Every major AI model can write a real estate listing description. Most of them will write one that sounds polished, professional, and confidently wrong, full of fair-housing violations the agent will never catch, invented facts the MLS will flag, and protected-class signals that read as steering under the legal standard courts have applied since 1988. This is a working guide to what fair housing compliance actually requires from any AI tool you use for listing copy, and the five specific failure modes to watch for before you hit publish.
The AI listing boom: and what agents are actually getting
Agents are using AI for listing descriptions in large numbers now. The use case makes obvious sense: a listing description is a well-defined, repeatable writing task with clear inputs (property data, features, neighborhood) and a predictable output format. AI handles repetitive structured writing well. The time savings are real.
The problem is that listing descriptions are not just marketing copy. They are regulated speech. Under the Fair Housing Act, a listing description that an ordinary reader would interpret as expressing a preference for or against a protected class is a violation, and the agent, broker, and brokerage can all be named in the complaint. Strict liability applies to the content; intent to discriminate is not required.
Generic AI tools, ChatGPT, Claude without a real estate compliance layer, Gemini, Perplexity, even AI embedded in listing platforms, are optimized to write convincing, readable prose. They are not optimized for fair housing. They do not know which phrases HUD has flagged in published enforcement actions. They do not apply the "ordinary reader" standard. They do not check against NAR Code of Ethics Article 10. They write what sounds good to a human, not what survives scrutiny by a fair housing organization or a state commission investigator.
The result is listing copy that an agent approves because it reads well, goes live on the MLS, and lands in a complaint six months later.
What the law actually requires
The Fair Housing Act prohibits advertising that "indicates any preference, limitation, or discrimination based on race, color, religion, sex, handicap, familial status, or national origin." HUD's enforcement guidance, particularly the 1995 Fair Housing Advertising memo, elaborates on what that means for listing copy in practice.
The legal standard is the ordinary reader test, articulated in Ragin v. New York Times Co., 923 F.2d 995 (2d Cir. 1991): would an ordinary reader, reading the listing, perceive it as expressing a preference for or against people in a protected class? The standard is objective, not subjective. It does not matter what the agent intended. It does not matter that the agent has sold homes to buyers of every background. What matters is how the language reads to someone who encounters it.
This matters for AI use because the question is never "did the AI intend to discriminate?" The question is whether the output, which you reviewed and published, reads as steering to an ordinary reader. You are responsible for what goes live.
State real estate commissions layer additional requirements on top of the federal floor. Most states have their own fair housing statutes that extend the protected-class list beyond the federal seven, common additions include source of income, sexual orientation, gender identity, marital status, age, and military status. The AI you use has no idea which state you're in unless you tell it, and even then it is unlikely to have current guidance from your state's commission.
The five failure modes of AI listing copy
1. Protected-class-adjacent language that implies preference
The most common failure. AI trained on real estate listing data has absorbed decades of language that was industry-standard before modern fair housing enforcement, and it reproduces it fluently. Phrases like "perfect for families," "great for empty nesters," "ideal for a couple," "senior-friendly," and "close to houses of worship" all signal a preference for or against a protected class under either the federal act or most state statutes. The AI does not flag these. It uses them confidently because they appear constantly in the data it was trained on.
What an unflagged AI will write
"This charming ranch is perfect for young families or retirees looking for a quiet, walkable neighborhood close to top-rated schools and houses of worship."
That sentence contains four fair housing problems: "young families" (familial status + age), "retirees" (age), "walkable" (disability-adjacent depending on context), and "houses of worship" (religion). A well-trained human reader notices most of them. An AI without a fair housing filter generates them in one pass.
2. Invented facts presented as verified
AI models are trained to be helpful and fluent. When they don't have a specific fact, they estimate, and the estimate reads like confident prose. Square footage gets rounded. School ratings get invented or outdated. HOA amenities that may or may not exist get listed as features. Year built gets approximated. This is not the AI lying; it is the AI doing what it was trained to do, which is produce readable, complete-sounding text.
On an MLS, invented facts are a material misrepresentation. Buyers and their agents rely on listing data. A discrepancy between the listing and the property, even one generated by an AI the agent never thought to double-check, is the agent's liability, not the AI's.
3. Neighborhood characterizations that imply steering
Phrases like "safe neighborhood," "up-and-coming area," "established community," and "charming, quiet street" can all read as demographic signals to an ordinary reader depending on context. HUD's guidance explicitly identifies terms that can indicate a racial, ethnic, or national-origin preference even without naming those characteristics. AI trained on real estate marketing copy uses these terms fluently because they appear in training data constantly.
"Safe neighborhood" is the most commonly cited example, it implies that the surrounding area is not safe, which courts have found can be a coded signal. The AI uses it because it reads well. It is flagged in HUD published enforcement guidance because it steers.
4. Language calibrated for a perceived buyer rather than the property
AI listing tools that take neighborhood data as input, nearby schools, walkability scores, demographic summaries, can generate copy that is implicitly calibrated to who currently lives in the neighborhood rather than to the property's features. This is steering at the output level. The listing reads as if it is written for a particular type of buyer, not as if it is describing a home anyone might want to purchase.
This is subtle, and it is almost impossible to catch without a filter specifically designed to look for it. The copy does not use any flagged phrases. It just happens to emphasize features that appeal to one demographic group while omitting features that would appeal to others.
5. Inconsistent compliance across the same agent's listings
This is a risk specific to AI use. When an agent writes listings manually, their language patterns are relatively consistent, for better or worse, they have habits. When an agent uses AI, the output varies based on the prompt, the model, the day, and what else was in the conversation context. The result is that compliance level can vary dramatically across listings from the same agent. One listing is clean; the next contains two violations. This inconsistency makes it hard to catch problems early and creates an audit trail that is difficult to defend.
What the law requires from your AI tool: specifically
Given these failure modes, here is what fair housing compliance actually requires from any AI you use to write listing copy:
A fair housing filter that runs on every output, not just on request. The filter needs to check against the federal protected classes under the Fair Housing Act and against the extended list for your state. It should flag both explicit references and ordinary-reader-test violations, the language that implies a preference without stating one. And it needs to run automatically, on every draft, not only when the agent thinks to ask.
A fact-grounding rule. The AI must be instructed to use only facts the agent provides, to flag when it lacks a fact rather than estimate, and never to invent MLS-verifiable information like square footage, school ratings, or year built. This needs to be a hard rule in the AI's instructions, not a suggestion in the prompt.
Jurisdiction awareness. The AI needs to know which state you operate in, and it needs to have current guidance from that state's real estate commission and any state-level fair housing statutes. A generic AI knows the federal seven protected classes. It does not know that your state adds source of income, or that your state commission specifically prohibits certain neighborhood characterizations, or that your MLS has a rule about school proximity language.
Explicit language prohibitions, not just general guidance. Telling an AI "write fair housing compliant copy" does not produce compliant output. The AI needs an explicit list of phrases to avoid, categories of language to flag, and what to do when it catches a problem. "Flag it and explain why" is more useful than silently removing the phrase, because the agent needs to understand what the violation was.
Human review before publication: built into the workflow, not optional. No AI should be publishing listing copy directly. The agent reviews every draft. But the review needs to be meaningful, not rubber-stamp. That means the AI's output should include a compliance note, what it checked, what it flagged, what it changed, so the agent is making an informed decision, not just reading prose that sounds fine.
What a compliant AI listing workflow looks like in practice
A compliant AI listing workflow has three stages, not one.
Stage one: input control. The agent provides the AI with MLS-verified property data, square footage, beds, baths, year built, features from the seller disclosure, and nothing else. No neighborhood characterizations. No school ratings from third-party sites. No demographic context. The AI's job in stage one is to draft copy from verified facts only.
Stage two: filter and flag. Before returning the draft to the agent, the AI runs its compliance check. It looks for protected-class signals, ordinary-reader-test issues, neighborhood language that could read as steering, and any statements it cannot verify from the input data. It returns the draft with a compliance note that lists what it checked and any flags.
Stage three: agent review and approval. The agent reads the draft and the compliance note. They make any edits, approve the copy, and publish. The AI's compliance note creates a paper trail, not a legal shield, but a record of the process the agent followed.
This is different from the typical AI listing workflow, which is: paste some notes into ChatGPT, read the output, copy it into the MLS. The typical workflow has no filter, no fact-grounding rule, no compliance note, and no paper trail. It is fast and it is the workflow most agents are using right now.
How to evaluate any AI listing tool
If you are evaluating AI tools for listing copy, whether that is a standalone AI, a platform with AI built in, or a custom prompt you have been using, here are the questions that matter for compliance:
- Does the tool run a fair housing filter on every output, automatically? Not "can I ask it to check", does it check without being asked?
- Does it know the fair housing rules for your specific state? Ask it. "What are the protected classes under [your state]'s fair housing laws?" If it gets this wrong or gives you the federal list only, it does not have the state layer.
- What happens when it lacks a fact? Ask it to write a description for a property and leave out the square footage. Does it flag the gap, or does it fill it in? If it fills it in, you have a fact-invention problem.
- Does it explain what it flagged? Ask it to write copy that includes "perfect for families." Does it catch it, explain why, and suggest alternative language, or does it silently produce the phrase?
- Is human review built into the workflow, or is the tool designed to publish directly? Any tool designed to publish listing copy without agent review is not a tool designed for compliance.
Why this matters more now than it did two years ago
Fair housing enforcement around AI-generated listing copy is an emerging area. HUD has not issued formal guidance specific to AI-generated listings as of this writing, but the legal framework that governs the content, the Fair Housing Act, the ordinary reader test, state commission rules, applies regardless of how the copy was generated. "The AI wrote it" is not a defense. The agent published it.
What has changed is volume and consistency. Before AI, an agent wrote listing copy occasionally, maybe one or two new listings a month. With AI, the same agent can produce listing copy for every property, every price reduction announcement, every social post, every email to the neighborhood about a new listing. The compliance surface has expanded dramatically, and the exposure has expanded with it. One compliant listing out of ten is not a compliance program. Consistent compliance across every piece of listing-related content is.
The agents who build a compliant AI workflow now, before enforcement focus sharpens on AI-generated listings, are the ones who will have defensible practices when it does. The agents who assume that polished-sounding AI output is safe are taking a risk that is hard to see until a complaint arrives.
The listing-description skill does this automatically.
The Real Estate Agent Toolkit includes a listing-description skill with a fair housing filter built into the guardrails block, it runs on every output, flags what it caught, and uses only the facts you provide. No unflagged "perfect for families." No invented square footage. Works with Claude, ChatGPT, Cursor, or any chat AI.
