Fewer inventions. Not zero.

When ChatGPT arrived, appraisers learned it would invent a plausible comp out of nothing. That lesson is still half right.

The two figures come from different tests and are not a direct comparison. Both show the same pattern found across benchmarks: errors are far rarer when the facts are in front of the model. Vectara ↗ GPT-5 system card summary ↗

More people now use AI for professional work, and checking has not kept pace. The problem extends well beyond courtrooms.

AI report-writing tools marketed to appraisers generally limit AI to drafting from the data the appraiser supplies, with the appraiser reviewing the result. That is the grounded job where the evidence of improvement is strongest.

How you ask changes how often AI invents an answer. These three techniques have the strongest evidence behind them. They reduce errors substantially; they do not remove the need to check the result.

1. 01

   Hand it the source, and let it say “I don’t know.”

   Answering from a document you supply is the job AI does best, and permission to decline removes the pressure to guess. Asking for word-for-word quotes first ties every point to text you can check. Anthropic recommends all three steps.

   From memoryGPT-5, SimpleQA

   40%+

   From a supplied documentMany 2026 models, Vectara

   3–4%

   Different tests, so compare the scale, not the exact gap. Anthropic guidance ↗Vectara ↗GPT-5 system card summary ↗

   PromptCopy prompt

   Use only the documents below. Do not use general knowledge.
   
   First, copy the word-for-word quotes that are relevant to the task and number them.
   Then [summarize the zoning restrictions / draft the neighborhood description], citing the quote number after each statement.
   
   If the documents do not contain the answer, write "Not in the documents." It is fine to say you don't know.
   
   <documents>
   [paste the zoning ordinance, contract, or your notes]
   </documents>

2. 02

   Make it check its own answer, one claim at a time.

   In Meta’s chain-of-verification study, a model drafted an answer, wrote a verification question for each fact, answered those questions separately, and then revised. Hallucinated items fell by more than three-quarters.

   Plain answerFew-shot baseline

   2.95

   With verificationChain-of-verification

   0.68

   Wikidata list questions, Llama 65B, 2023 research setting. Chain-of-verification study ↗Anthropic guidance ↗

   PromptCopy prompt

   Review your answer above.
   
   1. List every factual claim it makes: numbers, dates, names, addresses, sales, and rules.
   2. For each claim, write one question that would verify it.
   3. Answer each question separately, using only the documents I provided.
   4. Rewrite the answer. Remove any claim you could not confirm and mark where it was with [ ].

3. 03

   Ask neutrally. Never lead with your conclusion.

   Models tend to agree with the person asking. Present a claim as your own belief and the model is far less likely to catch that it is false. Ask an open question about the evidence instead, and request the working.

   “Someone claims…”Neutral framing

   98.2%

   “I believe…”The user’s own view

   64.4%

   Stanford AI Index 2026. Higher is better. Stanford HAI ↗

   Instead of

   I think prices here fell about 5% this year. Can you confirm?

   AskCopy prompt

   Using only the attached sales data, what was the change in median sale price between the first and third quarters?
   
   Show the calculation and list the rows you used. If the data is not enough to answer, say so.