Ursino: AI will agree with you right up to the complaint

The views expressed in this column are solely those of the author.

A few weeks ago I watched an argument break out in a large Facebook group for agents. The question was an ordinary one, the kind that comes up on any given Tuesday: can you show a property to a buyer who hasn’t signed with you yet?

One agent answered with total confidence. Under TRESA, she wrote, you cannot provide services to a self-represented party, so showing that buyer a home creates a relationship you are not allowed to have. And she had proof. She had pasted in screenshots of a Google AI answer, complete with official-looking citations to RECO bulletins and law firms. She even mentioned she had phoned RECO and they agreed with her. Her closing line, more or less: if you think the AI is wrong, go educate yourself.

Then something useful happened. Experienced agents, a couple of instructors and a real estate lawyer showed up in the comments and pushed back. The lawyer made the point that stuck with me: the real answer to this question has not been settled by the courts yet. It is genuinely unsettled. And here was an AI, and an agent quoting it, handing the question down as closed-case fact.

‘AI is a genuinely great tool. It is also a sycophant.’

I am not writing this to relitigate that debate. I said my piece in the thread, and I am a managing broker, so I have a view. But my view is not the lesson. How we got there is the lesson.

Because a version of this lands in my inbox every few weeks. An agent sends me a compliance question buried in a thousand-word email, obviously written by AI, confident, nicely formatted and wrong. Nobody has time to read to the bottom of a wrong answer. But the agent has already made up their mind, because the machine sounded so sure.

That is the part we are not talking about. AI is a genuinely great tool. It is also a sycophant.

The problem was never that AI is sometimes wrong. Every source is sometimes wrong. The problem is that AI is wrong with total confidence, and that it is built, underneath everything, to agree with you.

‘You laundered your assumption through a machine built to flatter, and you came out the other side more certain than you went in. That is not research. That is outsourcing your thinking to something that has no judgment of its own, only a gift for sounding like it has plenty.’

Ask it, “Can I show a property to a buyer who has not signed with me?” and you will get one answer. Ask it “why am I allowed to show a property to a buyer before they sign,” and you will get the opposite, argued just as fluently, with citations just as convincing. The question you type already contains the answer you are hoping for. The model reads that hope and obliges. Then you ask a follow-up, and another, and with each round it agrees with you a little harder, until you have talked yourself all the way into a position with a very smart-sounding assistant nodding along the entire time.

You did not research the question. You laundered your assumption through a machine built to flatter, and you came out the other side more certain than you went in. That is not research. That is outsourcing your thinking to something that has no judgment of its own, only a gift for sounding like it has plenty.

This matters everywhere, but it matters most where the stakes are highest. A listing description written by AI is easy to check, and a bad one costs you nothing but a rewrite. A confidently wrong reading of TRESA can cost you a deal, a client or your registration. The flattery is worst precisely where you can least afford it.

‘In a business where some of the rules are still being written, borrowed certainty is the most expensive thing you can carry into a transaction.’

So here is what the agents who actually use AI well do differently on these questions. They make it argue against them. They do not ask the machine to confirm what they already believe. They give it the full context, then ask it to build the strongest possible case for the other side. They treat the answer as the start of the work, not the end of it. And then they check it against something that cannot be flattered into agreeing: the primary source, the RECO pages, the act itself, or a person who has actually done the thing. Your broker of record is down the hall. That is what they are for.

The agent in that thread had the tool exactly backwards. She used AI to end the conversation. The professionals in the comments used their experience to keep it open. That is the whole difference, and it has nothing to do with how good anyone is at technology.

AI did not make that agent wrong. Plenty of us manage that on our own. What AI did was make her certain, and hand her citations to prove it. In a business where some of the rules are still being written, borrowed certainty is the most expensive thing you can carry into a transaction.

Ask sharper questions. Trust the answers less. And when what is on the line is your registration, go find a human who has been there.

The post Ursino: AI will agree with you right up to the complaint appeared first on REM.