Seven things no AI can know about your Medicare, however good it gets
September 22, 2026
Listen Now
Subscribe
Ready to talk with an advisor?
The advisors at American Retirement Advisors have been having these conversations for over 25 years. A no-cost discovery session is the fastest way to find out where you actually stand.
Schedule a Discovery Session
Show Notes
Eddie and Betty's Conversation
Welcome to The American Retirement Advisor podcast. I'm Betty, and Eddie is here with me in the studio today. We are on part thirteen of our Healthcare in Retirement series, and today we are taking what I'd call a detour, a really useful one, because this episode isn't about a specific Medicare situation the way the last twelve have been. It's about the tool a lot of you are quietly using to try to make sense of all of it. AI. And specifically, whether you can trust what it tells you about Medicare. Eddie, I feel like we have been dancing around this topic for a while.
We really have, and I'm glad we're finally sitting down with it, because the piece that sparked today's conversation is from Ian Schaeffer, our company's COO. And what makes his perspective interesting is that he is not an advisor. He's an engineer. He builds and runs an AI assistant himself, every single working day. So when he writes about this, he's not writing as a skeptic or as a cheerleader. He's writing as someone who actually knows what these tools are doing under the hood.
And the reason he wrote it, from what I can tell, is that he keeps hearing the same three complaints from readers. People who tried AI, got burned, and walked away. Which, honestly, is a very understandable response.
Right, and before we get into what went wrong for those folks, let's just acknowledge who we're actually talking about. Because there's a stereotype that older adults aren't using this technology, and the numbers don't quite back that up. Pew Research did a survey of over five thousand adults in February of 2026, and they found that twenty-three percent of adults sixty-five and older have used an AI chatbot. That's roughly one in four.
One in four. That's not a fringe group.
Not at all. And AARP's own technology survey, which they fielded in the fall of 2025 among adults fifty and older, found that AI usage had nearly doubled in a single year, going from eighteen percent in 2024 to thirty percent in 2025. So the adoption is real and it's accelerating. But here's the catch in those same numbers: only seven percent use it daily, and only six percent describe themselves as confident with it.
So a lot of people tried it, weren't impressed, and didn't go back. Which raises the obvious question: why weren't they impressed? What happened when they tried it?
Ian Schaeffer's article is pretty direct about this. The most common complaint he hears is that the AI made up sources. People would ask a Medicare question, the assistant would cite a study or a government document, they'd go look for it, and it just... didn't exist. A completely fabricated citation.
Which is terrifying when you think about what's at stake. Someone's trying to figure out their Medicare enrollment window or their Part D penalty, and they're getting citations to documents that were essentially invented.
And the reason that happened is actually pretty fascinating once you understand it. These AI language models are built on enormous amounts of text, and the way they work is by predicting the most likely next word, or the most likely next sentence. When there's nothing connecting them to the live internet, they can't look anything up. They can only recall. And when they recall a source, they may just assemble a plausible-looking title, author, and date the same way they assemble any other sentence. It sounds real. It reads real. But it was generated, not retrieved.
So it's not exactly lying, it's more like... confidently filling in a blank it didn't actually know the answer to.
That's a good way to put it. And according to Ian Schaeffer, in 2023, that was essentially every AI assistant. That was the default state of the technology. What changed is that the major assistants added live web search with links you can actually click and verify. He mentions one of them describing a March 2025 release where it said it provides direct citations so you can easily fact-check sources. And the others added the same capability in roughly that same window, from late 2024 into 2025.
So there's a meaningful difference between a source the AI is remembering and a source it's actually pulling up in real time.
A huge difference. And Ian Schaeffer gives a test for this that I think is worth writing down. Ask your Medicare question, and then follow up with: give me the link to the government page you're relying on. Then open that link. Does the page exist? And does it actually say what the assistant told you it said? If the answer is no, you're either using an old tool or an old setting, and nothing it tells you about Medicare should move any decisions until a human has reviewed it.
I like that test because it's not asking you to become a tech expert. You're just checking whether a link works. Anyone can do that.
Exactly. It takes about sixty seconds and it tells you almost everything you need to know about whether to trust that particular tool.
Now, before we get too far, I want to bring up something in Ian Schaeffer's piece that genuinely surprised me. There was a study about how accurate AI actually is on Medicare questions, and the results were not what I expected.
This one really is striking. In June 2024, three researchers, Emily Langston, Neil Charness, and Walter Boot, put seventy-seven real Medicare questions to four different AI assistants and published the results in a journal called The Gerontologist. The two chat-based models scored ninety-seven point four percent and ninety-three point four percent correct. That's remarkable accuracy on a topic that trips up a lot of professionals.
But here's the number that really got me. How did actual Medicare beneficiaries score on those same questions?
Fifty-five point four percent overall. So the chat models were scoring in the nineties, and the people the questions were originally written for were scoring right around fifty-five. And to be clear, Ian Schaeffer is careful to say hold those numbers loosely, because all of these tools have changed since 2024. But the shape of that finding has held up in his own use.
It makes a certain kind of sense, though, doesn't it? Medicare is genuinely complicated. There are rules and deadlines and penalty structures that most people only encounter a handful of times in their lives. A well-designed AI that's ingested all the government documentation might actually have better recall of the general rules than someone trying to piece it together from a letter they got in the mail.
That's the thing. Ian Schaeffer puts it really plainly: a modern chat model knows what a Part D penalty is better than most of us do. But, and this is in the same article, the little speaker on the counter does not.
Oh, that part caught my attention. The voice assistants, the ones you just talk to from across the kitchen, scored how?
Thirty point three percent and two point six three percent. So one of them was essentially guessing at random, and the other wasn't much better. The title of that paper was actually a question: Are Virtual Assistants Trustworthy for Medicare Information? And their answer was, it depends entirely which one.
Which is such an important distinction, because I think a lot of people lump all of this together. AI is AI. But a voice assistant you're shouting questions at from across the room is a very different thing from sitting down at a computer with a chat-based assistant that can search the web.
Very different. And the category matters more than any specific product name, which is something Ian Schaeffer comes back to at the end of his piece. He's not going to name specific tools, partly because they change so frequently. But he's clear that what you want is an assistant that can search the live web and shows you links.
Okay, but I want to make sure we don't paint too rosy a picture here, because there's a section in Ian Schaeffer's article that's a real counterweight to all of that. The accuracy numbers are impressive, but he's also pretty upfront about some current failures.
He is, and I appreciate that he includes it, because it would have been easy to leave it out. In April 2025, the maker of the most widely used AI assistant published its own test results showing that their newer reasoning model invented facts about real people thirty-three percent of the time on one particular benchmark. That's compared to sixteen percent for the model it replaced. So the newer version was actually worse on that specific problem, and the company said more research was needed to understand why.
Thirty-three percent. On a test run by the very people who built it. That's not a fringe critic saying the tool is unreliable.
No, that's the company's own data. And Ian Schaeffer also mentions that in January 2026, the largest search engine removed its AI summaries for certain medical searches after a newspaper investigation found the summaries were giving lab value ranges that didn't account for age or sex. These are current products from some of the largest companies in the world. The problem got smaller. It did not go away.
Which leads to his second test, which I think is even more elegant than the first one.
This one is clever. Ask the assistant what the Medicare Part B premium will be for 2027. The correct answer today is that nobody knows. Medicare announced the 2026 premium, which is two hundred and two dollars and ninety cents, back on November 14th of 2025. The 2027 number arrives on roughly that same schedule every year. So if an assistant tells you the 2027 premium with total confidence, it is guessing. And the assistant you actually want is the one that says, that hasn't been announced yet, here is the 2026 figure, and here is where it came from.
The willingness to say I don't know yet is a sign of trustworthiness. Not a weakness.
In Medicare planning, it might be the most important sign. Because if it will confidently fabricate a future premium, what else will it confidently fabricate?
Now I want to get into the third complaint Ian Schaeffer addresses, because I've heard this one from people too. Either the AI argues with them constantly, or it agrees with absolutely everything they say. And he explains why both of those things happen.
Both failures come from the same dial being turned too far in one direction or the other. On April 29th of 2025, the most widely used assistant's maker rolled back an update within days of releasing it, because the assistant had started offering what one reporter described as uncritical praise for virtually any user idea, no matter how impractical, inappropriate, or even harmful. The company's own explanation was that it had leaned too hard on short-term user approval signals and built a tool that leaned too far into affirmation without discernment.
And Ian Schaeffer specifically says, read that again with Medicare in mind.
Because the stakes are real. He gives a specific example: if you tell the assistant your Part B has no premium because a neighbor said so, you need it to say no, that's not right, and show you the page. That is not the AI being argumentative. That's the AI doing the job correctly.
There's a difference between an assistant that corrects you with a source and one that just digs in on a wrong answer with no evidence at all. The first one is useful. The second one is just frustrating.
And that's the old failure mode, the one from 2023. The current design goal, at least on paper, is an assistant that pushes back but brings you the source when it does. Which is why Ian Schaeffer has a third test: tell the assistant something you know is false about Medicare, and see if it corrects you politely with a link. If it just agrees with you, put it down.
Three tests. Check the source link. See if it knows what it doesn't know yet. And make sure it will correct you when you're wrong. Those are not complicated, and you don't need any technical background to run them.
That's the core of what Ian Schaeffer is arguing. The habit matters more than the tool. He uses a phrase in his article that I keep coming back to: an answer I cannot trace is a draft, not a fact. In his own work, the assistant gives him a paragraph and a link, and he opens the link before that paragraph goes anywhere. That is the whole difference, he says, between people who get burned by these tools and people who get real work out of them.
And he's pretty clear it's not about age or intelligence. It's just about whether anyone checks before the answer counts.
Which matters enormously in Medicare, because the cost of getting it wrong isn't abstract. He's pointed this out throughout this series: a Part D late enrollment penalty runs for life. The window to get a Medigap plan at 65 with guaranteed acceptance closes and locks behind you. Those are permanent consequences from a misunderstanding of the rules.
So let's talk about how to actually use this thing well, because Ian Schaeffer does give some guidance on that. It's not just here's what's broken and good luck.
Right, and the framing he uses is a good one. Use it the way you'd use a well-read friend. Ask it to explain a letter you got in the mail. Ask it to define a term. Ask it to help you draft the questions you're going to bring to a professional. That's the lane where it's genuinely useful.
And the part it doesn't belong?
Don't ask it which plan to pick. He's explicit about this. The honest answer to that question depends on your county, your specific doctors, your specific prescriptions, your specific pharmacy, and the plan list for your area for the coming year. An AI assistant, however good it gets, cannot know those things about you.
He flags that the next piece in this series is going to be exactly about that, the seven things no AI can know about your Medicare. That's worth calling out now, because that framing is helpful even before you read it. There's a category of Medicare question that is fundamentally personal, and no general-purpose tool can substitute for someone who knows your situation.
And that brings us to the last section of Ian Schaeffer's piece, which is about what a Certified Medicare Planner does differently. He's not dismissing the AI tools. He's placing them correctly in the ecosystem.
What struck me in that section is how he describes the advisor's approach in terms of the same three tests. They already run those tests, on your behalf, for a living.
They won't tell you a 2027 premium figure that hasn't been published. They pull the government page itself, not their memory of it, and they know which page. And when a client arrives certain of something a neighbor or a chatbot told them, they push back with the source.
But what they add, that no AI can add, is the part that isn't on the internet anywhere. Your actual doctors in next year's plan directory. Your actual prescriptions at your actual pharmacy. The plan list for your county the day it goes live.
And Ian Schaeffer includes a line that I think is really encouraging for anyone who has done their homework with these tools before meeting with an advisor. He says they love a client who did the reading. Bring them what the AI told you. That's a great starting point for the conversation.
I love that framing because sometimes people feel a little sheepish admitting they looked things up on AI first, like they should have just waited to ask a professional. But no, bring it. Walk in with your AI summary, your questions, your concerns, and let the advisor work through it with you.
For anyone who wants to connect with the team, the number is 602-281-3898. And if you want to get into a room with someone in person, there are workshops coming up. The article mentions Friday, October 2nd. Foothills Library in Glendale at 10:30 in the morning, and Mustang Library in Scottsdale at 2:30 in the afternoon. At no cost to you. You can register at 123easymedicare.com/medicare-workshop or call 877-220-1089.
And just to be clear, those workshops are presented by a brand of American Retirement Advisors. They are not sponsored by Medicare, the Centers for Medicare and Medicaid Services, the Social Security Administration, or any government agency. The libraries hosting them are not affiliated with the event. It's all educational.
One thing I want to circle back to before we close, because I think it deserves a moment. The pattern Ian Schaeffer is describing, where someone tries an AI tool, gets a fabricated citation or a confidently wrong answer, and then writes off the whole category forever, that response is understandable, but it may be costing people something.
Because the tool they tried in 2023 and the current tools are not the same thing.
Not remotely. And the three tests are a way to verify that for yourself without taking Ian Schaeffer's word for it, or our word for it. He's very explicit about that. He's not asking anyone to trust him on this. He's giving you the test so you can see it yourself.
I appreciate that about the piece. He's an engineer who works with this every day, and he's saying here's how I evaluate the tool, and you can use the same method. That's a lot more useful than just trust me.
And the underlying message, underneath all of it, is really about calibration. Use AI for what it's good at: explaining, defining, organizing your thinking, helping you show up prepared. And use your Certified Medicare Planner for what no AI can do: taking all of that general knowledge and running it through every last particular of your situation -- all those personal details we talked about earlier that no algorithm can just look up about you.
There's also something worth saying about the emotional side of this, because Medicare decisions carry a lot of anxiety. People are afraid of making a mistake that costs them for years. And I think that anxiety is part of why AI tools are appealing in the first place. You can ask a question at ten o'clock at night when you're worried, without feeling like you're bothering anyone.
That's a completely legitimate use. Ask it then, get the explanation, get the draft, get the questions organized. And then bring all of that to a professional who can tell you what it means for your specific situation. The two things work together.
And if you fall into that camp Eddie just described -- the ones who had a frustrating early experience and quietly closed the door on all of it -- Ian Schaeffer's piece is worth reading with fresh eyes. Not because you have to use these tools, but because understanding why that earlier experience went wrong is genuinely useful context. The article is from our COO, it's in the Healthcare in Retirement series, and it's the kind of piece you might find yourself sharing with a friend who had that same discouraging run-in.
The specific numbers and details from the studies he cites, the Pew data, the AARP survey, the Gerontologist paper, all of that is worth a close read. We hit the highlights today but the article has more texture than we can cover in one conversation.
One last thought before we wrap up. If you are navigating Medicare right now, or you have a family member who is approaching that sixty-five window, please don't rely on a single source, whether that's AI, a neighbor, something you read online, or even what we've talked about today. Sit down with someone who does this for a living and who can look at your actual situation. The advisors at American Retirement Advisors are at 602-281-3898, and those October 2nd workshops are a wonderful way to get a lot of your questions answered in person, at no cost, in a setting where you can just listen and learn. That's what we want for every listener: you go into Medicare informed, prepared, and with someone in your corner who knows the rules cold.
Well said. And thank you to Ian Schaeffer for writing something that was genuinely useful and genuinely honest about both the promise and the limits of these tools. It's the kind of piece that respects the reader.
Thank you for spending this time with us. We'll see you in the next one.