Try it right now. Ask ChatGPT, Claude, or Gemini to "make this listing sound warm and inviting" and there's a good chance it reaches for "perfect for families" or "great for young professionals."

Both are protected-class language. Familial status and age are both covered under fair housing law. The model isn't malfunctioning — it was trained on decades of real estate marketing copy, a lot of which wasn't compliant either. It learned the pattern.

This is the part almost every "AI for agents" guide skips entirely, because it's not the fun part. But if you're a solo agent using AI to draft listing remarks, captions, or social posts without a second set of eyes, this is the actual risk, not the model getting a square footage wrong.

Three things that help:

1. Build a personal "banned phrase" list — family-oriented, perfect for [any group], safe neighborhood, walkable to [religious institution] — and paste it into every prompt as a constraint.

2. Ask the model to audit its own output afterward: "Review this listing description for fair housing violations and flag anything referencing a protected class." It catches a surprising amount when asked directly.

3. Never publish AI listing copy unread. That's true for everything AI drafts, but it's non-negotiable here.

I put together a full phrase table and an audit prompt for this in a guide for solo agents — happy to share specifics if useful, just didn't want to make this post an ad. The bigger point stands on its own: check what your AI is actually writing before it goes live.