Lesson objective
Recognise the forms confabulation takes in property descriptions, rule explanations and citations, and flag each one for checking against the listing record or the official source before it reaches a client.
Plain-language explanation
Why this matters now
Module 1 showed why a language model can produce something that sounds right and is not. This lesson is about what that looks like in real estate work, because the failure has recognisable shapes. Once you can name them, you can spot them in a draft before a buyer, a seller or a regulator does.
NIST's Generative Artificial Intelligence Profile (NIST AI 600-1, released July 26, 2024) defines confabulation as "the production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications')" by which users may be misled or deceived. Section 2.2 says confabulations "are a natural result of the way generative models are designed," and that outputs "may also include confabulated logic or citations that purport to justify or explain the system's answer." The profile is voluntary, it is not written for real estate, and it does not measure how often this happens. Applying it to listings and rules is this lesson's analogy.
A real estate regulator has said the same thing in plain terms. The North Carolina Real Estate Commission's Tech Corner: What to Avoid When Using AI Tools (NCREC Bulletins, dated October 28, 2025 in the page metadata) warns that AI "can confidently provide incorrect market data, use outdated rules, or mislead you on legal interpretations," and that licensees should verify every fact before sharing it. It warns that AI-written listing descriptions may include language that violates fair housing laws or misrepresents features, and must be reviewed for accuracy, compliance and tone before publishing. That is guidance to North Carolina brokers, not a rule, and other states' commissions may say different things; but its warning describes a failure every licensee can meet.
This course holds no peer-reviewed study measuring how often AI-generated listing descriptions contain errors. No error rate appears here, and you should be wary of anyone who quotes one without a source.
A question to sit with
An assistant tells you a buyer "can cancel within the state's statutory three-day review period" and names a code section. The answer is confident, specific and formatted like a citation. What would you have to do before that sentence could go in an email to your client?
What goes wrong
Invented features. A draft adds a fireplace, a finished basement, "updated systems" or a view that the record never mentions, because those phrases often appear in listing remarks.
Wrong measurements and counts. Square footage, lot size, bedroom counts and year built are rewritten, rounded, or swapped with another property's figures, especially when the model is given several listings at once.
Stale or invented location claims. School assignments, zoning, HOA rules, flood status and "walking distance" statements are produced from patterns, not from a current official source.
Confabulated rules. A model states a deadline, a disclosure requirement or a licensing rule that does not exist, belongs to another state, or has changed.
Confabulated citations and logic. The model names a statute section, a rule number or a commission bulletin that does not exist or does not say what it claims, and wraps it in step-by-step reasoning. NIST notes models "sometimes provide logical steps for how they arrived at an answer even when the answer itself is incorrect."
Fair housing wording nobody asked for. A draft describes an ideal buyer, the people in a neighbourhood or its "character" in terms that can indicate a preference or limitation. NCREC names this risk directly.
The practical answer
Treat every sentence in an AI draft as belonging to one of five types, and flag each type for a specific check. Module 6 builds the full verification workflow; this lesson teaches you to see what needs verifying.
| Type of content | What confabulation looks like | Red flags | Where it must be checked |
|---|---|---|---|
| Property feature | A fireplace, pool, upgrade or view not in the record | Generic "selling" phrases; features you did not supply | The MLS record and the seller's own documents |
| Measurement or count | 2,400 square feet instead of 2,040; four bedrooms instead of three | Round numbers; figures that differ from your input | The record, and whatever source the listing licensee used for the figure |
| Location claim | A named school, zoning district, flood status or HOA rule | Named institutions; "top-rated", "walkable", "safe" | The official source that owns the fact, not the model |
| Rule or deadline | A review period, disclosure duty or licensing requirement | A number of days; "state law requires" | The statute, rule or commission publication for your state |
| Citation or explanation | A code section, rule number or bulletin title | Precise-looking references; tidy step-by-step logic | Open the cited source and read the text yourself |
Why supplied records still produce invented content. Giving the model the record helps, but does not stop confabulation. Models blend what you supplied with patterns from training. Three situations raise the risk:
- Thin records. If the record has few remarks, the model fills the gaps with typical features.
- Several properties at once. Figures migrate between listings in the same prompt.
- Requests for "compelling" or "vivid" copy. The more you ask for persuasion, the more the model reaches for common selling phrases.
A three-pass read for any AI draft about a property:
- Fact pass. Underline every feature, number, name, date and rule. Each underline needs a source in the file, or it comes out.
- Citation pass. For every reference to a law, rule, bulletin or data source, open the source. If it does not exist or does not say that, delete the sentence and note that the tool confabulated it.
- Fair housing pass. Read for any wording about who should live there, who lives nearby, or the neighbourhood's character. Module 5 teaches the Fair Housing Act's provision on discriminatory statements and advertisements (42 U.S.C. 3604(c)) and HUD's advertising regulation (24 CFR 100.75), which bind nationally.
Remember this
- Confabulation has shapes in real estate: invented features, wrong measurements, stale location claims, invented rules, and citations that do not exist.
- Supplying the record reduces invention but does not stop it, especially with thin records or several properties in one prompt.
- A precise citation or neat step-by-step logic is not evidence; open the source.
- NCREC guidance: AI can use outdated rules and incorrect market data, and fair housing responsibility stays with the licensee.
- Read every draft three times: facts, citations, fair housing.
Real estate example
Larkfield Realty Group is a fictional brokerage. Agent Samuel Oyelaran asks an assistant to draft remarks and a short buyer FAQ for a new listing, 61 Thornbury Court, fictional Tri-County MLS number TC-62390. He pastes the record, which gives 3 bedrooms, 2.5 baths, 2,040 square feet of finished living area, a 0.18-acre lot, built 2003, a gas furnace, and the remark "HOA covers common areas only." In the same prompt he pastes a second listing, 9 Kestrel Way, with 4 bedrooms and 2,400 square feet, to "match its style." The draft for Thornbury Court says "four spacious bedrooms across 2,400 square feet" (figures migrated from the second listing), "a cozy gas fireplace" (not in the record), and "assigned to the award-winning Maple Ridge Elementary" (no school field was supplied). The buyer FAQ adds: "Under state law, buyers have a 10-day inspection right after contract, per the state real estate code." It names no source the agent can open. Samuel runs his three-pass read. Fact pass: he restores 3 bedrooms and 2,040 square feet from the record, deletes the fireplace, and removes the school claim because only the official district source could support it. Citation pass: he searches his own state commission's website and statutes and finds nothing matching the FAQ claim; the inspection period in his brokerage's transactions comes from the purchase contract terms, not from a statute, so he deletes the sentence. Fair housing pass: the draft's closing line, "perfect for a young professional couple," comes out. He notes each removal in the listing file, and his broker adds "one property per prompt" to the team's prompting notes.
Responsible-use warning
Do not send a client, publish or repeat a rule, deadline or citation that came from an AI tool until you have opened the official source and read the words that support it. A section number, a rule name or a tidy explanation is exactly what confabulation produces. The same goes for features, measurements and school or zoning claims: a detail you did not supply is a detail you must source or delete. NCREC's guidance is blunt that AI can confidently use outdated rules and incorrect market data.
Human-review requirement
The listing or working licensee owns every property representation and client communication, and checks each feature, measurement, location claim, rule and citation against the MLS record, the seller's documents or the official source before use. The licensee also reads every draft for fair housing wording. The sponsoring or designated broker, under the supervision duty their state defines, reviews the file notes of removed content, confirms the three-pass read is being done, and refers any question about what a rule actually requires to the commission's own publications or to an attorney.
Practical exercise
Practise spotting confabulation on fictional material only.
- Write a fictional MLS record for an invented property with at least twelve fields, including bedrooms, baths, finished square footage, lot size, year built, heating, parking and two remarks. Leave school and zoning fields out.
- Write a second fictional record for a different property with different figures.
- Paste both records into an approved tool and ask for "compelling" remarks for the first property in the style of the second.
- Underline every feature, number, name, date and rule in the output (fact pass). Mark each as supplied, migrated from the second record, or invented.
- Ask the tool one question about a rule, such as a disclosure deadline in your state. Record the answer and any citation it gives.
- Open the official source for that rule on your state commission's or legislature's website (citation pass). Record whether the citation exists and whether it says what the tool claimed.
- Read the remarks for wording about people, an ideal buyer or neighbourhood character (fair housing pass) and list anything that must change.
- Rewrite the remarks using only supplied facts.
- Write a three-line file note of what you removed and why.
Knowledge check
-
How does NIST AI 600-1 define confabulation?
Answer guidance: As the production of confidently stated but erroneous or false content, known colloquially as hallucinations or fabrications, by which users may be misled or deceived. NIST says it is a natural result of how generative models are designed.
-
Name three situations that raise the risk of invented content even when you supply the listing record.
Answer guidance: A thin record that leaves gaps for the model to fill; several properties in one prompt, so figures migrate between them; and requests for compelling or vivid copy, which pull in common selling phrases.
-
What does NCREC's Tech Corner say about AI and market data, rules and fair housing?
Answer guidance: That AI can confidently provide incorrect market data, use outdated rules or mislead on legal interpretations, that listing descriptions may contain fair housing violations or misrepresent features, and that fair housing responsibility always rests with the licensee. It is North Carolina guidance, not a rule.
-
What are the three passes in the read this lesson recommends?
Answer guidance: A fact pass that sources or removes every feature, number, name, date and rule; a citation pass that opens every referenced source; and a fair housing pass for wording about people, ideal buyers or neighbourhood character.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- North Carolina Real Estate Commission — Tech Corner: What to Avoid When Using AI Tools (NCREC Bulletins) (2025-10-28). Accessed 2026-09-15.
- National Association of REALTORS® — Handbook on Multiple Listing Policy: C. Model Rules and Regulations for an MLS Operated as a Committee of an Association of REALTORS® (2026-01-01). Accessed 2026-09-15.