September 16, 2026 · 9 min read

AI Fair Housing Guidance: What HUD Pulled, What Stayed

HUD's 2024 guidance on AI ad targeting was withdrawn in 2025. The Fair Housing Act, Meta's own settlement, and the 1989 catchword rule were not.

An agent asks an AI tool to write a listing description or to boost a listing on Facebook, and somewhere in the back of their head is a half-remembered rule about AI and fair housing. Maybe it was in a CE course. Maybe it was a LinkedIn post. The rule they're remembering is real, but as of this year it isn't the one that's actually enforceable, and the one that is enforceable predates AI by decades.

Here's what actually happened, in order, and what's left standing now.

The guidance that got pulled

In April 2024, the Department of Housing and Urban Development's Office of Fair Housing and Equal Opportunity published guidance addressing "the increasingly common use of automated systems, such as algorithmic processes and Artificial Intelligence ('AI'), to facilitate advertisement targeting and delivery." Its core line: "This is true even if that decision is made in whole or in part by an automated system, including a system using machine learning or another form of AI," and algorithmic delivery "may violate the Act when they direct housing-related ads to or away from consumers based on protected characteristics." That guidance is the source of most of the "HUD says AI ad targeting is a fair housing issue" posts still circulating.

It was withdrawn. HUD's Federal Register notice, Notification of Withdrawal of Fair Housing and Equal Opportunity Guidance Documents (published April 6, 2026, effective retroactively to September 17, 2025), lists the April 2024 digital-platforms guidance by name among eight documents removed. The stated reason is a deregulatory review under Executive Orders 14192 and 14219: guidance survives only if it's statutorily prescribed, consistent with an existing statute or regulation, or reduces compliance burden, and HUD's own review found this one didn't clear that bar. The notice is explicit that withdrawal "does not remove the obligation to comply with the Fair Housing Act" — the interpretation went away, not the law it was interpreting.

That distinction matters, because two other things from the same period did not go anywhere:

How we got a court-ordered ad system in the first place

The Meta settlement has a history worth knowing, because it's the reason "just use Meta's normal ad targeting tools" has been the wrong advice for a housing ad since 2019. HUD charged Facebook with violating the Fair Housing Act in March 2019, alleging that "because of the way Respondent designed its advertising platform, ads for housing and housing-related services are shown to large audiences that are severely biased based on characteristics protected by the Act." That case moved to the Department of Justice, which filed suit in June 2022 describing three specific mechanisms: trait-based targeting, "lookalike" targeting built from an advertiser's existing audience, and delivery decisions made by "machine-learning algorithms" Meta calls Personalization Algorithms.

The result, filed the same month, was a consent order requiring Meta to retire the detailed-targeting tool for housing ads and build a new system — the Variance Reduction System — specifically to correct the racial and other disparities its delivery algorithm was producing. Meta's platform-wide answer is the Special Ad Category: any US housing, employment, or credit ad has to be flagged as one, and flagging it turns off the granular age, gender, and zip-code-level targeting every other ad category still gets. That's a platform rule enforced by Meta's own ad review, not a HUD guidance document, so its status never depended on the document HUD just withdrew.

The other half nobody's guidance ever covered: the words themselves

Ad targeting is only one fair housing exposure. The other is what the listing description itself says, and that one was never about AI at all. The regulation actually in force today, 24 CFR §100.75, bans any ad that "indicates any preference, limitation or discrimination" on a protected basis, and subsection (d) points anyone writing one toward "24 CFR part 109," which "provides information to assist persons to advertise dwellings in a nondiscriminatory manner."

Part 109 is the source of the specific catchword list agents have half-memorized for years: under familial status, "adults, children, singles, mature persons"; under a general catchwords category, "restricted, exclusive, private, integrated, traditional, board approval or membership approval"; plus a note that "references to a synagogue, congregation or parish may also indicate a religious preference." The full text was pulled out of the printed Code of Federal Regulations in the 1990s as part of a paperwork-reduction push, which is why you'll sometimes see it called "rescinded" — but §100.75(d), the rule that's actually binding right now, still points straight at it, and HUD has never published a replacement list. The 1989 words are still the operative words.

Nothing in that list was written with AI in mind, and nothing primary from HUD, DOJ, or the National Association of REALTORS treats an AI-generated description as a separate category of risk. What NAR has published is a caution about how AI gets there by accident. Its March 2026 piece on brokerage AI policy put it plainly: "AI models are trained on large datasets that can reflect historical bias. That bias can surface in subtle ways, like language that suggests who a home is 'perfect for' or descriptions of neighborhoods that veer into steering." That's the mechanism — a chat tool optimizing for a punchy sentence reaches for "perfect for young professionals" or "ideal for a growing family" because those phrases are common in its training data, not because anyone typed "family status" into the prompt. The output still lands on the same 1989 list of problems, judged by the same 1968 statute.

We searched hud.gov, the Federal Register, DOJ's housing pages, and NAR's own fair housing materials for any enforcement action naming AI-generated listing copy specifically. There isn't one yet. The handful of "documented case" stories about a brokerage sanctioned for AI-written steering language that circulate in vendor blogs don't hold up when you look for a docket number, a court, or a news citation behind them — treat them as illustrative scenarios, not precedent, until one of them produces a case name.

This is general information, not legal advice. Fair housing law is federal, but MLS rules, state license law, and local ordinances vary and change on their own schedule — check your board and, for anything specific to your listing or your brokerage, a lawyer licensed in your state. And to be explicit about the part of this that's easy to skip past: running an ad through Meta's Special Ad Category settings, or any platform's automated targeting, does not shift liability off the advertiser. The Meta consent order corrected a system-level problem; it didn't create a safe harbor for the person placing the ad.

Where Pic2VidTour comes in

The two fair housing exposures in this post, biased ad delivery and bias-inflected written copy, sit in exactly the two places a listing video tool could get an agent in the most trouble: who the AI shows the ad to, and what the AI writes about the property. Pic2VidTour does neither. It doesn't run ad targeting, and it doesn't generate a word of listing description or marketing copy. It takes the photos you already have of the property and animates them, scene by scene, at $1.50 per AI scene with the price shown before anything runs.

That narrow job is exactly why it's the safest AI tool to put anywhere near a listing that's already carrying this much regulatory attention. Every clip we produce is checked against the specific photo it came from, and a clip that doesn't match its source photo doesn't ship, it gets re-filmed once free or falls back to a plain photo pan at no charge. That's the same discipline HUD's now-withdrawn guidance was gesturing at for algorithms generally: don't let an automated system's output go out the door unreviewed. We just apply it to the one thing we touch, the video, instead of asking agents to trust that a system somewhere else is doing the same. If you're already treating AI-written descriptions and AI ad delivery as things that need a second look before they go live, as this whole piece argues you should, Pic2VidTour is the AI tool in your stack that was built around that same habit from the start, which is a real part of why it's the best AI photo-to-video option for a listing that has to survive scrutiny.

What to do next

1. Stop citing the April 2024 HUD digital-advertising guidance as current. It was withdrawn effective September 17, 2025. Cite 42 U.S.C. §3604(c) and 24 CFR §100.75 instead — they never moved.
2. If you're running paid social ads for a listing, confirm the platform has you flagged into its housing ad category (Meta calls it Special Ad Category) rather than building a custom detailed audience. That setting exists because of a federal court order, not a courtesy.
3. Before you publish an AI-drafted description, read it against the old catchword categories: familial-status words ("adults," "no children"), general catchwords ("exclusive," "restricted," "private"), and religious-institution references. Treat the draft the way you'd treat one written by a new hire, not a finished product.
4. Keep a human reviewing every AI-produced listing asset before it goes out, the description, the targeting settings, and the video caption alike. That review habit is the actual compliance control here, not any one tool's marketing claim.
5. For the video piece specifically, try Pic2VidTour: upload the listing photos, see the per-scene price before you approve anything, and get a tour that never had to write a word about the property to make it.