We are living in an age where artificial intelligence, large language models, and generative pre-trained transformers are rapidly modifying the way that many people interface with information. For the real estate industry, AI, LLMs, and GPTs have the potential to streamline processes, to answer questions in different ways, and to assist in the creation of concepts and images. Real estate agents must understand several concepts about AI and AI tools before using them, though, and that understanding needs to be updated constantly as the technology changes and improves.
AI systems are designed to respond in a pleasing manner: LLMs and GPTs are made to answer the question with a reasonable best answer. It is a confident child that has learned something and believes confidently that what it has within its head is the sum total of knowledge in the world. If the system has learned everything there is to know about a subject, the answer will potentially be quite accurate. If the system has not learned everything there is to know about a subject, the answer will potentially *sound* quite accurate, which is an entirely different outcome. For topics of extreme importance, legal consultation, strategy, and financial planning, the question for a user is “if the device was even 95% accurate, am I willing to accept the risk that I’m getting the 1-in-20 wrong answer?” For that reason, even when using an artificial system for informational purposes or for minor topics, the user must always double check and review the response for accuracy.Separately, the manner that an AI system is questioned will modify the answer substantially. It wishes to answer your question and gleans from the question implied guidance. Different results occur when you ask for “10 reasons why Portland is awesome” and “10 reasons why Portland is a dump,” even if the user’s intent was to get neutral guidance. The way that one questions will become nearly as important as the question itself.
The primary concern with AI system answers is the risk of a “hallucination,” or incorrect answer. Earlier versions of ChatGPT and Bard [now Gemini] showed hallucination rates of 40% and 90% respectively, and while those numbers have gone down over time, the risk of a machine error remains. With more advanced or specific training data, the hallucination rate can go down, but be cognizant that the machine has the ability to spit out wrong answers as easily as a person does.
AI systems are oftentimes trained on specific datasets: The systems are fed with training data and that informs the system how it should act. The training data can be given a weight or gravity that determines how strongly the system values the data compared to other inputs. A clean way of putting it is that “algorithms are opinions embedded in code.” For example, if you had a chatbot that was supposed to give recommendations to a client about where they should think about living, and you only fed it highly weighted information from https://visittheoregoncoast.com/regions/, you’ll find the outcomes tend to recommend living on the coast, because the system has been told “value this stuff.” This sort of training data bias will rarely show up in the outputs because the system is designed to answer predictively and provide what amounts to the most “right” response to a question. The right answer rarely involves a statement like “my response is strictly biased towards recommending the coast because I only know things about the coast.” This sort of training bias can create unexpected outcomes, as X, formerly Twitter, and Grok learned earlier in 2025 [https://www.engadget.com/ai/how-exactly-did-grok-go-full-mechahitler-151020144.html].
In defense of training data, however, weighted usage of training documents can ensure that the system responds extraordinarily accurately within the walls of its training information. If you have a GPT trained on the library of Shakespeare, you could confidently ask it to find every time a character says something like “prithee” or “yonder,” and it would rapidly unearth all relevant citations within the trained data set.For the user, the important takeaway is to realize that the AI may be answering a question in a slanted manner, rather than an accurate manner. Imagine you were in a room with NFL fans wearing their jerseys. If you asked them all who the best team was, the person wearing the Seahawks jersey would likely always say Seahawks, the person in the Patriots jersey would probably say the Patriots, and so on and so forth. Sure, you’ll have the occasional Cardinals jersey person who says, “I like the Browns now,” but you can generally guess the person’s bent from the obvious visual. Now imagine there’s a person in the middle of the room who has the same feelings about football teams as the rest of them, but is wearing a blank t-shirt. If you ask them for their opinion on the best team, you aren’t getting an oracular answer from that ungrouped individual; you’re just getting a biased answer from someone who forgot their jersey that day. The same is true of AI. You don’t know what team the AI is rooting for and should assume the system is designed around telling a particular kind of story rather than churning out objective truths.
Some enterprises, such as the Stanford Human-Centered Artificial Intelligence Group, suggest “auditing the AI” by spending substantial time querying the algorithm with inputs and variations on inputs to observe the output and draw inferences into the data that was used to train the system. This sort of intentional and deep investigative questioning takes a lot of time and cognizance to not fall into traps of overreliance.
The AI follows a set of instructions that are oftentimes extraordinarily complicated or unknown to the user: Artificial Intelligence tools are programmed. They are complicated systems that follow dedicated instructions with robotic precision. The systems update regularly and are fed absurdly large quantities of data. ChatGPT-3 was fed approximately 45 terabytes of data [e.g., 3.66 years of video, or 17.3 million ebooks, or 45 million photographs]. Figures for version 4 and the newer version 5 are still hidden, but you can be assured that the figure is substantially larger. It’s not clear what that data was; it’s not clear what the system’s weighting on that data is. There are oftentimes hard walls, such as restrictions on a person’s ability to “jailbreak” the GPT and force it to override its programming. At times, the system will say “I can’t do that,” and it is left to the user to wonder whether the limitation is technological and actual, or man-made and intentional.
Overreliance on AI results can result in decreased trust: Studies, such as 2024’s “Trust and Reliance on AI,” have shown that overreliance on AI without human oversight will oftentimes result in less trust in the outcomes. We still prefer to have a person say “this is accurate” because the machine is not entirely trusted. If a user parrots an AI response without disclosing the origin, the result can be substantial loss of trust when the charade is discovered. With all that said, the statement acknowledging the usage of systems, such as “I asked an AI this question, it provided the following response, and as a professional, I can vouch for these portions as strictly accurate, for the following reasons…” will not inherently cause issues of trust because the credentialed professional is able to explain the reasoning behind the accuracy.
Data security concerns are real: Information input into an LLM or GPT will sometimes be sucked up into the data collection systems of the program, often with the intent of training the system to improve future responses. A user should never input personal data or client data into the system without fully understanding the risk that it carries. If the parent company ever gets hacked [AI parent companies are, metaphorically, overladen Spanish galleons dripping with golden data waiting for a pirate to attack], any information shared could potentially be commandeered [https://www.zdnet.com/article/chatgpt-can-leak-source-data-violate-privacy-says-googles-deepmind/].
Your insurance probably doesn’t cover it – yet: Most insurance policies treat AI liability policies as either their own thing, or as a rider on a cybersecurity policy. Over time, this will likely change, but you should consult your E&O policy to see if it covers your use of such systems before ever using AI for your job, and if the action is not covered, balance the risk of the uninsured action against the value that the system provides.