Artificial intelligence (AI) has moved from novelty to near-default workflow tool in real estate marketing — and the stakes for getting it right are rising.
Two industry voices who spend their days inside the AI arena say the technology is only as useful as the human reviewing it.
Nathan Brannen, chief product officer at Restb.ai, and Troy Palmquist, real estate advisor and founder of HomeCode Reviews, each spoke with HousingWire about where AI tools help, where they fail and how agents can protect themselves and their clients.
Brannen’s framing is blunt in regard to AI-assisted listing accountability; the agent owns the outcome, not the software.
“The agent should be responsible for what they put,” he said. “Whether it’s something they put in the listing or whether it’s something they put in the remarks. The idea of these tools is for them to save time, and the amount of time it takes an agent to review what’s been pre-populated or pre-generated by the AI is going to be quicker than it is for them to do it from scratch.
“I think, fortunately for humans, we’re pretty good at flagging the thing that isn’t right.”
Per the National Association of Realtors’ 2026 Technology Report, 48% of all Realtors use AI regularly — 23% daily and 25% weekly. Only 12 percent say they have no plans to use it, down from 32% in 2025.
OpenAI’s ChatGPT dominates, used by 91% of agents who deploy AI, followed by Google Gemini at 42%, Microsoft Copilot at 27% and Anthropic’s Claude at 25%.
Listing descriptions are the single most common use case, at 75%, followed by social media copy at 56%, client correspondence at 52% and market reports at 30%.
Palmquist said the quality of output tracks the quality of input.
“The more you put into it, the better the output is,” he said. “I think what happens is a lot of people just put the property address in and they hit enter, and it’s obviously pulling from previous listing descriptions or what it knows about an area, but it’s never really writing what’s special about the house — what’s going to attract a prospective buyer.”
Where AI goes wrong
Brannen said the most common errors stem from photos that do not actually depict the property.
“You could have a finished bedroom with full-size windows in a basement, and it would technically not be appropriate to call that, to add that to the bedroom count, because it would be a below-grade bedroom,” he said. “But based on that photo, there may be no way that you can actually tell that it is in the basement.”
He also warned that AI models fill gaps with invented detail when direction is unclear.
“If it’s not kind of given clear direction on something, it may make something up that fits in with what the data that it does see,” Branned said. “That may or may not be incorrect. It can be a little bit tricky because depending on how it’s generated — different flexibility or freedom that it may take.”
On the visual side, Brannen flagged photos of shared amenities or surrounding areas — a mountain view from the best vantage point in town, rather than the backyard — as frequent sources of misrepresentation.
Palmquist sees the same hallucination problem bleeding into imagery.
“When you go into ChatGPT and ask it to stage a photo, it’s not doing it correctly,” he said. “It starts adding additional canned lighting. It changes the transition from hardwood floors to carpet and runs different flooring. It starts hallucinating with the photos.
“You can go into these AIs and start prompting it to create a video, but again, you run into the hallucinations of it adding a pot filler behind the stove. Now you get a fake pot filler there. The buyer is going to expect that to be there.”
Words that trigger liability
Palmquist said the trouble can often start with simple adjectives.
“It’s stuff like ‘quiet’ or ‘family friendly,’” Palmquist said. “I think there are a lot of words that are subjective to the reader and the writer, and how quiet is definitive to different things. You could say that a beach is quiet, but then you get the foghorn that goes off, and the inclement weather and you get the seals that are barking. Is that really quiet? It’s subjective to the author and to the writer. AI is never going to know that.”
Fair housing and trademark traps compound the problem, he added.
“You really shouldn’t say walking distance or [a certain number of] steps too, because it’s a gray area if that’s a fair housing violation for handicap and accessibility,” Palmquist said. Even prompting the model to avoid violations does not solve it. “Even saying something like Caesar Stone. Caesar Stone is a company, a trademark — versus saying that it’s a quartz countertop.”
Who pays when a listing is inaccurate?
Asked whether any precedent shields an agent who misses a false statement generated by AI, Palmquist was unambiguous.
“No one is going to go back and sue ChatGPT or Anthropic,” he said. “You’re going to go to the real estate agent, and then subsequently the broker would be responsible because it would fall under broker supervision purview.”
He said no specialized listing tool completely avoids the chance for inaccuracy.
“If somebody buys something and it says that the house is on sewer and it turns out the house is on septic, the buyer is going to sue the seller and sue the agent,” said Palmquist. “The buyer is going to sue the broker. Now where did the AI come back and say, ‘Oh, it’s not on sewer. It turns out it’s really on septic?’”
Brannen further stressed the notion of AI output matching the work ethic of its human user.
“The larger issues come from people just being lazy and not looking over what’s been done,” he said. “I’ve heard the idea of being a curator of a listing rather than being creator of a listing, with the thought of it being a lot more efficient when you’re just focused on making the listing look great — without having to manually type everything in.”
The bottom line
Asked for overarching guidelines that any agent can adhere to, Brannen offered the following.
“You would be upset if you got to a property and things weren’t as they seemed [in the listing],” he said. “There are more and more ways that different platforms are allowing you to disclose. You can say that action X has been done on this image — whether it’s been decluttered or the sky has been replaced because when they went and got the photos, it was overcast. If you hesitate, you probably shouldn’t be doing that.
“Things like removing power lines or making a sunset in the direction it doesn’t set are things where there’s a large agreement that they shouldn’t be done, but inevitably, they’re still happening with some agents.”
Palmquist questioned whether time savings offered by many AI tools are truly beneficial for their users.
“I think around 84% of the industry hasn’t done a transaction yet in 2026,” he said. “Yet we’re marketing these tools like ChatGPT and these other things like we’re going to get you time back. Well, if 84% of the industry hasn’t done a transaction, why do they need more time back?”
Agents who treat AI as a drafting assistant — and themselves as the curator — capture the efficiency without the liability. Those who skip the review risk fines, lawsuits and damaged reputations that follow the human broker, not the technology.
