Is Medicare already using AI to decide what it pays for? Arizona, now
September 23, 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. I'm here with Eddie in the studio today, and we are diving into something that I think is going to surprise a lot of you, because it surprised me. We are talking about Medicare and AI, specifically: what can an AI assistant actually get right about your Medicare situation, and where does it go completely off the rails? Eddie, you read Ian Schaeffer's piece on this?
I did, and I want to say for the record, I'm a numbers guy, Ian Schaeffer is an engineer, so this article spoke to me on a cellular level.
High praise from the numbers guy.
The highest. But what he found is genuinely useful, because most people either trust AI completely for this stuff or they've given up on it entirely. And his argument is that both reactions miss the point.
Right, because he's not saying AI is useless. He's saying there's a specific category of Medicare question where even a really good AI cannot give you a correct answer, and the difference between a helpful tool and a dangerous one is whether it tells you that.
Exactly. And he found seven specific things that fall into that category. Seven Medicare questions where the AI is going to let you down, no matter how sophisticated it is.
Okay, walk me through the logic before we get into the list, because I want to understand the why.
So he breaks it into three buckets. First: information that isn't public yet, because the government publishes it on a schedule and we haven't hit that date. Second: information that's public but only applies to your specific county or your specific state, so a national answer is technically accurate and completely wrong for you. And third: stuff that only exists in your own file. Your records, your months without coverage, your particular life situation.
That third one is the one that got me. Because the AI can search the whole internet and it still can't find information that's just... in a filing cabinet somewhere.
It doesn't even know to look for it, because it doesn't know what it doesn't know about you. But let's start with the one that I think is the most vivid illustration of the problem, because Ian Schaeffer actually built a little test around it.
Oh, the two-number test. Yes, tell people about this.
So he writes the article on September 21st, 2026. At that point, the 2027 Part B premium has not been announced. Medicare typically publishes it in November. Doesn't exist yet.
But the 2027 Part D base beneficiary premium is already out there. They released that one in July.
Right, $41.33, up from $38.99. That number is out there, sourced from a government memo, with the arithmetic shown. So his test is: ask the AI for both numbers.
And a careful AI gives you the Part D figure with the source, and then says, sorry, the Part B figure hasn't been announced yet.
And a confident AI delivers two numbers in the exact same tone of voice. One real, one fabricated. And you cannot tell the difference from how it sounds.
That is unsettling. Because it's not like the AI is being shifty about it. It just doesn't know that one of those numbers doesn't exist yet.
Ian Schaeffer puts it perfectly. He says the AI doesn't know the difference between remembering and looking up. It produces an answer. It sounds right and it's formatted like one, and one came from a government memo and one came from nowhere.
And the person sitting at the keyboard is going to use that number.
Potentially to calculate a penalty, or decide when to enroll. Which brings me to the piece of this article that made my engineering brain very happy.
The penalty table.
So it's 1% of the national base beneficiary premium for every full month you went without creditable drug coverage after your window closed. And it stays on your premium for as long as you have Part D.
Which is a long time.
Potentially the rest of your life. Now here's where it gets interesting. He lays out the math for two different scenarios, twenty months without coverage and thirty-six months.
And those numbers are already pretty striking on their own. But what's the thing about the table that made you stop?
There are four columns. Two of them can be filled in today, from public government memos, because we know this year's base premium and next year's. But the Part B penalty column for 2027? It just says 'cannot be computed yet.' Because that premium hasn't been released.
So literally half the table is blank.
On purpose. Because filling in a number would be a guess. And then there's the row that stopped me completely: 'your months.' It just says 'only you know the first number.'
Because the AI has no idea how long you went without coverage.
Right. And he says an assistant that produces a dollar figure for your penalty without asking about your months, and without telling you which year's base premium it used, has produced a number that looks like arithmetic and is actually a guess dressed up as one.
Okay, that line I'm going to be thinking about for a while. It sounds so precise and it's nothing of the kind.
It's a good line for an engineer.
It really is. So that covers the 'not public yet' problem. What about the second bucket, the stuff that's public but doesn't apply to you specifically?
This is where geography becomes everything. Medicare Advantage plans and drug plans are sold county by county. Not state by state. County by county.
And Medicare's own Plan Finder won't even show you a plan until you type in a ZIP code.
Because a plan available two towns over might not be available to you. So if you ask an AI what plans are in your area and it doesn't ask for your ZIP code, the answer is useless even if every fact in it is technically true.
He mentions in the article, sort of as a reminder, that a plan with no premium of its own still means you're paying Part B. So 'zero premium plan' doesn't mean free.
Right, and that's a conversation for another part of the series. The bigger point here is just that 'there are lots of zero-premium plans available' is both accurate and completely unhelpful unless it's specific to your county.
Okay, what about Medigap? Because he gave a really concrete example there that I want people to hear.
So federal law sets a floor for guaranteed-issue rights, meaning situations where you can't be turned down based on health. But states can add to that floor, and they vary a lot.
And he gives Nevada and Arizona as the example.
It has a birthday-month window, an annual window where you can switch Medigap plans without answering health questions. Arizona has not passed an equivalent law. Ask an AI 'can I switch Medigap without health questions?' and the correct first response is 'which state do you live in?' If it doesn't ask that, you might be relying on one state's rules when you live in the other.
And it goes the same way in reverse. Someone in Nevada could think they have fewer options than they do.
Both directions are a problem. And this is a case where I'd really say, the specific rules in your state are absolutely a question to take to one of our advisors at American Retirement Advisors, because the state-by-state detail is exactly what they track.
And then there's the drug price one, which honestly baffled me a little bit.
This one I love. So you're on a Medicare drug plan. You go to your pharmacy, you get a quote. Your neighbor is on the exact same plan, goes to a different pharmacy, gets a different quote.
Same plan.
Same plan. Because the regulation allows plans to charge different copays at preferred versus standard pharmacies. So 'what will my prescription cost on this plan' has no answer without also specifying which pharmacy counter you're standing at.
An AI can't know which one you use.
Not unless you tell it. And even then, it would need current information on that specific plan's preferred pharmacy network for your area.
So we've covered six of the seven. The one we haven't talked about is the special enrollment period question.
Which is the most personal one. It's triggered by things that happen in your life: you moved, your employer coverage ended, your plan is leaving the area. The regulation lists the triggers, and an AI can absolutely read you that list.
But it can't know which one is yours.
It can't know if your situation qualifies. It can tell you what a qualifying situation looks like. You have to match your life to the list, and if you get that wrong, you miss your window.
Which is a very expensive mistake.
Lifetime penalty, yeah. That's the kind of thing that follows you forever.
Okay so we've established the seven things. But Ian Schaeffer doesn't just leave you with 'AI is bad, don't use it.' He actually gives you a way to use it better.
Three questions. Ask them before you believe any Medicare number from an AI. First: what date is this figure from, and is next year's version published yet? Second: which state and county is this answer for? Third: what would you need to know about me to be sure?
That last one is the one I'd forget to ask.
Most people would. But a careful assistant, he says, answers all three plainly and often volunteers them. If it does, you have a research partner. If it breezes past all three, he puts it beautifully: it's a very well-read stranger with no calendar, no map, and no file on you.
Enjoy the conversation. Do not enroll on it.
That is his closing line on that section and it is a good one.
So what does a human actually do with these seven things? Because he talks about what a Certified Medicare Planner does differently.
That checklist is the job. They know the Part B fact sheet comes out in November and they won't quote you a 2027 number in September. They pull your county's plan list the morning it goes live, not from memory. They look up your state's Medigap rules from the state.
And the penalty piece.
They count your uncovered months from your actual records and run it with the right year's base premium. Twice, once for this year and once for next. Because the base changes.
And he has a suggestion I thought was really smart, which is bring your AI conversation to that first meeting.
Right. Show the advisor what the AI told you and ask the three questions together. He says that's a better first meeting than most people have.
Because you've already done some homework and you know what to be skeptical of.
You walk in knowing that the number it gave you for the Part B premium might not exist yet. That's a useful thing to know.
I want to circle back to something for a second, because I think the biggest takeaway from Ian Schaeffer's piece is not 'distrust AI.' It's that the AI isn't the dangerous part. The dangerous part is not knowing where its edges are.
He's very clear that it's not the tool's fault. It doesn't know the difference between a number it remembered from training and a number it actually looked up. It's not hiding that from you. It just doesn't have that self-awareness built in.
Which is why your job as the person at the keyboard is to ask those three questions and watch how it responds.
If it hedges, cites sources, asks where you live, admits it doesn't know something, that's the tool working correctly. If it just hands you two confident numbers when one of them doesn't exist yet, that's the tell.
The tell. I love that he calls it that.
He says it's the most useful thing in the article. And I think he's right.
Speaking of getting answers from actual humans, there are workshops coming up. This Friday, October 2nd. Two locations.
Foothills Library in Glendale at 10:30 in the morning, and Mustang Library in Scottsdale at 2:30 in the afternoon. No cost to you to attend.
You can register at 123easymedicare.com/medicare-workshop, or just call 877-220-1089. And if you want to talk to an advisor directly about any of this, the number for American Retirement Advisors is 602-281-3898.
And if you go in with Ian Schaeffer's three questions already written down, I think you're going to have a really productive conversation.
This whole series he's been writing, by the way, is worth reading. He thinks like an engineer, which means he shows his work, he cites the actual regulations, and he doesn't pretend things are simpler than they are. This piece in particular, part fourteen, is one I'd bookmark.
Especially before open enrollment. If you're going to use an AI to research your options, read this first and you'll know exactly what to watch for.
That's the whole point, right? It's not about avoiding the tool. It's about using it with your eyes open. And then sitting down with someone who has your file, knows your state, and can count your actual months. That's where the real answer lives. Go find that person. You deserve an answer that's actually about you.