AI listing descriptions are the most widely adopted use of artificial intelligence in residential real estate, and the one that sits closest to a fair housing complaint. NAR's 2025 REALTORS® Technology Survey, fielded in July 2025 among 1,241 respondents with a margin of error of 2.78 points, found 46% of respondents using AI-generated content, with listing descriptions as the named example. What it did not ask is who reads that output before it reaches the MLS.
That gap is the problem. A model trained on decades of real estate copy has absorbed the coded language the industry spent fifty years being sued over. Ask it for warm remarks and it offers "perfect for a growing family" or "quiet, safe street" without hesitation, because that is what listing remarks have historically said. It will also state a square footage it inferred from a photograph. Both are publishing decisions, and the license belongs to whoever published them.
This is a working guide for broker-owners, team leads, and marketing directors running five or more agents: where the exposure sits, why agent-level checklists stop working at volume, and what a review process looks like that holds up in front of a regulator or an insurer.
Table of Contents
- What AI Listing Descriptions Do Well, and Where They Break
- The Fair Housing Exposure Sits in Two Places
- Why an Agent-Level Checklist Is Not Enough
- A Review Process Your Brokerage Can Defend
- What Not to Automate
- A 90-Day Rollout for a Team or Brokerage
- Frequently Asked Questions
- Ready to Put a Defensible Process Behind Your Listings?
What AI Listing Descriptions Do Well, and Where They Break
AI listing descriptions are drafting tools: dependable for structure, tone, and length, unreliable for facts and for fair housing judgment. The useful mental model is a capable new hire who writes quickly, has never seen the property, and does not know the law. Turning a verified feature list into readable prose and holding one voice across thirty agents is real work, and none of it requires judgment.
Failure mode one: the copy describes the buyer instead of the house. The Realtors Association of Northeast Wisconsin publishes a questionable words and phrases list whose operating rule is the one to memorize: remarks should "describe the property and not the potential buyers." It flags phrasing tied to familial status (family, couple, empty nester, perfect for), disability (handicap), age (mature, retired), and neighborhood composition (exclusive, prestigious). Models produce this language constantly, because it reads as warm.
Failure mode two: the copy states things that are not true. Models fill gaps. Given eight photos and a bullet list, one will supply a room count, a lot size, or a walking distance drawn from pattern rather than the record. NAR's 2026 Code of Ethics does not ask where the sentence came from: Article 12 requires REALTORS® to "present a true picture in their advertising, marketing, and other representations."
Adoption without a process shows up in the results. In the same NAR survey, 17% of respondents reported a significant positive business impact from AI and 46% reported none they could notice.
The Fair Housing Exposure Sits in Two Places
Most coverage treats this as a vocabulary problem. It is two problems: what the ad says, and who the ad is shown to. Recent federal attention has concentrated on the second, which brokerages consistently under-manage.
The language side. 42 U.S.C. 3604(c) makes it unlawful to make, print, or publish any housing advertisement indicating a preference, limitation, or discrimination based on a protected characteristic. It is about what was published, not who typed it.
The distribution side. HUD's April 29, 2024 guidance on advertising through digital platforms applies that same statute to targeting and delivery. It names three mechanisms: audience targeting tools that segment along protected characteristics or their proxies, custom and mirror audience tools that reproduce the bias in whatever list you upload, and the platform's own delivery algorithm. HUD is direct about the part brokerages miss, noting this "can happen without the advertiser's direction or knowledge." It adds that any entity playing "a substantial role" in a discriminatory housing decision can be held responsible, which includes the advertiser, not only the platform.
That is not theoretical. In its June 21, 2022 settlement with Meta, the Justice Department alleged that "Meta's ad delivery system uses machine-learning algorithms that rely in part on FHA-protected characteristics such as race, national origin and sex to help determine which subset of an advertiser's targeted audience will actually receive a housing ad." Meta paid the statutory maximum of $115,054, retired its Special Ad Audience tool, and rebuilt delivery under court oversight.
The translation for a brokerage. A lookalike audience built from your past buyer list is a mirror audience, the category HUD names directly. Use the platform's housing ad category on every listing campaign so its safeguards apply, stop building lookalike or uploaded-list audiences, and read delivery reports for skew, not only cost per click.
Why an Agent-Level Checklist Is Not Enough
An individual checklist works when one person writes one listing. It degrades at volume, fastest in the busy weeks, which is precisely when supervision matters. HousingWire's May 2026 pre-flight checklist is sound advice, and it is scoped to the agent. It says nothing about who supervises, what gets sampled, what gets logged, or who absorbs the loss when someone skips a step.
Run the arithmetic for a mid-sized shop. Assumptions, all of which you should replace with your own: 40 agents, 12 listings each per year, so 480 listings. Each produces about five pieces of public copy (MLS remarks, property page, two social posts, one email), roughly 2,400 published assets a year. Assume 2% carry a problematic phrase or an invented fact past the agent's own review: 48 defective items a year, several involving protected-class language.
Now price the fix. A careful read against a fixed rubric takes about four minutes, so reviewing all 2,400 costs 160 hours, close to a month of one person's time, or roughly $4,800 at a $30 fully loaded hourly rate. Set that against the drafting time saved: at ten minutes per asset, the brokerage recovers about 400 hours. Review consumes 40% of the gain and still leaves 240 hours. The lesson is not that review costs too much, but that undifferentiated review is wasteful, and a brokerage buying AI for speed should budget the review hours before counting the savings.
A Review Process Your Brokerage Can Defend
A defensible process has four parts: a controlled input, a tiered review, a sampling audit, and a log. Process first, technology second: build all four before buying anything.
- Control the input, not just the output. Publish one approved prompt template that accepts only verified fields from the listing agreement or MLS record, carrying a standing instruction to describe the property and never the occupant. Agents fill in fields; they do not free-type neighborhood impressions into a chatbot and hope.
- Tier the review by risk. Tier one is an automated term screen against your MLS's questionable-words list, run before anything is saved. Tier two is the agent's own check on every asset. Tier three is named-reviewer approval, reserved for what is most likely to fail: anything referencing schools, neighborhoods, safety, or community character, anything describing an occupant type, and anything containing a number. Expect that to be a minority of your volume, perhaps a fifth, and measure it in the first month rather than assuming it.
- Sample and audit monthly. Pull 10% of published remarks at random and score them against a fixed rubric: describes property not people; every number traceable to the record; no unmeasurable adjectives (quiet, safe, exclusive); no school or crime claim; no protected-class reference or proxy. Record the failure rate by agent and by category. An error rate nobody measures is one nobody can show improved.
- Log the approval. Date, agent, tool, approver, and final text, retained for the advertising record period your state license law requires. That log is the difference between "we tell our agents to be careful" and evidence of a supervised process.
Signals to watch between audits. Copy that describes people rather than rooms. Adjectives with no measurable referent. Numbers absent from the MLS record. The same phrasing across unrelated listings, meaning agents share a prompt with a bad instruction baked in. Ad sets built from uploaded contact lists. Any of these in a spot check is reason to pull the month's sample early.
The principle underneath all of it: AI enhances human judgment, it does not replace people. The model drafts, a licensed human decides what gets published, and the brokerage can prove who that was.
What Not to Automate
Some categories should never reach publication without human approval, and two should not go to a general-purpose model at all.
Require human approval every time: any characterization of a neighborhood, any reference to schools, any statement about safety or how an area feels, and any sentence about who the home would suit. These four account for most fair housing exposure in listing copy, and no term screen catches them reliably, because the problem is meaning rather than vocabulary.
Do not delegate at all: audience targeting for housing advertising, for the reasons HUD set out in 2024, and any screening judgment about a prospective buyer or tenant. Keep client financial and transaction data out of consumer chatbots that train on inputs.
The test behind AI for real estate brokerages applies here: automate where a wrong answer is cheap and obvious, keep people where it is expensive and invisible. A clumsy sentence is cheap and obvious. A steering implication is neither.
A 90-Day Rollout for a Team or Brokerage
Ninety days moves a team from ad hoc use to a documented process, if the first thirty go to measuring rather than shopping.
Days 1 to 30, establish the baseline. Export the last 90 days of published remarks, score 50 at random against the rubric above, and write down the failure rate. That number is the only evidence you will have that anything improved. Draft a one-page AI marketing policy covering approved tools, the approval step, and prohibited categories. Nothing is purchased this month.
Days 31 to 60, pilot narrowly. Three to five agents, the approved template, tier-three review on every asset, and the log running from day one. Meet weekly and read the failures aloud; this is where you learn which instructions in your template are not working.
Days 61 to 90, roll out and instrument. Train the roster in one session built on real examples from the pilot, move to the tiered model, and put the monthly audit on the calendar with a named owner. If the failure rate is not falling by month four, the problem is the process, not the agents.
For where AI fits elsewhere in the operation, AI marketing for small business covers sequencing and getting cited by AI search engines covers visibility.
Frequently Asked Questions
Can a fair housing complaint really come from a description a computer wrote?
Yes. Section 3604(c) of the Fair Housing Act prohibits publishing housing advertisements that indicate a preference or limitation based on a protected characteristic. The statute addresses what was published, not what drafted it. The agent and the brokerage are the publishers, so the tool that produced the sentence is not a defense.Doesn't my MLS screen listing remarks for this already?
Many MLSs screen remarks against a questionable-words list and issue fines, but coverage varies and a screen is not a compliance program. Filters catch vocabulary, not meaning, so a phrase implying who belongs in a neighborhood can pass cleanly. They also cannot detect an invented square footage. Assume the remaining gap is yours.Do we have to disclose that a listing description was AI-assisted?
Nothing in the Fair Housing Act requires disclosure for listing remarks. What applies is accuracy under NAR Code of Ethics Article 12, which requires a true picture in advertising. Image disclosure is separate: virtual staging and altered photos carry labeling requirements under many MLS rules and state license laws. Confirm your local rules with counsel.Which carries more risk, the wording or the ad targeting?
Targeting, judged by the enforcement record. The Justice Department's 2022 Meta settlement and HUD's 2024 advertising guidance both concentrate on audience selection and delivery rather than vocabulary. Wording errors are easier to catch, because a person can read a paragraph in seconds. Nobody reviews an audience definition unless the brokerage requires it.How much review time should a brokerage budget?
Use your own volume. At roughly four minutes for a careful rubric-based read, a shop publishing 2,400 marketing assets a year would spend about 160 hours reviewing everything. A tiered model concentrates senior review on assets mentioning people, places, schools, or numbers, often about a fifth of the volume, and leaves the rest to a term screen and the agent's own check.Ready to Put a Defensible Process Behind Your Listings?
The question is no longer whether agents will use AI to draft listing copy, because they already are. It is whether anyone can show what was reviewed, by whom, and with what result. ANtreliX is a strategy-first AI and business intelligence consultancy based in North Myrtle Beach, South Carolina, working with real estate teams and brokerages locally and remotely, and we build the measurement before the tooling. If you want that process documented rather than assumed, see how we approach this work or Talk to ANtreliX.