Latest · Rules & compliance

Your AI Vendor Doesn't Answer for Fair Housing. You Do.

September 2, 2026

You are the responsible party for every word a model puts into your listing descriptions, whether you read it before it went live or not.

That's not a technicality. HUD enforces the Fair Housing Act against the party who published the ad, and the FTC has made clear in its guidance on AI that using a third-party tool doesn't shift responsibility for a deceptive or discriminatory outcome away from the business that deployed it. Your vendor's terms of service might disclaim their liability. That disclaimer does not touch yours.

So the practical question isn't whether AI-generated copy is risky. It's whether your review process can catch a violation at the volume you're now producing copy. Most teams built their review habits for a world where a leasing consultant wrote three listings a month. That habit doesn't survive a model that can write three hundred a day.

Here's where it actually goes wrong. Fair housing violations in listing copy are rarely the obvious ones anymore. Nobody's model is writing "no children" or "ideal for young professionals" verbatim. The failure mode is subtler: language that describes a "vibrant social scene" for every listing near a nightlife district, phrasing that consistently frames accessible units as suited to "anyone who needs a little extra help," or copy that describes a neighborhood as "up and coming" in a way that reads as coded steering when you look at which properties get that phrase and which don't. A model trained on a firm's existing listing archive will faithfully reproduce whatever bias already lived in that archive. It just does it at scale and without the hesitation a human copywriter might feel typing the same line for the fortieth time.

The volume problem means you can't review every listing line by line forever. You need a system, not a habit. Two things belong in it.

First, build a banned and flagged term list specific to your portfolio and update it. Not just the classic list of family-status and disability language. Add the softer phrases your properties actually use for age, culture, or neighborhood framing, and check that they're applied evenly across every listing rather than clustered on specific properties or unit types. A simple keyword scan across all live listings, run monthly, catches drift before it becomes a pattern.

Second, spot-check by protected-class proxy, not just by property. Pull a sample across your portfolio and sort by neighborhood demographics, unit accessibility features, and price point. If the copy for accessible units reads differently in tone or specificity than everything else, that's your signal, regardless of how any individual listing reads on its own.

Here's what I would not do: hand the review job to the same model that wrote the copy. Asking an LLM to audit its own output for fair housing risk is asking it to grade its own training data. It will miss the same blind spots it already has, because the audit and the generation are drawing from the same well. Use a human reviewer, or at minimum a separate tool with a different training source, for the actual fair housing check.

One more thing worth saying plainly: state and local fair housing rules vary, and some states have added protections beyond the federal list, including source of income or sexual orientation and gender identity in certain jurisdictions. A national listing template that's compliant in one state can be non-compliant in another. If you operate across state lines, your review process needs a state-by-state layer, not a single national checklist.

None of this means walking away from AI-generated copy. It means treating the output the way you'd treat copy from a new hire who's fast, cheap, and has never sat through fair housing training. You wouldn't publish that person's first draft without a second set of eyes. Don't publish the model's either.

This is not legal advice. Run your specific workflow, vendor terms, and state obligations past fair housing counsel before you act on any of it.

Source: