NAR Rules and AI: What Brokerages Must Know Before Deploying

The Broker Owns the Output

An AI tool does not hold a real estate license. Your brokerage does.

Every listing description the model drafts, every answer the chat agent gives at eleven at night, every property summary it stitches together goes out under a licensed name and, for most of the firms reading this, under a REALTOR membership. That makes an AI rollout a supervision decision that happens to involve software, not a software decision that happens to touch supervision.

That is uncomfortable for the way most brokerages buy technology. The tool gets evaluated by whoever runs marketing or operations, on speed and price, and the compliance question arrives after the thing is already answering consumers. By then the disclosure language, the data flows, and the audit trail are all vendor defaults nobody chose.

This post walks the rules that actually apply, what each looks like when AI is the author, and the questions to ask before signing. It is general information rather than legal advice, and brokers should confirm current requirements with their association, their MLS, and their own counsel before deploying anything. Federal fair housing guidance in particular has been issued and then withdrawn inside the last two years, and more of it is under review, so verify the current status of anything cited here rather than trusting the date on this page.

Three Code of Ethics Articles Do the Heavy Lifting

The 2026 Code of Ethics and Standards of Practice contains no provision that names artificial intelligence. That is not a gap you can stand in. The Articles govern conduct and communications, and the Code does not care whether a human or a model produced the sentence. Three Articles carry most of the AI exposure.

Article 10: Equal Professional Service

Article 10 provides that REALTORS shall not deny equal professional services to any person for reasons of race, color, religion, sex, disability, familial status, national origin, sexual orientation, or gender identity. Standard of Practice 10-1 adds that in the sale or lease of a residence, REALTORS shall not volunteer information regarding the racial, religious, or ethnic composition of any neighborhood. Standard of Practice 10-3 prohibits advertising that indicates a preference, limitation, or discrimination based on those characteristics.

Now put a language model in the seat. A consumer asks a website chat agent which neighborhoods are "good for families like mine." A helpful model answers. It pulls school ratings and demographic-adjacent signals from whatever it was trained on, and produces a paragraph that reads as steering because it is steering.

The second failure is quieter. AI listing copy reaches for atmosphere, and atmosphere in housing copy is where "perfect for a young professional couple" and "safe, family-oriented street" come from. Those are the constructions Standard of Practice 10-3 addresses, and a model produces them fluently because they run through the marketing text it learned from.

Fair housing risk in AI deserves its own treatment, and we published one: fair housing compliant AI covers prompt design, blocked topics, and output review in detail.

Article 12: A True Picture

Article 12 requires that REALTORS be honest and truthful in their real estate communications and present a true picture in their advertising, marketing, and other representations. Several Standards of Practice extend that obligation to the internet specifically.

Standard of Practice 12-9 requires REALTOR firm websites to disclose the firm's name and the state or states of licensure in a reasonable and readily apparent manner. Standard of Practice 12-10 applies the true picture obligation to internet content, images, URLs, and domain names, and prohibits deceptive framing, manipulating listing and other content in ways that produce a misleading result, and deceptive use of metatags and keywords. Standard of Practice 12-12 addresses URLs and domain names that present less than a true picture, and Standard of Practice 12-13 limits members to credentials they are legitimately entitled to use.

The AI scenarios are concrete:

• A generative image tool "enhances" listing photos by removing a utility pole or furnishing an empty room, with no virtual staging label.
• An AI content engine spins out neighborhood pages at volume, keyword-dense and thin on verified fact, on a domain that carries no firm identification.
• A chat agent, asked about qualifications, invents a certification because the prompt told it to be persuasive.

None of those require bad intent. They require a fast tool and no review step. The honesty obligation travels with the output, which is worth reading alongside how AI home tours are presented to consumers.

Article 2: Pertinent Facts

Article 2 states that REALTORS shall avoid exaggeration, misrepresentation, or concealment of pertinent facts relating to the property or the transaction.

AI trips this in a specific way: summarization drops things. Ask a model to condense a disclosure packet, an HOA document set, or a long remarks field into three friendly sentences, and it drops whatever it scored as less important. Sometimes that is the flood history. Sometimes that is the special assessment. The consumer receives a confident summary with a hole in it, and nothing signals that material was omitted.

The other Article 2 pattern is confabulation. A model asked about square footage, permit status, or lot lines will often produce a plausible number rather than decline. Any AI surface that discusses property specifics needs hard retrieval from your system of record and an explicit refusal path when the fact is not there.

Disclosure: When the Consumer Must Know It Is Software

There is no single national rule that says "label your chatbot." There are overlapping obligations that add up to the same practical answer.

Start with Article 12. If a consumer reasonably believes they are conversing with a licensed professional and they are not, the interaction is not presenting a true picture. Add Standard of Practice 12-9's firm identification requirement, which means a consumer-facing agent should make the brokerage name visible, not just the friendly bot name.

Then add state law. California's Bolstering Online Transparency Act, in effect since July 2019, requires clear and conspicuous disclosure when a bot is used to incentivize a sale or transaction, which has generally been read to reach customer-facing service bots. Utah amended its AI Policy Act in 2025: general disclosure narrowed to situations where a consumer clearly and unambiguously asks whether they are dealing with AI, while individuals providing services in regulated occupations must prominently disclose generative AI use in high-risk interactions, verbally at the start of a verbal interaction and in writing before a written one. Other states have moved and will keep moving.

The design that satisfies the widest set of obligations is not complicated:

• Identify the agent as automated at the opening of the conversation, not in a footer.
• Name the brokerage and the state or states of licensure in the same view.
• Answer honestly and immediately when someone asks whether they are talking to a person.
• Give a visible path to a licensed human, and hand off on request without friction.
• Never let the agent claim a license, a designation, or fiduciary representation it does not have.

Our own approach to these defaults is documented on the AI agents for real estate page and in our published AI policy.

MLS Rules and What You May Feed a Third Party Model

This is where AI tools most often create a violation without anyone noticing, because it happens in a data pipeline rather than in front of a consumer.

NAR's IDX policy is the starting point. Its provisions include that MLS participants may not use IDX-provided listings for any purpose other than IDX display. Downloads and the displays fed by them must refresh at least every twelve hours. Modifying or manipulating information relating to other participants' listings is prohibited. Where a display mixes MLS data with outside property information, that non-MLS data must be clearly separated and its sources identified in immediate proximity.

Read that against a typical AI feature list. Building an embedding index of listing content for a "search anything" assistant is a use other than display. Rewriting another broker's remarks into cleaner marketing copy is modification of another participant's listing content. Blending an automated valuation estimate into the same card as MLS fields, unlabeled, collides with the separation and attribution requirement. Sending listing text to a vendor whose terms allow training on customer content raises a licensing question NAR has been active on, given its position that copyrighted real estate content including listings, photos, and MLS data should not be used for AI training without authorization.

Local MLS rules may be stricter than the national floor, and they are the rules that fine you. Two further points to check:

1. Under NAR's Multiple Listing Options for Sellers policy, effective March 25, 2025 with implementation required by September 30, 2025, sellers may choose a delayed marketing exempt listing. Those listings are filed with the MLS and available to participants, but are restricted from IDX display and from syndication for the stated period, with the seller signing a disclosure consenting to that. Any AI system that syndicates, republishes, or emails listing content has to honor that status flag, and has to be tested against it.
2. If your AI vendor is a third party rather than your MLS data licensee, confirm in writing who holds the license, what the permitted purposes are, and what happens to the data at termination.

The IDX solution page and the security page describe the controls we hold ourselves to on listing data flow.

Written Buyer Agreements and the Current Practice Rules

NAR's practice changes took effect August 17, 2024, and the settlement received final court approval in November 2024. Two elements matter for AI deployment.

First, MLS participants working with a buyer must enter into a written agreement with that buyer before touring a home. The agreement must specify compensation in an amount that is objectively ascertainable rather than open-ended, must prohibit the broker from receiving more than that amount, and must include a statement that broker fees are not set by law and are fully negotiable.

Second, offers of compensation are no longer communicated through the MLS. Compensation remains negotiable off-MLS.

For an AI deployment this is a workflow constraint, not a talking point. Any assistant that books showings needs a gate before the tour, not a note in a follow-up email. Any assistant that fields compensation questions should route to a licensed human rather than characterize terms, because a model speculating about who pays what is generating a representation about the transaction. Any AI-written nurture sequence needs its compensation language checked against the firm's current buyer agreement, since older marketing copy in the training data reflects the prior environment. Configuring those gates is a CRM and workflow question as much as an AI question.

Accountability When It Goes Wrong

Three layers determine who absorbs the problem.

Broker supervision. State license law puts supervision of licensed activity on the broker. A model that drafts public marketing or converses with consumers is performing activity the broker is responsible for reviewing: a named reviewer, a defined standard for what gets checked before publication, and evidence the review happened. The fair housing side works the same way, and it does not run on guidance. HUD withdrew its 2024 guidance on advertising through digital platforms, which addressed algorithmic and AI-driven ad targeting and delivery, by a memorandum dated September 17, 2025, and published notice of that withdrawal in the Federal Register on April 6, 2026. A separate FHEO memorandum dated September 16, 2025 pulled a list of items out of HUD's guidance repository, including the Department of Justice press release on AI discrimination in tenant screening and the United States statement of interest in the SafeRent tenant screening algorithm case. The statute did not move, and neither did the regulations: 42 U.S.C. 3604 is unchanged, and 24 CFR 100.75 and 24 CFR 100.500 are both still in the Code of Federal Regulations. The published notice states that conduct not complying with the text of the Fair Housing Act remains subject to enforcement by the Department, and that a complainant may file a civil action in federal or state court within two years of the alleged discriminatory housing practice. The September 2025 memorandum says FHEO will deprioritize enforcement against parties whose conduct does not conform to the withdrawn guidance while the withdrawal is pending. Both are HUD's own words. Sixteen states and the District of Columbia are challenging the September 2025 guidance in State of Illinois v. HUD, number 3:26-cv-02262 in the Northern District of California; that case is pending and no injunction has been granted, so the withdrawals are operative today. Liability follows the discriminatory outcome, not the existence of an explanatory document. Our fair housing compliant AI post covers the current status and the controls in more detail.

Vendor contracts. Read for four things: whether your content and consumer data can be used to train models, who owns the output, what the indemnity actually covers (most exclude the customer's own use of the output), and what logs you can obtain if a complaint is filed. A vendor that cannot produce a conversation transcript on request has made your defense harder.

Audit trails. If you cannot reconstruct what the AI said, to whom, on what date, from which prompt and model version, you cannot defend the interaction and you cannot fix the pattern. Retention should match your state's record retention period for transaction records, with access limited to people who need it.

Pre-Deployment Checklist

1. Write the AI policy before the first tool goes live, and name an owner. Publish it alongside your other policies.
2. Inventory every AI feature already in the stack, including ones bundled into tools bought for other reasons. Most firms find more than they expected.
3. Classify data. Decide what may never be pasted into a model: client financials, disclosure packets, anything under a confidentiality obligation.
4. Confirm MLS and IDX permissions in writing for every system touching listing content, including delayed marketing exempt status handling.
5. Set disclosure defaults: automated identification at the opening, firm name and states of licensure visible, honest answer when asked, human handoff on request.
6. Constrain the model. Blocked topics for neighborhood and demographic questions, retrieval from your system of record for property facts, refusal when the fact is not available.
7. Require human review before anything carrying a licensee's name is published, and record that the review happened.
8. Turn on logging with retention matching your record retention obligations, and test that you can export a single conversation on demand.
9. Train the agents, not just the admins, and keep the completion records. Our training programs exist for this.
10. Schedule a quarterly output review that samples real conversations and real listing copy, and route findings back into the prompts and the policy.

Questions to Put to Any AI Vendor

Where to Start

Deploy AI the way you would onboard a new unlicensed assistant: define what they may say, what they may never touch, who reviews their work, and how you would reconstruct a conversation if someone complained. The tooling question follows the supervision question, not the other way around.

If you are scoping a rollout, the brokerage solutions overview and the full solutions catalog outline how the pieces fit together, and our AI policy documents the defaults we ship. When you are ready to review your specific data flows and MLS obligations, get in touch.

Frequently Asked Questions

Does the NAR Code of Ethics have a rule that specifically covers artificial intelligence?

The 2026 Code of Ethics and Standards of Practice does not contain a Standard of Practice that names artificial intelligence or automated tools by that term. The existing Articles still apply to anything published under a REALTOR's name, including copy and answers produced by software. NAR has published general guidance on AI use covering accuracy, privacy, fair housing, and transparency, and brokers should confirm current requirements with their association and counsel.

Can a brokerage feed MLS listing data into a third party AI tool?

Only if the MLS rules and the data license permit it. NAR's IDX policy states that MLS participants may not use IDX-provided listings for any purpose other than IDX display, which is narrower than most AI product roadmaps assume. Before any listing content reaches a vendor's system, check the specific data license, the MLS rules, and whether the vendor uses customer content for model training.

Who is responsible when an AI chat agent gives a fair housing problem answer?

The licensee and the supervising broker carry the exposure, not the software vendor, and that is true whether or not federal guidance on the point exists. HUD's 2024 guidance on advertising through digital platforms, which covered algorithmic and AI-driven advertising, was withdrawn by a HUD memorandum dated September 17, 2025, and HUD published notice of that withdrawal in the Federal Register on April 6, 2026. The Fair Housing Act itself did not change, and neither did 24 CFR 100.75. That published notice states that conduct not complying with the text of the Act remains subject to enforcement by the Department, and that complainants may still bring private civil actions. The September 2025 memorandum separately says FHEO will deprioritize enforcement against parties whose conduct does not conform to the withdrawn guidance while the withdrawal is pending, and neither statement addresses AI. Vendor contracts can allocate cost between the parties, but they do not move license law, the Fair Housing Act, or Code of Ethics accountability off the brokerage.

How do written buyer agreements change what an AI assistant is allowed to do?

AI tools that schedule tours or route buyer leads have to respect the requirement that MLS participants working with a buyer enter a written agreement before touring a home, which took effect with NAR's August 17, 2024 practice changes. That means the tour booking path in your CRM or chat agent needs a checkpoint, not a straight line from inquiry to showing. It also means AI-written copy should never describe buyer representation compensation in terms that contradict the signed agreement.

What should a brokerage AI policy actually contain?

At minimum: the list of approved tools, the data classes that may never be pasted into a model, a named human reviewer for anything published, the standing disclosure language for consumer-facing agents, log retention rules, and an escalation path when output goes wrong. Keep it short enough that agents read it and specific enough that a compliance reviewer can test against it. Pair it with training records so you can show the policy was communicated, not just written.