Health and AI,
for the people who teach it
Three questions your students are going to bring you, and answers you can hand them.
David Marek, DropDev. Ninety minutes: one hour together, thirty for whatever you want to ask. Everything on this page is yours to keep, in English and in Spanish.
It is already in the room when your student sees a doctor
Your students are doing more of their health online every year: the patient portal, the telehealth visit, the symptom they looked up last night. And they don't have to open a chatbot for AI to be part of it. It's already there in four ordinary places they pass through every time they see a doctor, and most of them have no idea.
The reply from your doctor’s office
At one large health system, clinicians used an AI-written draft to answer patient portal messages about 20 percent of the time. In a separate look at more than 55,000 messages, when a draft contained an error it was still sent completely unedited 35 to 45 percent of the time. JAMA Netw Open 2024;7(3):e243201. Frontiers in Digital Health 2025.
The answer above the search results
The summary at the top of a search page is generated. Your students read it as "what the internet says," because it sits where the answer has always been, and it does not look like an opinion.
The note your doctor typed
Ambient AI scribes listen to the visit and write the note. Across five academic medical centers and about 1,800 clinicians, the ones who chose to adopt them recorded roughly 16 fewer minutes per eight hours of patient care. That is an observational study of self-selected adopters, not a trial. The actual randomized trial of two commercial scribes found about a 10 percent cut in documentation time. JAMA, 2026, prospective observational. RCT: NEJM AI, DOI 10.1056/AIoa2501000.
The decision about whether Medicare pays
Medicare began an AI-assisted prior authorization pilot called WISeR on January 15, 2026, in six states. Texas is one of them. In May 2026 the GAO found it meets the legal definition of a rule, and repeal resolutions were introduced. As of today it has not been repealed. CMS; GAO 5/12/2026; CRA resolutions 5/20/2026.
None of that announced itself. There was no banner. This matters for your classroom because you can't teach someone to be careful with a thing they don't know is in the room. The first skill is noticing.
Being wrong here costs more than money
Your students are already asking these tools about their bodies. The only question is whether anyone prepared them first.
Here's the finding that should slow all of us down. In a randomized study of 1,298 people, the chatbots by themselves identified the right condition 94.9 percent of the time. But when actual people used those same chatbots on the same situations, they got under 34.5 percent, while the group who used no AI at all got 55 to 67 percent. The people without the tool got it right nearly twice as often. The tool can be right and the person can still come away worse off, because people don't type in the whole story and don't always act on the part the tool got right. Bean et al., Nature Medicine 2026;32:609-615.
Is this real?
Not "is AI good or bad." That one has no useful answer. Your students need a narrower question: is this specific thing in front of me true?
It is not looking anything up
If your students take one idea home, make it this one. Almost every mistake with these tools comes from a wrong picture of what the tool is doing.
What people think it does
They picture a very fast librarian. Someone who goes and finds the true answer in a big trusted book, and reads it back. That is the mental model almost everyone starts with, and it is the wrong one.
What it actually does
It predicts the next likely word, over and over, based on patterns in an enormous amount of text. It is producing what an answer to your question usually looks like. Most of the time, what an answer usually looks like and what is true are the same thing. That is why it works at all, and also exactly why it fails without warning.
This is why it can hand you a study that doesn't exist, with real-sounding authors, in a real journal, with page numbers, formatted perfectly. It isn't lying to you. It's producing the shape of an answer, and a citation has a very predictable shape. When researchers checked the references ChatGPT produced for medical questions, 41 of 59 were completely fabricated in one study, and 54 of 115 in another. They looked entirely normal. Gravel et al., Mayo Clin Proc Digit Health 2023. Bhattacharyya et al., Cureus 2023;15(5):e39238.
Confidently wrong, and getting less hedged
Most of these are normal-use numbers, nobody jailbroke anything, and I'll flag the one stress test when we reach it. This is just what happens when people ask health questions the way people do.
Here's the shift that matters most, and it's the uncomfortable one: confidence in the writing is no longer a signal of anything. As these models got more accurate, they got less careful about saying so. The share of answers carrying a safety disclaimer fell from about 26 percent in 2022 models to under 1 percent in 2025. Better answers, and the warnings nearly gone. Sharma, Alaa and Daneshjou, npj Digital Medicine 2025.
Try it
Real, or made up?
Call it out loud before the reveal. Some of these are true and sound absurd. Some are false and sound reasonable. That gap is why "use your judgment" isn't something you can teach.
"People who used an AI chatbot to work out a medical problem did worse than people who used no AI at all."
Heard from: a news story about a study"This fruit repairs your vision while you sleep." Said by a doctor in a video, in Spanish, with a white coat and a clinic behind him.
Heard from: YouTube, forwarded by a family member"AI checking mammograms alongside radiologists found more cancers, in a randomized trial of over 100,000 women."
Heard from: someone at a health conferenceA video of a well-known Spanish-language TV health personality, in his own voice, recommending a product he has never mentioned before.
Heard from: a Facebook ad"I asked the chatbot and it gave me a study to back it up. It had the authors, the journal name, and the page numbers."
Heard from: a student, proud of checking"A honey-based supplement reverses memory loss and Alzheimer's." Advertised with video of famous people recommending it.
Heard from: an ad that followed them aroundDon't study the page. Leave it and look elsewhere.
The instinct is to stare harder at the thing, hunting for the flaw. Fact-checkers do the opposite: they leave the page right away and check what everyone else says about the source. That's lateral reading, and it's the habit that transfers best.
Four questions, in order
- Who is saying this? Is there a real name, a real hospital, a real agency? Or nobody at all?
- Why are they saying it? Are they telling you something, or selling you something? If there is a product at the end, start over.
- Does anywhere you already trust say the same thing? This is the whole trick. Leave the video. Go look.
- When was it written? Health guidance changes. A page with no date is a page with no accountability.
For a claim that came from a chatbot
- Copy the title of the study and search it. Not the claim, the title. If the paper is real it appears immediately.
- If nothing comes up, or a different paper comes up, the citation was generated. This happens constantly and it is not a sign of a broken tool.
- Ask it a second time in a new chat. If the answer changes materially, you have learned something about how much it knows.
- For anything that would change a medication, a dose, or a decision to go to the emergency room: the chatbot is not the last step. A person is.
You're going to teach this, so you deserve to know how solid the ground is. Lateral reading has real evidence behind it, and it's the habit that transfers best to a tablet. Lateral reading: Wineburg and McGrew, Stanford History Education Group.
Where to send someone instead
"Don't trust that" is only half an instruction. Your students need somewhere to go. All of these were live, and genuinely in Spanish, when I checked on August 23, 2026.
MedlinePlus en español
- The National Library of Medicine. Written in Spanish, not machine-translated from English.
- The drug and supplement lookup at medlineplus.gov/spanish/druginformation.html is the direct answer to "is this miracle cure real."
- This is the one to make your students bookmark. If they only keep one, keep this.
Cancer.gov en español
- National Cancer Institute, genuine Spanish translation.
- Has a live phone and chat consultation line: 1-800-422-6237.
- Useful precisely because cancer claims are what the fake doctor videos push hardest.
Factchequeado
- Spanish-language fact-checking built for US Latinos, IFCN-affiliated.
- Actively checks health claims and AI fake-doctor content, which almost nothing else in Spanish does.
- Good place to send someone who arrives with a forwarded video.
Close to home
- Texas DSHS Public Health Region 9/10, the regional office headquartered here, covering the 35-county border region.
- spanish.umccares.org, UMC El Paso patient education in Spanish.
- AARP en español on health fraud, written for the older adults these scams target hardest.
Is this a scam?
This is the one that scares people first, and it deserves a straight answer, even where that answer is "nobody knows yet."
A voice used to be proof
For your students’ whole lives, a familiar voice on the phone meant it was that person. That stopped being true recently. Nobody sent a notice.
Most fraud still has no AI in it at all. The FBI tags what victims report, it doesn't confirm AI was used, and victims often can't tell. So when someone gives you a number for how much fraud is now AI, they're guessing. The right thing to teach isn't "AI is everywhere," it's "your ears are no longer proof, and here's what to do instead."
Walk it through
It is 11 at night and the phone rings
One choice at a time. There's a right answer, but watch how reasonable the wrong ones feel in the moment. That's the whole design of the scam.
The number on the screen is local. You answer. It's your granddaughter's voice, and she's crying so hard you can barely make her out. There was an accident. She's at the police station. She says: "Grandma, please don't tell mom."
Three habits, and they beat every version of this
Your students don't need to detect AI. Detection is a losing game. They need a routine that works whether the voice is real or not.
Hang up. Call back.
- It does not matter how real the voice sounds, because you are not going to argue with the voice. You are going to end the call.
- Call back on a number you already had: from the bill, the back of the card, the official website. Never the number they gave you and never the one on your screen.
- Caller ID can be faked. A call can show your bank’s real number, or a real government number. That is called spoofing.
Pick a family word
- One unusual phrase every family member knows. Not a name, not a birthday, not a pet, not anything posted on Facebook.
- Anyone calling with an emergency that involves money has to say it first.
- This is the rare defense that gets stronger as the fakes get better, because it is a secret rather than a judgment call.
Watch the payment, not the story
- Gift cards, cryptocurrency, wire transfer, a courier coming to the house for cash. No real agency or bank ever asks for these.
- If the story is urgent and the payment is strange, that combination is the scam, every time.
- You do not have to work out whether the story is true. You only have to notice how they want to be paid.
Where to report it
- ReportFraud.ftc.gov, and in Spanish, ReporteFraude.ftc.gov.
- Texas Attorney General: 1-800-621-0508, and texasattorneygeneral.gov/es in Spanish.
- IC3.gov for internet crime. Be aware: the FBI’s complaint form is English only, which is a real barrier for your students. Start them at ReporteFraude instead.
Where does what I type go?
Here's a question your students almost never think to ask, and it matters. Telling your doctor something and typing it into a chatbot aren't the same act. They're not even under the same law.
HIPAA does not follow you into the app
HIPAA binds your clinic. It doesn't bind a chatbot, because a chatbot isn't your doctor and doesn't work for your doctor.
That's not a loophole somebody is abusing. It's how the law is written. HIPAA covers covered entities, meaning health plans, clearinghouses and providers, plus the business associates they hire. A general chatbot is neither. So what protects you there is not federal health-privacy law, it is the company’s own privacy policy plus general consumer-protection law. That is a real but much weaker and differently enforced bar. HHS, Covered Entities and Business Associates; 45 CFR 160.103.
Try it
The same sentence, two doors
Pick something a student might say, then follow it down both paths. The words are identical. What happens next is not.
- It becomes protected health information the moment it enters your chart.
- Only people involved in your care and payment may look at it. The clinic has to be able to tell you who saw it.
- It cannot be sold, and it cannot be used to advertise to you without your written say-so.
- If it leaks, the clinic is legally required to notify you, and HHS can fine them.
- It becomes ordinary company data governed by a terms of service, not by health-privacy law.
- On a normal free account it is used to train the model by default at every major vendor. You have to go turn that off.
- A human reviewer may read it. Google’s own page says: "Please don’t enter confidential information that you wouldn’t want a reviewer to see."
- Delete does not always mean gone. A court order in the New York Times case forced OpenAI to preserve chats users had asked to delete.
Use it with the lid on
A blanket ban is bad advice and your students will ignore it. What works is a line between what's safe to type and what isn't, plus the one setting that matters most.
Safe to type
- "What does high blood pressure mean?"
- "What are the side effects of ibuprofen?"
- "Help me write a message to my doctor's office."
- "What questions should I ask at this appointment?"
- "Translate this medical term into Spanish."
Do not type
- Your name together with a diagnosis, a medication list, or a test result.
- Mental-health history, yours or anyone else's.
- A child's or family member's health, without them knowing.
- Anything you would not want read back to you by a stranger, a company employee, a court, or an insurer.
The one setting worth changing
- ChatGPT: Settings, then Data Controls, then turn off "Improve the model for everyone." Or use Temporary Chat.
- Gemini: myactivity.google.com/product/gemini, then turn off "Keep Activity." Or use a Temporary Chat.
- Claude: Settings, then Privacy, then turn off model training. Or use Incognito chat.
- Copilot: in-app privacy settings, turn off the training toggle.
Meta AI is the exception
- As of December 16, 2025, Meta uses AI chat interactions to personalize content and ads, and US users cannot opt out of that sharing.
- Meta says it will not use a short list of sensitive topics, health included, to target ads directly. The interactions are still used and may still be reviewed.
- Practical advice: do not put health details into Meta AI at all. The /reset-ai command clears the visible chat, it does not undo what was already used.
Down the line
Here's what's coming, without the hype. This field is producing real results and real failures at the same time, and telling them apart is the skill.
So what do I expect? Not a robot doctor. What's arriving is quieter and closer to your students: AI writing the note, drafting the reply, sitting in the insurance decision, answering the health question at 2 a.m. when the clinic is closed.
The local ground you're standing on
One thing this room has that no national guide does: the numbers for your own county. This is the ground your students are actually on.
The gap here is real, and it is measurable
- 14.75 percent of El Paso County adults 65 and older have no computer of any kind at home. Statewide it is 7.87 percent. Nearly double.
- 77.02 percent of adults 65 and older here have broadband, against 84.54 percent statewide. Roughly seven and a half points behind.
- 13.29 percent of households here are smartphone-only, no laptop, no desktop, no tablet. Higher than Texas and well above the national rate.
Two minutes
Tell me how this actually landed
You're the ones who have to teach this. If a piece of it doesn't survive contact with a real classroom, I'd rather hear it from you than keep presenting it. Everything here is optional and goes straight to me.
The form needs JavaScript, which looks like it is switched off. Email me instead at david@dropdev.co, or text (915) 234-1444. Same inbox either way.
Got it, thank you.
This genuinely changes the next version. If you left an email I'll write back. If something else comes up later, david@dropdev.co or text (915) 234-1444.
You're teaching confidence, not fear
Your students aren't going to stop using these tools, and they shouldn't. The three questions aren't here to scare anyone off the portal or the telehealth visit. They're here to give people the footing to take part in their own care, and to know, when something feels off, exactly what to do: stop, and check with something you already trust. A source. A number you already had. A real person. You're the ones who turn "scared of getting scammed" into "I can handle this."
David Marek · DropDev · david@dropdev.co · (915) 234-1444
This page stays up. Send it to anyone who could not make it, and switch it to Spanish for the ones who need that.
The claims that carry the talk, and where they came from
You'll teach this to people who trust you, so check me first. These are the claims the talk rests on. Where a finding is contested, an allegation, or a single small study, it says so. Everything else carries its source inline.
People using chatbots did worse than people using none
Randomized, 1,298 participants. Chatbots alone identified the right condition 94.9 percent of the time. Participants using them identified conditions in fewer than 34.5 percent of cases, against 55 to 67 percent for the control group, who had 1.76 times higher odds (95 percent CI 1.45 to 2.13, p less than 0.001).
Spanish-language AI fake doctors
60 Spanish-language YouTube channels, 4,254 videos, more than 203 million views. Separately 70 TikTok profiles with about 6 million followers impersonating 9 identified providers.
Elder fraud totals, and the AI share
201,266 complaints and $7.75 billion in reported losses among adults 60 and older in 2025, up 59 percent. Texas ranks third nationally. AI-tagged losses in that group were $352.5 million.
HIPAA does not cover consumer chatbots
HIPAA binds covered entities and their business associates. A general-purpose chatbot is neither.
Texas is investigating AI mental-health claims
The Texas Attorney General opened an investigation into Meta AI Studio and Character.AI, alleging they presented as confidential mental-health support while interactions were logged and used for advertising and development.
AI mammography, the win and the failure
MASAI, over 105,000 women, randomized: sensitivity 80.5 percent against 73.8 (p=0.031). A separate triage trial cut workload 63.6 percent but raised recalls 14.8 percent.
One in three adults ask AI about health
32 percent of US adults have used an AI chatbot for health information in the past year. 77 percent are concerned about the privacy of medical information shared with AI, and so is 65 percent of the narrower group who entered personal medical information into one.