The LLM Can Write the Ad. It Can't Be the One Who Gets Sued

Language models make marketing creative fast and cheap, and housing and lending are among the worst places to let that speed run unchecked. Fair housing, lender advertising rules, and consent laws attach to the words. The fix is a designed gate with a human on it, not a better prompt.

A close portrait of a man with a mustache and dark hat fixing a hard, unblinking stare past the camera
The Good, the Bad and the Ugly (1966), dir. Sergio Leone

Ask a language model to write an ad for a home in a nice suburb and it will happily hand you a line like “Perfect for young families in a safe neighborhood.”

It sounds great. It is also the kind of sentence that can land a housing advertiser in front of a regulator. The model has read about fair housing. It was also trained to be helpful and fluent, and it does not reliably apply those rules to a nuanced line of copy on a busy Friday afternoon.

I use language models every day. I think they make marketing creative faster and better, and I would not want to run a real estate or lending program without them. I also think this is one of the verticals where letting the speed run unchecked is a mistake, and I want to lay out why.

The speed is real

Let me say the good part first, because it is genuinely good. A model can produce 40 variants of a headline in the time it used to take to argue about one. It can write a local angle for every neighborhood a client serves. It can rework a message for a first-time buyer, a move-up buyer, and a refinancer without anyone starting from a blank page. It makes iteration cheap, and cheap iteration is how you find out what actually works.

Teams that ignore this will lose ground to teams that do not. I am not arguing for fear. I am arguing for putting the speed somewhere safe.

Why this vertical is different

Most categories can let a model write ad copy and fix problems after the fact. A weak line gets low clicks and gets rewritten.

Housing and lending do not work that way, because the rules attach to the words and the targeting, and the exposure is real money. At a high level, here is what sits around this work.

Fair housing law restricts how you can describe a property or an audience in relation to protected classes. Language that implies a preference for or against certain kinds of buyers, even innocently, can be a violation. “Walking distance to churches” and “ideal for families” are the sort of phrases that read as harmless and are not always treated that way.

Lender advertising has its own regime. Once an ad states certain loan terms, additional disclosures can be triggered. Mortgage ads often have to carry licensing information, in the United States an NMLS number among other things, and exactly what is required depends on the advertiser, the product, and the state. A rate or payment mentioned casually in a headline can trigger obligations the writer never noticed, depending on the exact terms.

Consent rules govern calls and texts. A lead who filled out a form has not necessarily agreed to be texted about everything, and a model drafting a follow-up sequence has no idea which permissions exist for which person.

RESPA adds its own limits on what can be paid or offered in exchange for referrals, with room for real, documented services, and any program that touches it needs legal review. A cheerful incentive idea from a model can quietly cross a line.

I am not a lawyer, and this is not legal advice. Every one of those regimes has details, and they change. The point is the shape: the rules live in specific words and in who sees them, and a fluent sentence is not the same thing as a compliant one.

The model cannot own the risk

A language model has no license and cannot be held accountable the way an advertiser can. It has no memory of the update that changed a rule last quarter, and it cannot be sued, fined, or asked to explain itself in a letter.

The advertiser can be. That is the asymmetry that matters. The model produces the words, and the person or company that runs them owns the consequences. When something goes wrong, “the AI wrote it” does not move the responsibility anywhere.

So the question is not whether the model is smart enough. A very capable model can still produce a line like the safe-neighborhood sentence, because plain, ordinary phrasing is what it leans toward, and the subtle cases are the ones that slip through. The question is who is accountable for each line that goes live, and whether the system makes it hard to skip that accountability when you are in a hurry.

Design the gate, not the prompt

The tempting fix is a better prompt. Tell the model about fair housing, tell it to be careful, tell it to avoid protected-class language. That helps at the margin. It is not a control.

A prompt is a request. It does not stop a bad line from reaching the ad platform on a Friday afternoon when someone is trying to launch. What does is a gate that is built into the process, with a named person who owns it.

At the level of principle, here is how I think about it. The model drafts. A human who understands the rules approves. Some categories of copy never ship without that approval, no matter how routine they look. And the review is a step in the workflow with an owner, not a habit that depends on someone remembering. I will not describe the specific checks, tools, or lists I use. They are part of how we run accounts, and a competitor can build their own. What I will say is that the value is in the structure. If review is optional, it will be skipped in the exact weeks it matters most.

There is a cost. A real gate slows the fastest part of the process. You give back some of the speed the model bought you, and a client who loves the idea of instant creative sometimes has to be talked into the idea of a checkpoint. I think that trade is correct here, and I think a client who gets sued for a model-written ad would agree.

The platforms are a second fence

There is another layer that the writer does not control. The ad platforms themselves restrict how housing, credit, and employment ads can be targeted. They put these categories under special rules that remove some detailed targeting options, and the specifics differ by platform and change over time.

That changes what the copy can even be aimed at. A line written for a very specific audience may be impossible to deliver to that audience. And it means copy and targeting have to be reviewed together, because a perfectly fine sentence can become a problem once it is attached to the wrong audience.

I am cautious about specifics here because the rules shift. Platforms revise their policies, and regulators revise their guidance. A system that assumed last year’s rules is a liability. A system that treats the rules as something to recheck on a schedule is not. If you take one thing from this piece, make it that: the process has to catch changes in law and platform policy before the next campaign runs on the old rules.

Where I land

Use the model for what it is great at, which is speed, variety, and the first draft nobody wanted to write. Keep a human accountable for every word that goes live, and build that accountability into the system so it survives a busy week.

The model can write the ad. It cannot be the one who answers for it. Someone has to, and it should be someone who decided to, on purpose, before the first line was drafted.