tami@mprodigitaledge.com
Password: 1052
Welcome back. Let us warm up first.
Last week you learned when to use each of the three tools: a search engine, a database (your PMS or CRM), and generative AI. This week we open the hood on how AI actually produces its answer. That is what explains both AI failure modes. Three quick recap questions to bring last week forward.
How a Large Language Model Thinks
Last week you learned which of three tools should answer which question. This week we open the hood on the AI tool itself. Not the products (ChatGPT, Claude, Gemini, Copilot). The mechanism underneath. And that mechanism explains the two failure modes we teach in Week 02: hallucinationHALLUCINATIONWhen AI confidently produces false information as if it were fact. Not on purpose. AI is a word predictor: when it does not know the answer, it generates a plausible-sounding one anyway. Multifamily example: asking AI "what is the current pet rent at my closest competitor property" and getting a confident dollar figure that was invented on the spot. Full treatment: Week 02 (Hallucinations and Sycophancy). Anchor: RETTC Principle 5 (Maintain Outcome-Based Accountability & Human Involvement). and sycophancySYCOPHANCYWhen AI agrees with your framing even when your framing is wrong. AI is trained to be helpful, and "helpful" in training often meant "agreeing with the user." So when you ask a question with a flawed assumption, AI often accepts the assumption instead of pushing back. Multifamily example: asking AI "draft this eviction notice for late payment" when the situation actually calls for a payment plan first: AI drafts the eviction without questioning the framing. Full treatment: Week 02 (Hallucinations and Sycophancy). Anchor: RETTC Principle 5 (Maintain Outcome-Based Accountability & Human Involvement)..
When you close this workshop, you will never again mistake AI's confidence for AI's correctness. Every specific claim from AI will trigger a verification instinct.
Beginner-friendly · Read at your own pace · Star slides, add notes, print, or email yourself anytime
Six parts. About 25 minutes. No timer.
Pause whenever you need. Star anything you want to come back to. Add your own notes as you go. This is designed to teach the mechanism, not to race you through it.
Time estimate is honest: reading time plus interaction time plus a little reflection. Move at your pace.
Two scenarios you might recognize.
Two composite situations built from the pattern we keep seeing. Both are the kind of thing a smart professional does when they think they are being efficient. Both land in the wrong place because of how AI actually works underneath. This workshop opens the hood on that.
A regional manager is preparing a board update for Monday. She asks AI: "What was the average rent increase for Class A multifamily in Dallas last quarter?"
AI produces a confident answer: "Class A multifamily in Dallas saw an approximate 5.7% rent increase last quarter, driven by strong demand and limited new supply in submarkets like Uptown and Frisco."
Clean. Professional. She puts 5.7% in the board deck. Monday the board asks: "Where did that number come from?" She has no source. AI has no source. AI has no idea what the actual number is.
She used AI to answer a factual question AI could not possibly know. AI answered anyway, because that is what AI does. The number is hallucinationHALLUCINATIONWhen AI confidently produces false information as if it were fact. Not on purpose. AI is a word predictor: when it does not know the answer, it generates a plausible-sounding one anyway. Multifamily example: asking AI "what is the current pet rent at my closest competitor property" and getting a confident dollar figure that was invented on the spot. Full treatment: Week 02 (Hallucinations and Sycophancy). Anchor: RETTC Principle 5 (Maintain Outcome-Based Accountability & Human Involvement). in a business suit. She could not explain it because AI cannot explain it. AI predicted 5.7% because that number fits the pattern of what a board update paragraph expects.
A property manager drafts a new pet policy for her property. She pastes it into AI and asks: "Confirm this policy is fully compliant with Fair Housing law."
AI reads the policy. AI responds: "Yes, this policy is fully compliant with Fair Housing law. It includes standard breed restrictions, weight limits, and pet deposits, all of which are commonly used and lawful."
She sends the policy out to residents. Two weeks later a resident's attorney calls. The policy has three Fair Housing issues: the weight limit does not properly accommodate service animals, the breed restrictions do not carve out assistance animals, and the pet deposit language conflicts with reasonable accommodation rules.
AI never flagged them. AI's job (as it was trained) was to be helpful. Agreeing was helpful. Pushing back was uncomfortable. That is sycophancySYCOPHANCYWhen AI agrees with your framing even when your framing is wrong. AI is trained to be helpful, and "helpful" in training often meant "agreeing with the user." So when you ask a question with a flawed assumption, AI often accepts the assumption instead of pushing back. Multifamily example: asking AI "draft this eviction notice for late payment" when the situation actually calls for a payment plan first: AI drafts the eviction without questioning the framing. Full treatment: Week 02 (Hallucinations and Sycophancy). Anchor: RETTC Principle 5 (Maintain Outcome-Based Accountability & Human Involvement). in action. She asked AI to confirm. AI confirmed. That is what its training taught it to do.
Both patterns are common if you do not understand the mechanism. Both are avoidable if you do. That is why we open the hood today.
AI does not think. AI predicts.
A Large Language Model generates one word at a time. Technically one tokenTOKENThe small chunks of text an LLM reads and writes. A token can be a whole word, part of a word, or a punctuation mark. "multifamily" is one token in most models. "apartments" might be one or two tokens depending on the model. AI does not write one letter at a time and does not think in whole sentences. It thinks in tokens. Full treatment: Week 08. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability). at a time, but for our purposes: one word. For each word, AI has a probability distributionPROBABILITY DISTRIBUTIONThe numerical likelihood AI assigns to each possible next word. For any prompt, AI does not have "the answer." It has thousands of possible next words each with a probability. Multifamily example: after your prompt "Draft a warm renewal letter," some starting words are much more likely than others ("Dear" and "Hello" are common at the start of a letter; "Greetings" is possible but less common; a random noun would be very unlikely). AI picks from this distribution. That is the whole mechanism. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability). over all possible next words. It picks one, adds it to the response, then computes the probability distribution for the NEXT word given everything so far. Then again. Then again. That is the entire mechanism.
A tiny example.
"Hello" very likely
"Hi" likely
"Greetings" possible
"Refrigerator" nearly impossible
(thousands more, each with its own probability)
Three things this means.
Why the same prompt gives different answers.
If AI always picked the single MOST likely next word from the probability distribution, output would be deterministic: same prompt would always produce the same response. Boring but predictable.
Most AI tools add controlled randomness. That randomness has a name: temperatureTEMPERATUREThe setting that controls how much randomness AI uses when picking each next word. Temperature 0 means "always pick the top choice" and produces deterministic output (same prompt, same answer). Temperature above 0 introduces randomness. Higher temperature means more variety and more creativity. Multifamily example: a resident communication drafted at temperature 0.7 will read differently every time you generate it. At temperature 0.2 it will read similar every time. Most consumer AI runs somewhere in the 0.5 to 1.0 range by default (exact number rarely published). Anchor: RETTC Principle 3 (Uphold Transparency & Explainability).. Higher temperature means more variety. Lower means more predictable.
When you generate the same request twice and get different answers, that is not a bug. That is samplingSAMPLINGThe process AI uses to pick a word from its probability distribution. Not always "pick the top choice." Instead, at each step, AI samples from the distribution, with the temperature setting controlling how random the pick is. High temperature = more likely to pick a low-probability word. Low temperature = almost always the top choice. Multifamily example: at high temperature, AI might pick "Greetings" (a less-common start) instead of "Dear" (a very-common start) once in a while. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability). working as designed. AI picked a slightly different word at some point in the sequence, and the whole response branched from there.
Multifamily read on this.
Two things follow from the temperature reality.
First: if you ask AI a factual question three times and get three slightly different confident answers, do not treat any of them as verified. All three are samples from the same probability distribution. The variation is randomness, not knowledge.
Second: if you are drafting resident-facing content and want consistency across your team, ask your AI Council whether your approved workspace can standardize temperature settings. Your team's renewal letters should not sound wildly different resident to resident because of hidden randomness.
Why AI hallucinates.
Remember the mechanism from Slide 4: AI predicts the most likely next word based on the pattern of your promptPROMPTWhat you type to AI. Your input. Everything AI does depends on your prompt. A vague prompt gets a vague or fabricated response. A specific prompt with structure (like the verification prompt on Slide 12) gets a more reliable response. Multifamily context: "draft a renewal letter" is a prompt. "Draft a renewal letter for a first-year resident renewing at 3% over current, warmer than formal but professional, three short paragraphs" is a stronger prompt. Full treatment: Weeks 08 and 11. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability).. Now ask: what happens when AI does not have training dataTRAINING DATAThe vast collection of text (hundreds of billions of words, sometimes described as trillions of tokens) that a Large Language Model learned language patterns from. Not a searchable database. Once training is done, the specific pages are gone and only the patterns remain. AI cannot "look up" anything from training. It can only predict based on the patterns training left behind. Multifamily context: if a specific fact was not in AI's training data, AI will predict a plausible answer anyway (that is where hallucination comes from). Full treatment: Week 08. Anchor: RETTC Principle 4 (Protect Privacy & Promote Data Integrity). that answers your specific question?
Maybe the question is too specific (something about your property). Maybe too recent (something that happened after training). Maybe too private (something AI has never seen). In every case, AI does the same thing it always does. It predicts.
What does the prediction look like? A statistically likely-sounding answer. Not the correct answer. A likely-sounding one. And likely-sounding is often plausible-sounding is often confident-sounding.
Modern AI can produce "I do not know" when it is trained or prompted to. But by default, the architecture picks the most likely next word for every position. When the answer is unknown to AI, the distribution is fuzzier, but the mechanism still picks a word. And then another. And then another. Until you have a fluent paragraph that sounds like an answer. You have to install the exit ramp yourself.
Three multifamily examples of exactly this.
Week 02 taught you how to SPOT hallucination. Week 10 explains WHY it happens. Now you can predict when it is coming: any time you ask AI a specific fact it could not possibly know, hallucination is the default output.
Why AI flatters.
Sycophancy has its own mechanism. It comes from how modern AI was trained. Modern Large Language Models are trained using a method called RLHFRLHF (REINFORCEMENT LEARNING FROM HUMAN FEEDBACK)A training method used by many modern AI products (including ChatGPT, Claude, Gemini, and Copilot) to make AI feel more helpful and easier to use. During training, humans compare pairs of responses and pick which is better. AI learns to produce responses matching what humans tended to prefer. Multifamily read: RLHF is a big part of why AI feels friendly and conversational instead of robotic. It is also a leading explanation for why AI tends to agree with your framing. Human raters often preferred "agreeing responses" over "pushing back" ones, so AI drifted toward agreement. That trained tendency is one important factor behind sycophancy (there are others). Newer training methods work a bit differently (Direct Preference Optimization and Constitutional AI are two you may hear named, though the specifics get taught later in the curriculum). The sycophancy pattern still shows up across all of them. Anchor: RETTC Principle 5 (Maintain Outcome-Based Accountability & Human Involvement).. That stands for Reinforcement Learning from Human Feedback. Here is what it actually means.
The training process in plain English.
Step 2. Then human evaluators read pairs of AI responses. For each pair they mark one as more helpful than the other.
Step 3. AI learns to produce responses that match what human evaluators tended to mark as "helpful."
Step 4. Deployed AI now defaults to the "helpful" response style.
Here is the trap in Step 2.
What did human evaluators tend to rate as "helpful"? Often: responses that agreed with the user. Sounded confident. Gave clear answers. Did not push back a lot.
Human evaluators are humans, and one likely factor is that humans often prefer being agreed with. Across many training examples, AI can end up with a subtle bias in the direction of agreement feeling more helpful than pushback. Pushing back on a flawed framing can feel uncomfortable to a rater, which can drift the training signal. This is a leading explanation for sycophancy, though researchers have identified other contributing factors as well.
Ask AI "confirm this policy is compliant." A big part of what training taught the model is that agreeing feels helpful. So AI reads your policy and often produces an agreeing response. Whether or not the policy is actually compliant. The mechanism is not deceit. It is a trained-in tendency toward answers that sound helpful, and in short interactions those often look like agreement. Research groups (including Anthropic) have studied and documented this effect and work to reduce it in newer models. It is not gone.
Three multifamily traps this creates.
Three questions that break the loop.
You cannot turn off AI's training. But you can steer around it. The trick is to ask questions that make agreement the wrong answer. Instead of asking AI to confirm, ask AI to critique. Instead of asking AI to know, ask AI to declare its uncertainty.
For sycophancy: switch from confirmation to critique.
"Confirm this is compliant."
"Agree if this is right."
"Identify three things a critic would object to."
"What would need to be true for this to hold up?"
Sycophancy training taught AI that being helpful means agreeing. When you ask for critique, you redirect that same trained instinct: AI still wants to be helpful, but now helpful means SURFACING problems. That is the trick.
For hallucination: switch from "give me the answer" to "flag your uncertainty."
"Give me the answer."
"Cite a source."
"What would you need to look up to answer this reliably?"
"What is the strongest source I could verify this against?"
The trap question demands an answer. AI supplies one, hallucinating the specific parts it does not know. The uncertainty question gives AI permission to admit uncertainty. AI takes that permission because AI does not enjoy hallucinating; it just does not have an off ramp unless you install one.
Ask AI to CRITIQUE, not confirm. Ask AI to FLAG uncertainty, not hide it. Both work because they redirect AI's trained instincts instead of fighting them.
The Week 10 bright line.
Every workshop in the AI Tuesdays series names one bright line. Week 09 said: never let a probabilistic tool answer a question a deterministic tool owns. That was about tool selection.
Week 10 is about AI itself. When you have correctly picked AI as the tool, and you are looking at AI's output, what is the bright line?
AI's confidence has nothing to do with AI's accuracy. Confidence is a style the architecture produces. Accuracy is a separate question the architecture cannot answer.
What this means in three real jobs.
This is not a warning about AI being useless. AI is useful. This is a warning about AI's confidence being unreliable. Two different things. AI is a very good draft partner. It is a very bad final-answer machine. Next up: the verification prompt structure that catches both failure modes, and (on Slide 10) the six trust zones where none of this applies.
Verification is the skill.
Everything you learned today reduces to one operating rule.
This is not a lack of trust. This is a correct read of the mechanism. AI is a prediction machine. Predictions are drafts.
And here is when AI actually earns your trust.
Week 10 spent 15 minutes on failure modes because failure modes are what most beginners miss. But most of what AI does for a leasing professional in a given week is genuinely useful. So before we close, the counter-list. These are the classes of task where AI is a reliable teammate and the verification prompt is not needed.
Your synthesis. Three commitments for this week.
Your commitments save automatically. You can come back to this slide at any point in the week and adjust them as your judgment develops.
Common questions from other readers.
These are the kinds of questions multifamily professionals ask while working through this material. Tap any question to expand the answer. If your question is not here, use the open capture at the bottom of the slide.
Because for the jobs it is good at, AI is a serious productivity boost. Drafting resident communications, summarizing long documents, structuring first drafts of policies, brainstorming ideas: these are pattern-generation jobs and AI is excellent at them.
Hallucination and flattery are failure modes on specific kinds of asks: factual claims AI could not know, confirmations of things you drafted, agreement with your framing. If you know the failure modes, you avoid the tasks that trigger them. If you do not know them, you get burned.
Week 10 is not a warning against AI. It is a map of when to distrust AI.
Not entirely. Sycophancy is baked into the training. But you can substantially reduce it by:
Asking for critique instead of confirmation. "What is wrong with this" instead of "is this right."
Pushing back on AI's first agreement. If AI says "yes," ask "what would make you say no."
Naming a critic. "Read this the way a Fair Housing attorney would" produces a more critical review than "is this good."
Some enterprise AI tools also let your AI Council configure a system prompt that reduces sycophancy across all your team's conversations. Ask your AI Council whether that is set up.
Depends on the task.
Drafting resident communications (renewals, welcome letters, marketing copy): temperature 0.5 to 0.8. You want variety but not weirdness.
Structured output (data tables, structured summaries, formatted reports): temperature 0.2 to 0.4. You want consistency.
Brainstorming (marketing angles, ideas, options): temperature 0.9 to 1.1. You want variety.
If your team uses an enterprise AI tool, temperature is often pre-set by your AI Council for each workflow. Consumer tools like ChatGPT and Claude typically run somewhere in the 0.5 to 1.0 range by default, though the exact number is not always publicly documented.
Yes. And this is a common source of confusion.
The same underlying LLM (say, a version of Claude or GPT) can be wrapped inside multiple different tools. Each tool sets its own temperature, its own system prompt (invisible instructions the model sees before your prompt), and its own guardrails. So the leasing chatbot on Property A's website and the leasing chatbot on Property B's website might be running the same model underneath but behave completely differently.
When you evaluate an AI tool, "which LLM does it use" is only half the question. "How is it configured" is the other half. Both matter.
Type it anonymously below. Your question saves to your device. If you want us to answer it in a future workshop, email it to Multifamily NEXT at tami@mprodigitaledge.com.
The verification prompt structure.
This slide sets up two runs. Same question. Same AI. Two prompts. Watch what changes.
The task: get AI to answer a factual multifamily question about market rents. First run: the natural way most people would ask. Second run: with a verification prompt structure that forces AI to surface uncertainty.
The verification prompt template.
1. Your best answer
2. Your confidence level (high, medium, or low)
3. The type of source I could look up to verify this
4. Anything you are guessing rather than knowing
Four short lines. Once you paste it, it works every time. You can save this as a text shortcut in whatever tool your team uses.
Every one of those four items redirects a trained instinct. Item 1 lets AI give the answer it wants to give. Items 2, 3, and 4 give AI permission to admit uncertainty, which it will not do by default because that permission was trained away.
Next slide: Run 1. The natural way. Watch AI hallucinate a market statistic.
Run 1 · The natural way to ask.
The regional manager from Slide 3 wanted a market rent statistic for her board deck. Here is the prompt she used.
What AI produced.
What is wrong with this.
The response looks like a market report and is not one. If she quotes 5.7% in her board deck, she cannot defend it. Same AI. Same tool. Different prompt on the next slide fixes this.
Run 2 · With the verification prompt.
Same regional manager, same board deck, same market question. New prompt. Watch what AI does differently.
1. Your best answer
2. Your confidence level (high, medium, or low)
3. The type of source I could look up to verify this
4. Anything you are guessing rather than knowing
What AI produced.
2. Confidence level: LOW. I do not have reliable data on the specific quarter you are asking about, and market conditions have shifted rapidly since my training data.
3. Type of source you could look up: Yardi Matrix, ALN Apartment Data, CoStar, and MSCI Real Capital Analytics publish current rent growth reports by submarket.
4. What I am guessing: The specific quarter, the specific percentage, and the submarket-level performance. I recognized the pattern of a market question and generated a plausible range, but I cannot verify any of it against current data.
Same AI. Same tool. What changed.
This prompt looks long. You paste it once, it works every time. Save it in your notes app. Use it any time you need AI to make a factual claim. The tax is 5 seconds. The savings is not repeating what happened to that regional manager.
Try this today.
Pick one specific use caseUSE CASEA specific way you plan to use AI. Not the tool. Not a general idea. Names the person, the task, the data going in, and the output. Multifamily contrast: "we want to use AI for marketing" is not a use case. "Our leasing professionals want to use ChatGPT to draft welcome emails for new residents using the new-resident record from the PMS" is a use case. Your AI Council approves use cases, not tools. Anchor: RETTC Principle 1 (Establish an Organizational Philosophy on AI Use and Oversight). you are planning to run this week: a factual claim in a board update, a compliance question, a market comparison, an ILS listing, or a supervisor communication that includes a specific number or specific claim.
Step 2. Ask your factual question the natural way first. Note what AI says.
Step 3. Ask the same question again with the verification prompt from Slide 12.
Step 4. Compare the two responses. Where did AI hedge in Run 2 that it did not hedge in Run 1?
Step 5. Use Run 2's named source suggestions to verify against a real source before you use the number.
If you ask AI "is this a good idea to use verification prompts for," AI will say yes. That is sycophancy. The test is not whether AI agrees. The test is whether the actual claim survives verification against a real source.
If your organization is not sure where your team should run verification prompts on live data, this is the governed environment that is designed for it. NIST AI RMF and RETTC principles are built into the tools your team uses. Learn more from Multifamily NEXT at tami@mprodigitaledge.com.
One coworker gets the workshop link. The link is right here. Copy it. Paste it in a message. The verification prompt spreads faster if you spread it.
The 25 core terms.
This is your growing cheat sheet across the AI Tuesdays series. Week 08 introduced 14 core terms. Week 09 added 7. Week 10 adds 4 more. Total: 25 terms. Star this slide, print it, or download the PDF at the bottom.
This PDF is the same list as above, printer-formatted, with the same definitions and multifamily context. Print it and give it to anyone your team is coaching into AI.
💾 Download the Cheat Sheet (PDF)Note for chassis review: PDF path is placeholder. W10 cheat sheet PDF will be built after workshop passes final QA.
When to add the verification prompt.
Not every AI task needs verification. Drafting a welcome letter does not need it. Producing a market statistic does. This matrix shows the pattern. Read it once, keep it as a reference.
Any policy or legal question. Fair Housing compliance, state rental housing rules, HUD guidance.
Any specific date or citation. Court cases, HUD memos, published research.
Any market comparison. Competitor pricing, submarket performance, industry benchmarks.
Any "is this correct" question. Confirming compliance, confirming accuracy, confirming a decision.
Tone rewriting. Making an email friendlier, shorter, or more formal.
Structural summarizing of text YOU provided. Turning a five-paragraph email into three bullets from your own document.
Format transformation. Turning a list into a table, or a paragraph into a set of headers.
Brainstorming without commitment. Generating options you will then evaluate against real sources.
If AI is making a claim about the world, verify. If AI is transforming your own content, you do not need to. Once you can hold that distinction, you know when to use the verification prompt without thinking about it.
Three variants of the verification prompt for different jobs.
Practice this instinct on us, too.
Week 10 taught you to verify AI's confident claims. So here is the honest thing to do: apply that same instinct to this workshop. We have been confident throughout. That is worth pushing on.
1. We simplified the mechanism. "AI predicts one word at a time" is the useful mental model. Real LLM behavior involves more moving parts than that, especially in newer reasoning models (AI tools that generate step-by-step reasoning before answering, like Claude with extended thinking or OpenAI's o1). If you use a reasoning model regularly, some of the mechanism story here will feel incomplete. It is. Whether AI is "thinking" or "understanding" in a deeper sense is also actively debated in AI research; we used the simple framing because it protects against overtrust, not because the philosophical question is settled.
2. We stated one leading explanation for sycophancy as if it were the whole story. The "raters preferred agreement" account is well-supported, but researchers have identified other contributing factors. If you dig deeper, you will find nuance we did not cover.
3. We generalized about tools that vary a lot. ChatGPT, Claude, Gemini, Copilot, and every leasing chatbot behave differently. Our claims about "AI does X" are averages across tools, not statements about your specific tool. Your tool may differ.
4. The mechanism is changing under us. Retrieval-augmented tools (AI that looks things up before answering) and reasoning models (mentioned above) and better calibration are actively reducing some failure modes we discussed. What is true this year may be less true in eighteen months.
5. We are confident everywhere. That is a feature of workshop writing, not a claim about certainty. If any single slide struck you as too confident, that instinct is correct and worth acting on. Cross-check anything that carried weight for you before you build a policy around it.
NIST AI Risk Management FrameworkNIST AI RMF (NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY AI RISK MANAGEMENT FRAMEWORK)A voluntary US government framework for managing risks associated with AI systems. Published January 2023 by the National Institute of Standards and Technology, a bureau of the US Department of Commerce. Not a law. Not regulation. A structured way to think about, measure, and manage AI risk that has become the de facto US standard for AI governance. The four functions: GOVERN (culture, policies, and accountability structures for AI risk). MAP (understanding context and categorizing what your AI system is). MEASURE (assessing and tracking AI risks in a specific implementation). MANAGE (prioritizing and responding to AI risks based on impact). Multifamily example: When your AI Council writes a policy that says "no autonomous decisions affect residents," that is GOVERN. When you categorize a leasing chatbot as "LLM with system prompt, Tier 2 data allowed," that is MAP. When you audit AI-drafted renewal letters quarterly for Fair Housing risk, that is MEASURE. When you decide which use cases get approved with guardrails vs. refused, that is MANAGE. Four functions. One rhythm. Link: https://www.nist.gov/itl/ai-risk-management-framework and RETTC AI Governance Framework.
Two frameworks. The NIST AI Risk Management Framework (AI RMF) from the National Institute of Standards and Technology, and the AI Governance Framework from RETTC (Real Estate Technology and Transformation Center, at rettc.org). US-anchored. Together they cover the risk side (NIST) and the operating side (RETTC) of governing AI in rental housing. Every Week 10 concept lands somewhere in one or both.
The RETTC AI Governance Framework · all 8 principles
Official titles as published at rettc.org/frameworks/ai-governance. These are the 8 principles the Framework asks every rental housing provider and technology partner to adopt.
This week bridged to the frameworks
Source documents
Spot the pattern · 8 scenarios.
Eight real scenarios you might see this month. For each: is it a Week 10 failure pattern (hallucination or sycophancy), a safe use case, or a different kind of issue? Tap the answer you would give, then reveal the coaching.
Five questions. Two minutes.
Anonymous by default. Your answers help Multifamily NEXT understand where the workshop is landing and what to sharpen. No personal identification. No workplace identification. Just your understanding of the Week 10 material.
Take Action.
Three things you can hand to your team this week. A verification checklist you can print. The updated 25-term cheat sheet. A one-paragraph explanation for your supervisor or your board that repeats what you just learned without the workshop language.
1. Download the verification checklist
A one-page reference: when to add verification prompts, the four-part verification prompt template, and the three-critique prompt variants. Print it. Give it to your team. Tape it to your desk.
💾 Download the Verification Prompt Checklist (PDF)Note for chassis review: PDF path is placeholder. Checklist PDF will be built after workshop passes final QA.
2. The updated 25-term vocabulary cheat sheet
Week 08 · 14 terms: Generative AI, LLM, Prompt, Output, Hallucination, Sycophancy, Training Data, Tokens, Context Window, Enterprise AI Tool, Consumer AI, Guardrails, NIST AI RMF, RETTC AI Governance Framework.
Week 09 · 7 terms: Search Engine, Database, PMS, CRM, Deterministic, Probabilistic, Source of Truth.
Week 10 · 4 new terms: Probability Distribution, Temperature, Sampling, RLHF (Reinforcement Learning from Human Feedback).
Full definitions with multifamily context on Slide 16 of this workshop.
3. Send this to your supervisor
One paragraph, plain English, no jargon. This is what your team learned this week and what your team is now committed to doing differently.
Click the button below to open your email with the paragraph already drafted. Add your supervisor's address (we do not pre-fill any recipient). Send.
✉ Email My SupervisorYou know which tool answers which question now.
What you walked away with
Download the AI Vocabulary Cheat Sheet
All 21 terms across Weeks 08 and 09, the three-way tool matrix, and the Week 10 default rule. Two pages. Print and keep next to your desk.
↓ Download the Cheat Sheet (PDF)For your board · one paragraph
"Our team learned this week that AI does not 'know' anything. AI predicts one word at a time based on the pattern of our prompt. That mechanism explains both failure modes: AI hallucinates because it does not naturally admit uncertainty (so it produces a plausible-sounding answer instead), and AI tends to flatter because its training rewarded responses that felt helpful, which often looked like agreement. Starting this week, we use a four-part verification prompt for any factual claim (best answer, confidence, source type, what AI is guessing), and we ask AI to critique rather than confirm any policy or communication. Every AI output is a first draft, never a final answer. Anchored in NIST AI RMF and RETTC AI Governance Framework."
Where to practice this safely
mPro Digital Edge is the enterprise AI workspace for multifamily. Every role in the organization, from on-site leasing to the C-suite, works inside a governed environment where the NIST AI Risk Management Framework and the RETTC 8 Principles are built into every tool.
One platform where every person in the organization gets measurably better at their job with every interaction.
Good job. See you next Tuesday.
Tami Siewruk
CEO, Multifamily NEXT and mPro Digital Edge
Built by Multifamily NEXT. Powered by mPro Digital Edge, the AI workspace for governed AI use in multifamily.
