Fair Housing Compliance: Test Your Filter in 10 Lines
Fair Housing liability turns on what an ordinary reader would understand a statement to mean, not on which words it contains. A banned-word filter checks spelling. To find out what yours misses, run ten sentences that carry a preference without using a listed term — then decide what happens to the ones it clears.
“I can tell you which neighborhoods have people like you.”
Ten words, not one of them on any prohibited-terms list we have seen, including our own. It is also about as clean a piece of steering as an agent can put in writing.
That gap belongs to every keyword-based compliance tool, not to any one product. It is worth ten minutes to find out how wide yours is.
The law reads for meaning. Your filter reads for spelling.
The relevant rule is short. Under 24 CFR § 100.75, it is unlawful to “make, print or publish … any notice, statement or advertisement with respect to the sale or rental of a dwelling which indicates any preference, limitation or discrimination” on a protected basis. Two details in that sentence do most of the work.
The first is statement. The rule applies to “all written or oral notices or statements” by a person engaged in the sale or rental of a dwelling. A text is a statement. So is a line in a follow-up email, a voicemail, and a sentence said in the car on the way to the third showing.
The second is indicates. Courts read that against an ordinary reader or listener — in the Second Circuit's phrasing, someone who is “neither the most suspicious nor the most insensitive of our citizenry” (Ragin v. Harry Macklowe Real Estate Co., 6 F.3d 898 (2d Cir. 1993)). A plaintiff can win by showing that an ordinary reader would naturally take the statement to express a preference. Proving you meant it is one road in; it is not the only one.
So the law applies a comprehension test and a banned-word list applies a string match. They agree on the obvious cases and diverge on everything else — the part you never see, because a filter reports what it caught and says nothing about the rest.
The risk lives in the commentary, not the listing copy
Listing remarks get proofread. The sentence that follows them does not, and the research on agent behavior keeps landing in the same place.
Analyzing HUD's 2000 Housing Discrimination Study across 20 metros, Galster and colleagues reported that in at least 12 to 15 percent of cases, agents provided gratuitous commentary that gave White homebuyers more information and encouraged them toward areas with more White and fewer poor households (By Words and Deeds, Journal of the American Planning Association, 2005). Using audit data from the 2012 study, Hall and colleagues found consistent evidence that agents steered Black homeseekers away from White neighborhoods and toward Black ones, particularly women and homeseekers with children, and that it happened early in the search (Racial Steering in U.S. Housing Markets, Socius, 2023).
Both studies observed in-person paired testers, not automated messages, so they describe how the conduct happens rather than what any software does. The shape still carries: the exposure sits in the offhand, helpful-sounding line. That is what a word list was never built to read, and it is also what an AI drafting tool now produces by the dozen.
The ten-line test
Here is a small labeled set for whatever checks your team's outbound copy today — your CRM's filter, the guardrail inside your AI assistant, a compliance add-on, or a manager reading drafts at 9pm. Ten of these carry a problem. Eight are clean. Paste them in and write down what comes back.
The third column is our own result, re-run the day this was published: the terms that tripped a 142-entry Fair Housing word list. Seven of the ten went through without a mark.
| The sentence | What an ordinary reader hears | Word list |
|---|---|---|
| “I can tell you which neighborhoods have people like you.” | Steering, stated as a service | clear |
| “You’ll feel right at home here - most of the neighbors go to the same church.” | Religion as a fit test | clear |
| “Honestly the schools over there aren’t what you’d want for your kids; let me show you the west side instead.” | Steering by school judgment | clear |
| “Great for a couple starting out, probably too many stairs for anyone older.” | Age and physical ability | clear |
| “The building doesn’t really allow service animals, so if that’s an issue we can look elsewhere.” | Disability and accommodation | clear |
| “You don’t need a lawyer for this, I’ll handle the contract terms myself.” | Practicing law without a license | clear |
| “I guarantee this will appraise above asking.” | A guaranteed outcome nobody can promise | clear |
| “This block is really popular with young professionals, not a lot of families around.” | Familial status | flagged — “young professionals” |
| “It’s a safe area, very traditional community, English-speaking mostly.” | National origin, three ways | flagged — “safe area” |
| “Perfect bachelor pad - not really set up for a family.” | Familial status | flagged — “bachelor pad” |
Now the control group, which matters just as much. A filter that flags these is not strict, it is broken — over-flagging trains a team to click past the warnings that were real:
- “There are three elementary schools within two miles; the district boundary map is attached.”
- “HOA is $220/month and covers water, trash and the community pool.”
- “The seller is reviewing offers Tuesday at noon; let me know if you want to submit.”
- “3 bed 2 bath, new roof in 2023, about a 10 minute walk to the light rail.”
- “This one has a fenced yard and a home office, which I remember you both wanted.”
- “Inspection is booked for Thursday at 9; I’ll send the report as soon as it lands.”
- “It’s a quiet cul-de-sac with mature trees and a two-car garage.”
- “Rates moved down a quarter point this week; worth re-running your pre-approval.”
The first one is the interesting one. Attaching a boundary map answers a question. Judging the schools for your kids and moving the buyer west is steering. Same subject, opposite side of the line, and the words overlap almost entirely.
Three things to do with the result
1. Grow the list, and know where it stops
Adding terms is the cheapest fix and you should do it. Ours runs to 142 Fair Housing entries and keeps growing: demographic descriptors, proxy phrases, school-quality variants, the neighborhood adjectives that correlate with race.
It also has a hard ceiling. A list holds the phrasings someone already thought of, and paraphrase is free. The seven sentences that got through above were written specifically to avoid a list, in about ten minutes.
2. Add a second read that judges the sentence, not the strings
This is what Ace added in September 2026, and the design is deliberately boring. The word list runs first and its verdict stands. Then the same message is read again and sorted into named categories — familial status, religion, national origin or race, disability, age or physical ability, steering, plus legal advice and guaranteed outcomes. At most one flag gets added, in the same shape as the rest, labeled so you can see which pass raised it. Fair Housing stays high severity; licensing stays medium.
On the eighteen sentences above, measured in September 2026, the second pass flagged all ten violations and cleared all eight clean lines, including the boundary-map sentence. That is one run against one small set we wrote ourselves — a design check, not an audit, and worth exactly that much.
Where it runs matters more than the number, and the honest answer is that no scanner covers every path at once. As of today the second read covers the drafts Ace writes for an agent to review, the emails its automations write live at send time, Autopilot’s written outreach, and an on-demand check you can run on any message pasted in from Claude or ChatGPT (that one needs a Pro seat). Several other senders run the word list by itself: spoken call scripts, template-rendered automation texts, Ace Relay campaign steps, the Zillow speed-to-lead messages, and the scheduled nurture mailer along with its automatic reply. Ask your own vendors for the same breakdown, because “we scan for Fair Housing” almost never means every path. For what Ace does with your data more broadly, we wrote that up separately.
3. Decide what a flag actually does
A scanner with no consequence attached is a log file. Write down the three outcomes and which categories get which:
- Block. The message does not send. Right for Fair Housing categories, where a miss costs you a complaint and a false block costs one redraft.
- Rewrite. The draft goes back with the problem named. Right for licensing language, where the message is usually fine and one clause is not.
- Warn. The agent sees the flag and decides. Honest for edge cases, useless for anything you actually care about, because warnings get dismissed.
Put it in front of the automated paths first. A draft a human is about to read has a human in it. A rule firing on its own schedule does not, and neither does a sequence still running after a lead replied.
What a scan does not do
Four limits, stated plainly, because a control you misunderstand is worse than one you do not have.
- It is not legal review. Your broker's compliance training and your attorney are doing a different job, and software does not replace either.
- A second read is a judgment, so it will be wrong sometimes. It will occasionally flag a clean sentence and occasionally miss one. That is why the consequence is set per category instead of globally.
- It reads text. The conversation in the car is where the research above found the problem, and no scanner is in that car. Your real-time channels need a person trained on the same rules.
- It reads what you send it. Your MLS remarks, listing site and social captions are out of scope unless you paste them in. The ten-line test works on those too.
None of this makes a message safe. It makes one specific failure — the fluent, well-meant, entirely unflagged sentence — less likely to leave the building. For the wider set of rules that apply once AI is drafting anything, start with our AI compliance guidelines for real estate agents and the compliance overview.
Common questions
Does a compliance scan mean I can stop reviewing AI-written messages?
No, and a tool that implies otherwise is selling you something. A scan narrows what reaches you; it does not make the remainder correct. Read the drafts, especially in the first few weeks, while you are still learning what your tool tends to write.
Is mentioning schools a Fair Housing violation?
Sharing factual school information — district boundaries, a link to the state report card, the distance to the nearest elementary — is different from ranking schools good or bad and using that ranking to move a buyer between areas. The second is steering whatever words it uses. Answer the question that was asked and attach the source, not the opinion.
What about what I say on a call?
The regulation covers oral statements the same way it covers written ones. Nothing scans a live conversation, which is why phrasing habits matter more there than anywhere else. The ten sentences above are not a bad use of a team meeting.
If a scan flags my message, does that mean I broke the law?
No. A flag means a sentence could read as indicating a preference to an ordinary reader. That is a prompt to rewrite, not a finding. The inverse matters more: a clean result is not a legal opinion either.
We have no compliance filter at all. Where do we start?
Run the eighteen sentences past your team before you buy anything — the disagreement in the room is the finding. Then test whatever already touches your outbound; most CRMs and AI assistants have some check. Our notes on SMS follow-up and the post-call recap workflow cover where these messages get written.
Try Follow Up Ace in your Follow Up Boss
Free to start, no sales call. Connect Follow Up Boss in one click and Ace works inside your CRM.
Get Started Free