Key takeaways
- AI can treat people unequally without anyone intending it, through proxies, uneven accuracy and inconsistent answers.
- The Fair Housing Act protects seven characteristics in buying, renting and related activities, and state laws add more[1][2].
- Audit outcomes, not just rules: compare response times, booked appointments and answer quality across areas, sources and languages.
- Fair housing training is required of REALTORS each cycle, and it is the natural place to review AI rules with the team[3].
How bias gets in without anyone intending it
Most agents would never knowingly treat a buyer differently because of race, religion or family status. AI raises a different kind of risk: unequal treatment that nobody chose, produced by rules, data or technical limits that affect some groups more than others. Because AI applies its rules to every lead, consistently and at volume, a small problem in design can become a pattern quickly.
The Fair Housing Act prohibits discrimination because of race, color, national origin, religion, sex, familial status and disability[1], and it protects people when renting or buying a home, getting a mortgage, seeking housing assistance or in other housing-related activities[2]. Lead handling, how quickly someone gets a response, what they are told and which homes they hear about, sits squarely inside that. Many states and cities protect additional characteristics such as age, source of income, marital status, sexual orientation or gender identity.
Where AI touches lead handling
| Stage | How bias can appear | What to check |
|---|---|---|
| Lead sources and ads | Targeting that reaches some groups and not others | Who sees your ads and which areas your leads come from |
| Answering and transcription | Speech recognition that works worse for some accents or languages | Summary accuracy by caller language and accent |
| Answers to questions | Different information about areas, schools or availability depending on how a question is phrased | Paired test questions with different personas |
| Qualification and scoring | Scores that use proxies such as area code, ZIP code or name | Every input to the score and its source |
| Routing and speed | Some leads routed to slower queues or less experienced agents | Response time and booked appointments by area and source |
| Rental screening | Automated criteria that exclude groups disproportionately | Criteria, exceptions and outcomes, reviewed by counsel |
Six risk points in detail
Proxies in scoring. A lead score that gives weight to the caller's area code, ZIP code, email domain or name can track race, national origin or religion closely in some markets. The safest score uses only what the person said about their plans: timeline, financing status, representation and the property they asked about.
Uneven accuracy. Speech recognition and summarization can be less accurate for callers with some accents, for non-native speakers or for people with speech disabilities. If their summaries are more often wrong, their callbacks start from worse information and the agent may judge the lead less promising. Measure accuracy by group rather than on average.
Inconsistent answers. An AI that answers questions about neighborhoods may give different responses to different phrasings, and phrasing often reflects who is asking. The fix is to remove the topic: no characterizations of areas, schools or safety for anyone, with a pointer to public data sources instead.
Steering in suggestions. If an assistant recommends other listings, it must base suggestions only on stated criteria such as price, size and location the buyer named, never on inferences about who the buyer is or who lives nearby.
Speed and routing. Routing rules that send leads from some sources or areas to a slower queue can produce unequal service even when each rule looks neutral. Monitor response times and appointment rates by area and source.
Channel gaps. People reach you differently. Pew Research found Americans aged 18 to 29 are more likely than older adults to take calls from unknown numbers[4], and 16% of adults rely on a smartphone for internet access with no home broadband[5]. A process that answers web forms within minutes but returns voicemails the next day treats groups differently through channel alone.
Do not forget the top of the funnel
Bias can enter before a lead ever calls. AI-assisted ad tools choose who sees your listings and your recruiting posts, and AI writing tools draft the copy. If the audience is narrowed by characteristics that track protected groups, or the copy describes an ideal buyer, the problem starts long before the phone rings.
- Target housing ads by geography and property criteria you can justify, not by audience traits such as age, family status or interests that stand in for them.
- Review where your leads come from each quarter. If a whole part of your market never produces inquiries, ask whether your advertising reaches it.
- Check AI-drafted ad copy and listing remarks for phrases that describe people rather than the property.
- Keep a copy of each housing ad and its targeting settings, so you can show what ran and to whom.
The same principle runs through every stage: decide on the basis of what the property is and what the person asked for, and check the results to make sure that is what actually happened.
Why caution is warranted
The public is wary of AI risks. In Pew Research's 2025 survey, 57% of Americans rated the societal risks of AI as high, compared with 25% who rated its benefits as high[6]. A fair housing complaint involving an AI tool would land in that climate, and the brokerage, not the software vendor, is the one with the license and the relationship with the client.
Source: Pew Research Center, 2025[6]
Show the data
| Item | Value |
|---|---|
| Rate the risks as high | 57% |
| Rate the benefits as high | 25% |
Governance is often missing. Gallup found that while 44% of US employees say their organization has begun integrating AI, only 22% say it has communicated a clear plan[7]. For a brokerage, a clear plan includes written fair housing rules for every AI tool that touches leads and a named person who checks them.
A monthly audit you can run
Paired test calls and chats
Write pairs of identical inquiries that differ only in one signal, such as a name, an accent, a mention of children or a mention of a wheelchair. Run both and compare the answers, questions asked, summaries and next steps.
Outcome comparison
From your CRM, compare time to first response, share of leads with a booked appointment and share handed to an agent, by ZIP code, lead source and language.
Summary accuracy sample
Check ten summaries against recordings for callers whose first language is not English, and ten for callers whose first language is English. Compare error rates.
Rule review
List every input used in scoring and routing. Remove anything that could act as a proxy for a protected characteristic.
Answer review
Search transcripts for words such as safe, family, church, school district and neighborhood to find any characterization that slipped through.
Record and fix
Write down what you found, what you changed and when you will retest.
Questions to ask AI vendors
- What inputs does any lead score or priority use, and can we see and turn off each one?
- How is speech recognition accuracy measured across accents and languages, and can we review our own results?
- How do we stop the assistant from describing neighborhoods, schools or safety, and how is that tested?
- Can we export transcripts, summaries and routing decisions for our own audits?
- What changes in the product when the vendor updates its models, and how are we told?
Training and accountability
REALTORS must complete at least two hours of Fair Housing training in each three-year cycle, from January 1, 2025 through December 31, 2027[3], and those who miss it face suspension and then termination of membership[8]. Use that training to walk the team through how your AI tools answer, score and route leads, and what agents should do when they notice something that looks wrong.
Name one person, usually the broker or compliance lead, who owns the AI fair housing rules, reviews the monthly audit and decides on changes. Agents should know they can raise a concern with that person without it becoming a disciplinary matter.
Example: what an audit caught
Example (fictional): a brokerage ran paired test calls in which one caller mentioned three young children and the other did not. The answers about the listing were identical, but the assistant suggested a different nearby listing to the caller with children, one with a larger yard. The suggestion came from a rule that offered alternatives based on details mentioned in the call.
Example result: the broker changed the rule so alternatives are offered only from criteria the caller explicitly states as search requirements, such as price range, bedrooms and area, and reran the paired calls. Both callers then received the same suggestions. The change and the retest were recorded in the audit log.
Frequently asked questions
Can AI tools violate fair housing law?
The brokerage using them can, if the tools produce unequal treatment based on protected characteristics, even unintentionally. That is why rules, testing and outcome monitoring matter.
What is paired testing for AI?
Running identical inquiries that differ only in one signal, such as a name or a mention of children, and comparing the answers, questions, summaries and next steps.
Should our AI answer questions about neighborhoods?
It should describe the property and factual amenities you have verified, but not characterize areas by who lives there, safety or school quality. Point callers to public sources.
Who is responsible if a vendor's AI discriminates?
The brokerage remains responsible for how it treats consumers. Vendor contracts and audits help, but they do not transfer that responsibility.
References
U.S. Department of Housing and Urban Development (2026). Housing Discrimination Under the Fair Housing Act (opens in a new tab) ↑
U.S. Department of Housing and Urban Development (2026). Housing Discrimination Under the Fair Housing Act (opens in a new tab) ↑
National Association of REALTORS® (2025). Fair Housing Training Requirement (opens in a new tab) ↑
Pew Research Center (2020). Most Americans don't answer cellphone calls from unknown numbers (opens in a new tab) ↑
Pew Research Center (2026). Americans' Use of Mobile Devices in 2026 (Mobile Fact Sheet) (opens in a new tab) ↑
Pew Research Center (2025). How Americans View AI and Its Impact on Human Abilities, Society (opens in a new tab) ↑
Gallup (2025). AI Use at Work Has Nearly Doubled in Two Years (opens in a new tab) ↑
National Association of REALTORS® (2025). Fair Housing Training Requirement (opens in a new tab) ↑