What Is Wrong With The Appraisal Profession

[Essexfenwick]

They are trying the pound a square art into a round science hole. AI may be able to do that.

 

[Retired-70]

Who's claude?

Claude is a family of large language models from Anthropic, offered in tiers that trade off capability, speed, and cost (currently Haiku, Sonnet, Opus, and the higher Mythos/Fable tier). Beyond conversation, it can use tools: search the web, read and produce files (Word, Excel, PowerPoint, PDF), run code, and carry out multi-step tasks with some autonomy. Developers use it through the Claude API and through Claude Code, an agentic coding tool. Anthropic trains it to follow a published set of principles emphasizing helpfulness, honesty, and avoiding harm.

What to add so expectations stay realistic:

• It can make mistakes, including stating wrong things confidently, so important facts should be checked.
• Its built-in knowledge stops at a training cutoff; for current information, it has to search.
• It doesn't remember past conversations unless memory features are turned on.

I primarily use Claude Code, which can access my local files and do work for me. I may use Grok and ChatGPT to verify Claude findings. Claude Code is the primary choice of software developers.

 

[Essexfenwick]

The NAR has over 1.4M members. Appraisers don't have the numbers

And still.. realtors got destroyed in the buyer broker compensation laws. They have also been deeply effected by tech and it's accelerating. The percentage of realtors that actually make a good living from the job is microscopic.

Governments especially on the left hate all facets of the real estate business because it's a traditional American capitalist business that Joe blow can become wealthy in. It's the number 1 industry of individual millionaires in forever. Everybody in the business is a target of capitalism hating and envious leftists. The general independence and can-do personality types drawn to real estate as a career drive leftists insane. So here come ..laws, taxes, licences, accusations, lawsuits, fees, character assassination, coddling of squatters and deadbeats, nitpicking, elimination of income streams...

Appraisal is the asses end of a industry full of asses in the eyes of blue leaning government at all levels and a huge slice of the general public. It's been a steady temp increase for 40 years in appraisal job and now the pot is a boiling and more untenable than ever.

 

[ucbruin]

Blame the left....
I'm sure the horse whip manufactures did the same thing back in their day....

 

[Vermonter]

I primarily use Claude Code, which can access my local files and do work for me. I may use Grok and ChatGPT to verify Claude findings. Claude Code is the primary choice of software developers.

Not a huge Anthropic fan, but I dig Opus 5.5. So much more efficient than past models

 

[Retired-70]

Not a huge Anthropic fan, but I dig Opus 5.5. So much more efficient than past models

I harden my papers, typically running them through Claude Code, ChatGPT, and Grok. They each add their own criticism and input. Claude gladly accepts the critiques from Grok and ChatGPT, critiques their critiques and then integrates suggestions it also approves of. Now when I publish a paper and someone else runs my articles through AI, I can be pretty sure they are not likely to find too much to criticize.

 

[Elliott]

God knows the appraiser 'craft' has problems and issues that have taken many decades to put it in the soup. But sometimes AI gives answers because it or others think its the smartest person in the room. If only we could sprinkle magic fairy dust and they would go away.

Perplexity: "AI hallucination is one of the most important limitations to understand about tools like the one you are using right now. The core explanation is straightforward: an AI language model does not actually “know” facts—it predicts what words are statistically likely to come next, optimizing for plausibility rather than truth.

Why Hallucinations Occur
There are several overlapping causes, each contributing to the problem:
• Pattern matching without truth verification: The model generates whatever sequence of words fits the statistical context—correct and incorrect outputs can look identical to the model because it has no “truth checker.”[ibm +1]
• Sparse or low-quality training data: For obscure topics, rare people, specific dates, or niche facts, the training data is thin or contradictory. The model fills the gap with a statistically plausible guess rather than admitting it doesn’t know.[openai +1]
• Wrong incentives during training: Research by OpenAI found that benchmark evaluations reward confident, complete-sounding answers—and penalize saying “I don’t know.” This teaches the model to guess rather than express uncertainty.[pmc.ncbi.nlm.nih +2]
• Incorrect inhibition circuits: A 2025 Anthropic interpretability study found that Claude has internal circuits that normally prevent it from answering when knowledge is insufficient. Hallucinations occur when that inhibition misfires—for example, when the model recognizes a name but lacks enough facts about that person, it may still generate a plausible-sounding biography."

 

[Retired-70]

Hallucinations happen when users ask AI questions without providing the correct source information. If you want to ask Claude about USPAP, you will get the best answers if you provide a PDF version of USPAP for it to read. That is true of everything. If you want Grok, ChatGPT or Claude to give you the best possible answers for UDP 3.6, then provide them with ALL of the UDP 3.6 docs from Fannie Mae.

Secondarily, running prompts through several LLMs or having one LLM critique the output of another is also a good idea.

But mostly, hallucinations are caused by asking LLMs questions and not allowing them to say, "I don't have the answer" - and then not giving them the information they really need to answer the questions - THINKING that they should just know everything. In other words, by forcing them to make up **** to your satisfaction.

 

[Zoe]

There are so many variables in real estate and biases among shoppers (buyers and sellers). AI can only handle so much variables in a very heterogeneous real estate market. No two parcels of real estate are identical even in a very homogeneous (cookie cutter) market. What if your best comps are 20 miles away from the subject and each has tons of differences in site and other elements of comparison?

 

[Retired-70]

There are so many variables in real estate and biases among shoppers (buyers and sellers). AI can only handle so much variables in a very heterogeneous real estate market. No two parcels of real estate are identical even in a very homogeneous (cookie cutter) market. What if your best comps are 20 miles away from the subject and each has tons of differences in site and other elements of comparison?

MARS can handle all of that. It is VERY good at it. You just need some experience using it. I can go as far back in time as I need or as far out. Of course, you are asking for big adjustments. But the good thing is that going back in time, as far out as necessary, allows MARS to extract intelligence from more property types. Invariably, market areas that are very far apart share similarities. A large house in Tennessee will likely sell for more than a smaller one in Tennessee, all things being equal - and the same is true in California.