Hey there, folks. Happy Friday. This week the industry got handed a genuinely important reality check, and it deserves your full attention. Researchers at Columbia University ran the leading AI models — ChatGPT, Claude, and Gemini — through basic mortgage-file tasks. The results were rough: none broke 80% accuracy, and one landed at 51%. Worse, the study found what researchers called "an extreme level of bias," with models flagging transactions as foreign based on whether a name sounded English. Meanwhile, a new Delta Media analysis shows brokerage leaders' AI worry scores climbing again after dipping last year — driven specifically by agentic AI. And two more states are moving to regulate AI-edited listing photos. After two years of "adopt or die," this week's theme is different: trust, verify, and know where the tools break. Let's dig in.
1. Leading AI Models Failed Basic Mortgage Tasks — and Showed "Extreme" Bias Against Non-English Names
This is the story of the week, and it's a genuinely important one. A study conducted by Columbia University's DAPLab and AI workflow startup Tidalwave — using a benchmark called MortarBench — tested the latest commercial AI models on simple, common mortgage-application tasks: matching account IDs, identifying large deposits on bank statements, verifying employer names in direct deposits and paystubs. As RISMedia's AI Pulse reported this week, the models overwhelmingly struggled. Gemini and ChatGPT's latest iterations landed a little over 75% accuracy. Claude came in at 51%. Even with a framework designed to correct for models' oversensitivity and false positives, the best any model managed was 80%.
The accuracy problem is concerning. The bias finding is alarming. Per the research, the models displayed what the authors described as "an extreme level of bias" — consistently classifying transactions as foreign when they came from a company or person with a non-English name, while never flagging those same transactions when the names were English. The researchers' conclusion is blunt and should stop anyone in their tracks: professionals interacting with non-English speakers "are likely to receive unequal treatment" from AI models. For an industry governed by fair housing and fair lending law, that's not a technical footnote — it's a legal and ethical red flag. Worth noting for context: the same research program previously found that a purpose-built, mortgage-trained model substantially outperformed general-purpose LLMs on these tasks, scoring 84% overall versus 71%, and 95% versus 42% on yes-or-no compliance checks. The gap reflects a real difference in design — general-purpose models treat a loan file as raw text, while domain-trained systems work with structured mortgage data.
Why It Matters: Read this one twice, because it cuts against a lot of the momentum we've been reporting all year. Three practical takeaways. First, general-purpose chatbots are not underwriting tools, and they're not document-verification tools. If you or your lending partners are pasting loan files or bank statements into ChatGPT to "check something," stop — a 51% to 80% accuracy range on basic verification isn't a productivity gain, it's a liability generator. Second, the bias finding demands attention from anyone whose AI touches consumer-facing decisions. Disparate treatment based on the ethnicity of a name is precisely what fair housing and fair lending law prohibits, and "the AI did it" has never been a defense — as California's DRE reminded agents earlier this year. If you serve diverse communities, this is a direct warning. Third, the purpose-built versus general-purpose distinction is the actionable insight: domain-trained tools integrated with real mortgage data meaningfully outperformed the consumer chatbots on the same tasks. When you're evaluating AI for anything consequential, ask what it was trained on and what systems it's connected to. The right tool for the right job — and a human checking the output — is not old-fashioned advice. It's the current best practice.
2. Brokerage Leaders' AI Worry Score Is Rising Again — and Agentic AI Is Why
Here's the perfect companion data point. A new analysis of the Delta Media Real Estate AI & Leadership Survey, covered by HousingWire this week, found the average AI "worry score" among brokerage leaders rebounded to 6.38 out of 10 in 2026, after falling to 5.80 in 2025. The share of leaders assigning AI risk a high worry score of 8, 9, or 10 followed the same pattern — dropping from 42.4% in 2024 to 33.7% in 2025, then climbing back to 37.9% this year. Concern went down as leaders got comfortable with AI. Now it's going back up.
The likely explanation is right in the survey design: 2026 was the first year in the three-year series to ask brokerage leaders specifically about agentic AI tools capable of automating tasks. The survey doesn't establish causation, but as the analysis notes, agentic AI's emergence adds new questions around compliance, data security, liability, and accountability when AI systems take action rather than just generate text. The breakdown by leader is also striking. Among female brokerage leaders, 71.4% identified compliance with real estate regulations as an AI concern, versus 39.1% of male respondents. Data privacy and security followed the same pattern — 65.7% of women versus 48.4% of men. Male leaders were more likely to cite uncertainty around AI costs and ROI (51.6% versus 28.6%). Smaller firms worried most about regulation; larger firms cited integration and security. The analysis draws on 100-plus brokerage leader responses collected annually from 2024 through 2026.
Why It Matters: Two weeks ago we covered this same survey's finding that AI holdouts are nearly extinct — 98% of brokerages are adopting. This is the other half of that story, and together they tell you exactly where the industry is: everyone's in, and everyone's nervous. That's not a contradiction; it's maturity. The worry is rising for a legitimate reason — when AI moves from drafting your emails to prepping and signing your paperwork, the stakes change. Liability questions that were theoretical when AI wrote listing copy become concrete when AI takes action on a transaction. The gender split in concerns is worth sitting with, too: if women in leadership are flagging compliance and data privacy at nearly double the rate of men, that's a signal worth heeding rather than dismissing, especially given what the mortgage study above just found about bias. The practical move for brokers: if your AI worry score is high but your AI policy is thin, close that gap. Write down what tools are approved, what data can go where, who reviews AI output before it reaches a client, and who's accountable when something goes wrong. Worry without policy is just stress. Worry with policy is risk management.
3. More States Move to Regulate AI-Edited Listing Photos
The regulatory response is starting to show up in state law. Per RISMedia's AI Pulse, two more states are considering rules on AI-edited listing photos, joining a movement that Wisconsin has already acted on. California's approach is instructive: agents and brokers there must include the original photo along with a disclosure if an image was significantly edited — whether by AI or by any other means. It's a straightforward standard, and one that translates easily into practice.
This lands on a practice that's become nearly universal. We covered the virtual staging boom back in the spring, when tools dropped to roughly $7 per room and adoption surged. The economics made AI staging irresistible for vacant listings, and it works — staged listings drive more clicks and showings. But the same capability that adds a virtual sofa can enhance a sky, remove a power line, brighten a dated kitchen, or quietly minimize a flaw. The line between "marketing" and "misrepresentation" is exactly what these state rules are trying to draw. The disclosure-plus-original-photo model is emerging as the practical answer: edit freely, but show the buyer what's real.
Why It Matters: This is the easiest item in this week's issue to act on, and there's no reason to wait for your state to legislate. Adopt the California standard voluntarily right now: if a listing photo has been significantly edited — AI staging, sky replacement, object removal, anything material — disclose it clearly and make the original available. Three reasons. First, it's very likely coming to your state anyway, and getting ahead of it costs you nothing. Second, it protects you: the gap between the photo and the in-person walkthrough is where buyer trust dies and complaints get filed. Third, and most underrated, it's a competitive advantage in a market where consumers are increasingly suspicious of AI-generated everything. Being the agent whose photos are honest — and who says so proudly — is a differentiator right now. That's the through-line of this entire issue: as AI capability races ahead, verified accuracy and disclosed honesty are becoming the scarce goods. Sell those.
That's the wrap, folks. After two solid years of "adopt or get left behind," this week delivered the necessary counterweight: the leading models flunked basic mortgage verification and showed real bias, brokerage leaders are getting more worried as AI starts taking actions instead of just writing text, and states are stepping in on edited photos. None of that is a reason to retreat — it's a reason to get deliberate. Use purpose-built tools for consequential work, keep a human on the verification step, write down your AI policy, and disclose your edits. The agents who pair AI's speed with genuine accountability are the ones who'll still be standing when the trust reckoning sorts everyone out. Have a great weekend, and I'll see y'all next week.
Disclaimer: AiRE Update is an independently produced newsletter that curates and summarizes publicly available news. I don't write the original articles featured here — I summarize them in my own words and add commentary on why they matter. All original reporting, content, and intellectual property remain the property of their respective authors and publications, including RISMedia (Jesse Williams), HousingWire (covering Delta Media Group research), and the Columbia University DAPLab and Tidalwave research team. Each story links back to its original source, and I encourage you to read the full articles there. The summaries and opinions in AiRE Update are my own and are provided for informational purposes only; nothing here should be taken as legal, financial, or professional advice.