Lesson objective

Refine an AI-drafted listing, buyer or transaction document through controlled revisions that keep the fact table intact and trigger re-verification after every change.

Plain-language explanation

Why this matters now

First drafts are rarely final. The seller wants the garden mentioned first, the team leader wants 50 fewer words, the buyer wants the tour sheet in a different order. Each request sends the draft back to the model, and each round is a chance for a verified fact to change without anyone noticing.

The OWASP project describes the stakes well. Best practice (voluntary framework): in the OWASP Top 10 for LLM Applications 2026 (dated August 4, 2026), the project leads note that misinformation was placed near the bottom of the list by voters but near the top by the incident record, and write: "When a model's fluent, confident output drives a decision or a tool call, a wrong answer turns into a wrong action." OWASP's list is a voluntary community document, not a regulation. In a brokerage, the "action" is a published listing, a sent client update or a signed timeline.

Best practice (voluntary framework): NIST AI 600-1 (July 26, 2024) describes automation bias as "excessive deference to automated systems" that "can exacerbate other risks of GAI, such as risks of confabulation." By the third revision, a reviewer who has already checked the facts once tends to assume they are still right.

A question to sit with

You verified every fact in version 1. The model has since produced versions 2, 3 and 4 at your request. Which facts are you still sure of, and why?

What goes wrong

Silent drift. Asked to "make it punchier", the model changes "roof replaced in 2019" to "newer roof", and later to "new roof". Each step looked like style.

Compression that merges facts. Asked to cut 50 words, the model combines "two-car garage" and "detached workshop" into "two-car garage with workshop", which describes a different property.

New content during revision. Asked to "add a line about the neighbourhood", the model supplies features nobody gave it, or describes the people who live there.

Editing the wrong version. The agent revises version 3 while the approved text was version 2, and publishes a mix.

Re-verifying only the changed sentence. The model rewrote other sentences too, and the reviewer checked only the one they asked about.

The practical answer

Treat every revision as a new draft that must pass Stage 5 again. The loop below is an AI Coalition Network recommendation.

Step What you do What you check
1. Freeze the facts Keep the verified fact table from version 1 as the reference The table matches the sources
2. Ask for one change One instruction per round: order, length, tone or emphasis The instruction adds no new facts
3. Repeat the sources Send the sources and constraints again with the revision request Nothing is left to memory
4. Require a new fact table Ask the model to list every fact in the new version with its source line Table produced
5. Compare tables Line up the new table against the frozen one Any fact added, changed, merged or dropped
6. Re-run fair housing and advertising checks On the whole new version, not just the changed sentence Wording about people, value, identification
7. Label the version v1, v2, v3, with the instruction that produced it One approved version, clearly named

Edit by hand when the change is small. A reordered sentence or a trimmed adjective is often faster and safer to make yourself. Sending a small change back to the model invites it to rewrite everything else. Best practice: hand-edit small changes and re-check the edited sentence; regenerate only when the structure changes.

Ask for style changes that cannot touch facts. Phrase revision requests so the facts are explicitly fixed: "Reorder the paragraphs to open with the garden. Do not change, add, merge or remove any fact in the table."

When a new fact is genuinely needed, add it to the sources first, verify it, update the frozen table, then ask for the revision. Never let a revision request be the place a new fact enters.

Accuracy duties do not pause during revision. Professional rule (MLS participants and subscribers): NAR's model MLS rules, Section 1.2.0 (marked mandatory), require participants and subscribers to submit accurate listing data and to correct any known errors. A drifted fact published in version 4 is an error in the listing, whichever round introduced it. Your local MLS rule governs. Law / rule (Texas): 22 Tex. Admin. Code §535.155(d)(1) treats an advertisement inaccurate in any material fact as misleading; applying it to AI revisions is an application of the rule, and other states differ.

Remember this

  • Every revision is a new draft: re-send the sources, require a new fact table, compare it with the frozen one.
  • One change per round, phrased so facts cannot move.
  • Hand-edit small changes; regenerate only when structure changes.
  • New facts enter through the sources and verification, never through a revision request.
  • Re-run fair housing and advertising checks on the whole new version before release.

Real estate example

Copperline Properties is a fictional California brokerage. Listing agent Dana Whitaker verified the version 1 remarks for the fictional listing at 312 Quarry Lane: 1,880 square feet, three bedrooms, roof replaced in 2019, two-car attached garage, detached workshop, 0.18-acre lot, HOA dues of $140 a month. The sellers asked for three changes over two days: open with the garden, cut to 150 words, and "make it sound more exciting". Dana ran each change as a separate round, re-sending the property file and constraints each time, and asked for a new fact table with each version. Version 2 moved the garden up and changed nothing else. Version 3, the cut, merged the garage and workshop into "two-car garage with workshop" and dropped the HOA dues. Version 4, the "exciting" round, turned "roof replaced in 2019" into "brand-new roof" and added "a friendly community of young families". Dana compared each table with the frozen version 1 table. She restored the roof wording and the detached workshop by hand, decided to restore the HOA dues because buyers rely on them, and deleted the neighbourhood sentence as both unsupported and a fair housing flag. She then re-read the whole version 5 for fair housing and advertising problems, not just the lines she had fixed, and confirmed the approved identification block was in place. The sellers approved version 5, Dana approved it as the licensee, and her transaction AI use record names version 5 as the released text, lists the four rounds and notes the three restored facts.

Responsible-use warning

Do not re-check only the sentence you asked the model to change. Models often rewrite other sentences in the same round, and fluent revisions hide drift. Never let a revision request introduce a new fact; add it to the sources and verify it first. Do not publish from a file whose version you cannot name. If the fact table for the new version is missing, the revision has not been verified, however good it sounds.

Human-review requirement

The licensee who approved version 1 approves the final version as well, on the same standard. They compare the final fact table with the frozen table, confirm every added, changed, merged or dropped fact was resolved against a source, re-run the fair housing and advertising checks on the whole text, confirm the identification block, and record which version was released. Where a seller requested the changes, the licensee still decides what the listing says; a seller's wish does not make an unsupported fact accurate.

Practical exercise

Use a fictional listing only.

  1. Write a fictional verified property file with ten facts, including one measurement in acres and one improvement with a year.
  2. Write version 1 remarks of about 180 words from the file, and build its fact table with a source line for each fact.
  3. Freeze the table.
  4. Write three revision requests: reorder, shorten to 120 words, and change the tone.
  5. For each request, rewrite it so that it explicitly forbids adding, changing, merging or removing facts.
  6. Run the three rounds on an approved tool, re-sending the sources each time, or write the versions you would expect, planting one drift, one merge and one addition.
  7. Build a fact table for each new version.
  8. Compare each table with the frozen table and label every difference as same, changed, merged, dropped or added.
  9. Convert the acreage to square feet and check any converted figure in the drafts.
  10. Decide which differences to restore, which to accept, and which to delete, with a reason for each.
  11. Re-read the final version in full for fair housing and advertising problems.
  12. Write the record entry naming the released version and the rounds that produced it.
  13. Note which round introduced the most problems, and write one revision instruction you would phrase differently next time.

Knowledge check

  1. Why must every revision be treated as a new draft?

    Answer guidance: A model can change, merge, drop or add facts in any round, often in sentences you did not ask it to touch, so each version needs a new fact table compared against the frozen verified one.

  2. When is it better to edit by hand than to ask the model to revise?

    Answer guidance: When the change is small, such as reordering a sentence or trimming a word. Sending a small change back invites the model to rewrite everything else, which then needs re-verification.

  3. How should a genuinely new fact enter a draft during revision?

    Answer guidance: Through the sources: add the document, verify the fact, update the frozen table, and only then ask for the revision. A revision request is never where a new fact enters.

  4. What does the OWASP 2026 project say about misinformation that matters for brokerage drafts?

    Answer guidance: That when a model's fluent, confident output drives a decision or action, a wrong answer becomes a wrong action. In a brokerage that action is a published listing or a sent client update.

Sources and further reading