Will Your AI Assistant Make Up a Number in Front of a Client? Anti-Fabrication, Explained
An AI that confidently invents a sale price, an RCW citation, or a closing deadline is worse than no AI at all. Here is how fabrication happens, and how to tell a grounded tool from a guessing one before you repeat its answer to a client.
One fabricated fact is all it takes
You are on the phone with a seller. You say the last comparable sale closed at $812,000, that the inspection contingency runs 10 days, and that the excise tax on their sale comes to a specific dollar figure. If any one of those is a number your assistant invented, you do not get a do-over. The client either acts on bad information or catches the error — and once they catch one, they stop trusting all of it.
Real estate runs on facts that are checkable: a parcel's last sale price, a Form 17 delivery window, a graduated REET bracket. A wrong comp can blow a list-price strategy. A wrong deadline can blow a contingency. A wrong statute citation can put your license in the conversation. The cost of a fabricated answer is not measured in awkwardness — it is measured in deals and in trust that does not come back.
This is why anti-fabrication is not a nice-to-have feature of an AI assistant. For a licensed agent, it is the whole ballgame. A tool that is right 95% of the time and silently wrong the other 5% is a liability you carry into every client conversation.
How chatbots drift into confident wrong answers
A general-purpose language model is a prediction engine. Ask it for the last sale price of a specific address and it will produce a plausible-looking number — formatted correctly, in the right range, stated with total confidence — whether or not it has ever seen that record. This is what people mean by AI hallucination in real estate: the output looks authoritative and is simply made up.
The failure is sneakiest on exactly the questions agents ask most. Property specifics (beds, baths, square footage, owner tenure), market stats (median, days-on-market, months of supply), contract deadlines, and RCW or WAC citations are all things a model can fake fluently. It does not know it is guessing, so it does not warn you. And financial math is a known weak spot — graduated brackets, business-day counting, and net-sheet arithmetic are precisely where an LLM is most likely to be confidently off.
Worse, an unconstrained bot will paper over its own gaps. Instead of admitting it lacks data, it fills the hole with something. The more natural and assured the writing, the easier it is to repeat the made-up part to a client without a second thought.
The anti-fabrication approach: ground every claim in a tool
The fix is architectural, not cosmetic. Instead of letting the model answer property and market questions from memory, you wire it to call real tools — county assessor records for property data, market feeds for stats, a deterministic calculator for tax and deadlines — and you let it speak only from what those tools return. If the data is not there, the tool returns nothing, and the assistant says so rather than inventing a value.
Deterministic calculations should run outside the language model. SENTINEL routes supported REET and seller-net calculations to code, and calculates transaction dates from the terms the user enters. Rates, form language, and contract inputs can change, so the source, effective version, and executed documents still need verification.
Sourced claims are the other half. A grounded answer identifies the assessor record, market observation, or legal source it used and states the effective period where available. If a retrieved fact cannot be traced to a source, treat it as unverified.
"I don't have that — here's how to get it" is a feature
The instinct is to want an assistant that always has an answer. The opposite is what protects you. An assistant that says "I do not have a recorded sale for that parcel — want me to pull the assessor record a different way?" has just saved you from repeating a guess to a client. That plain admission is the single most valuable behavior a real estate AI can have.
Compare two replies to the same dead-end. One invents a number and states it confidently. The other says it does not have the data and offers the next step. The first feels smoother in the moment and can quietly cost you a relationship. The second feels slightly less magical and keeps you safe. A tool built for licensed agents should default to the second every time, and it should never dress up a missing-data situation as a fake system error to avoid admitting the gap.
So treat "I don't have that" as a green flag when you are evaluating a tool. A vendor whose demo never once says it lacks data is not showing you a smarter product — they are showing you one that hides its uncertainty, which is exactly the failure mode that burns you in front of a client.
What to test before you trust a bot on property data
Test it on questions where you already know the truth. Pull up a parcel you know cold and ask for its last sale price and square footage. Ask for the median and days-on-market in a ZIP you farm, then check the source and effective period instead of accepting a remembered figure. Ask it to compute REET on a known sale price and verify the brackets by hand. If it handles the facts you can check, you have a reason to trust its process on the ones you cannot.
Probe the edges on purpose. Ask about an address that does not exist, a market with no data, or a statute you made up. The right answer is a clean "I do not have that." A wrong answer — a confident, fabricated reply to an impossible question — tells you everything about how it will behave on the real ones. Then ask any contract-deadline or RCW question and confirm it cites a source you can open, rather than asserting from thin air.
If you want to see how a tool should reason about its own data and limits before you rely on it, SENTINEL publishes its full methodology — sources, math, and known gaps — so you can check the line between what is looked up and what is computed. That same discipline is what separates a sourced, checkable answer from a confident guess.
See the assistant work · no login
Bring Sentinel the real question behind this article.
Start with a prepared Western Washington example, then replace it with the work on your desk. Sentinel keeps the geography, source dates, assumptions, and verification boundaries visible.
Build and inspect the question publicly. Starting a private workspace afterward is optional.
Frequently asked
- What is AI hallucination in real estate?
- It is when an AI assistant produces a confident, plausible-sounding answer — a sale price, a market stat, a contract deadline, or a statute citation — that it actually made up rather than pulled from a real source. The danger is that the output looks authoritative, so an agent can repeat a fabricated number to a client without realizing it was never grounded in data.
- How do I know if my AI assistant is making up property data?
- Test it on facts you can already verify. Ask for the last sale price or square footage of a parcel you know, then check the answer against the county assessor record. Ask it about an address or statute that does not exist — a trustworthy tool says it has no data, while a fabricating one invents a confident reply. If every claim it makes traces back to a source you can open, it is grounded; if it cannot show its sources, treat its numbers as unverified.
- Why are AI tools bad at real estate math like excise tax and deadlines?
- Language models predict text, they do not calculate, so graduated brackets and business-day counting are exactly where they go wrong. Washington's REET uses tiered state rates plus a local portion, and contract deadlines follow specific rules — for example, periods of five days or fewer skip weekends and state holidays. The reliable approach is to run that math in deterministic code and use the AI only to explain the result, never to compute it.
- Is it a problem if an AI assistant says it doesn't know something?
- No — it is a sign the tool is built correctly. An assistant that admits it lacks a record and offers a way to get it has just protected you from repeating a guess to a client. The real problem is the opposite: a tool that always has an answer is likely hiding its uncertainty behind fabricated confidence, which is the failure mode that costs agents trust.
- Can I rely on an AI assistant for RCW citations and contract deadlines?
- Only if it sources them. A grounded tool will cite the specific statute or form and let you verify it, rather than asserting a citation from memory. Treat any legal, contract, or deadline output as a draft to confirm with your broker or a Washington-licensed attorney before you advise a client — verification is the standard, not an optional extra.