails, and texts. Second, the statute reaches both express statements ("families only") and statements that "indicate" preference (the proxy phrases we covered all week). Third, intent is irrelevant — the test is whether the statement, read by an ordinary reader, indicates preference.

HUD Advertising Guidance and Algorithmic Bias

HUD's longstanding interpretive list of problematic words and phrases (originally in 24 CFR Part 109, withdrawn but still cited by enforcement) is the operating dictionary regulators use. HUD's 2022 guidance on Section 8 voucher discrimination and the 2023 Office of Fair Housing and Equal Opportunity guidance on tenant-screening algorithms extend the framework to AI tools — meaning if you use an AI to score, rank, or filter prospects, the AI's output is your output, and disparate impact analysis applies.

NAR Code of Ethics Article 10

Article 10 binds every REALTOR®. The headline: "REALTORS® shall not deny equal professional services to any person for reasons of race, color, religion, sex, handicap, familial status, national origin, sexual orientation, or gender identity. REALTORS® shall not be parties to any plan or agreement to discriminate against a person or persons on any of these bases." The Standards of Practice that bite hardest in AI work:

• SOP 10-1 — REALTORS® shall not volunteer information regarding the racial, religious, or ethnic composition of any neighborhood. (Strip neighborhood demographic narrative from every CMA, listing, and email.)
• SOP 10-2 — REALTORS® shall not engage in any activity that would result in discrimination as proscribed by Article 10. (Includes prompts that produce discriminatory output if you publish that output.)
• SOP 10-3 — REALTORS® shall not print, display, or circulate any statement or advertisement that indicates any preference based on a protected class. (Mirrors §3604(c) and applies to AI-generated content.)
• SOP 10-5 — REALTORS® shall not use harassing speech, hate speech, epithets, or slurs based on protected classes. (Prompt-injection content from training data can produce this; review every output.)

The Compliance Preamble

Paste this into the system or instructions field of your AI tool, once. Use it for every real-estate prompt going forward.

You are assisting a licensed real estate
professional in the United States. Every output
you produce must comply with:

1. The Fair Housing Act, 42 U.S.C. §3604(a)-(d),
   especially §3604(c) (advertising) and the
   prohibition on indicating any preference,
   limitation, or discrimination based on race,
   color, religion, sex, handicap, familial
   status, or national origin.
2. HUD's interpretive guidance on advertising,
   tenant screening, and algorithmic decision
   support.
3. NAR Code of Ethics Article 10 and Standards
   of Practice 10-1, 10-2, 10-3, 10-5.
4. Equal Credit Opportunity Act prohibitions on
   discrimination in lending-related referrals.
5. CAN-SPAM (15 U.S.C. §7704) for any email
   output and TCPA (47 U.S.C. §227) for any SMS.

DO NOT produce:
- Phrases naming or implying preference about a
  protected class ("perfect for families,"
  "great for kids," "active adult community,"
  "exclusive," "private," "Christian
  neighborhood," "ethnic community").
- Demographic or religious-composition
  commentary about neighborhoods.
- Steering language that directs prospects
  toward or away from properties based on
  perceived protected-class membership.
- Inferences about a lead's family, religion,
  national origin, or disability status from
  name, photo, school choice, or any other
  signal.

If a user request would produce non-compliant
output, refuse and propose a compliant
alternative.

That preamble is doing two jobs at once. It constrains generation, and it gives you a written record that you instructed your AI to comply — useful if a complaint ever comes.

The 90-Second Pre-Publish Checklist

Before any AI-generated content goes live, run this on it. Literally read each line and check.

1. Does any sentence describe who the property is for? (Strike or rewrite.)
2. Does any sentence reference family, kids, parents, retirees, or empty nesters? (Strike unless the property qualifies under HOPA's 55+ exemption with documentation.)
3. Does any sentence name a place of worship as a feature? (Strike. Schools and parks fine.)
4. Does any sentence describe the neighborhood's residents? (Strike.)
5. Does any sentence use "exclusive," "private," "traditional," "established," or "desirable community"? (Replace with property-specific facts.)
6. Are pricing and adjustment claims tied to property and market data, never to demographic narrative? (Verify.)
7. For emails: Is there a working unsubscribe sentence and a physical postal address? (CAN-SPAM.)
8. For SMS: Did the recipient give prior express written consent? (TCPA.)
9. For ads on Meta or Google: Is the campaign in the platform's Housing special-ad category? (Required for housing.)

Ninety seconds. Do it every time. Build the muscle.

Documentation: The Audit Trail That Saves You

The lightest-weight system: a single shared folder per listing, with subfolders for "prompts," "drafts," "published." It takes thirty seconds per piece of content and can save you a HUD investigation. The minute you publish without a record is the minute you have nothing to show a regulator.