Welcome to the DRIVE series, complete with a gimmicky name for my pleasure. At the end of every month, I’ll provide some combination of discoveries, resources, ideas, values, or experiments involving AI, technology, and healthcare. Let’s go!

If March was busy, April was even busier. So I didn’t write any posts. However, it was creatively fruitful so I have a bunch of cool projects I’ll be releasing in the next few months. Here are my favorite links and articles in no particular order.

Project Glasswing

• By Anthropic

Assessing Claude Mythos Preview’s Cybersecurity Capabilities

• Red Anthropic

Anthropic’s New AI Model Sets Off Global Alarms

• By Paul Mozur and Adam Satariano - The New York Times

We start with the most important news of the month. Anthropic’s announcement that they have developed a general-purpose, unreleased frontier AI model that has such high coding capability, that it can discover and exploit software security vulnerabilities. They call the model Mythos and claim that it has already discovered thousands of security vulnerabilities in “every major operating system and web browser.”

Some of these vulnerabilities are called zero-day vulnerabilities, or software flaws that software developers were unaware of, that were decades old. Anthropic has not released the model for widespread use, but has instead convened a consortium of companies that include Apple, Google, Amazon, JPMorganChase, NVIDIA, Microsoft, Linux, and others to use the model on their own systems to protect themselves before models like these become more widespread. They call the initiative Project Glasswing. Unfortunately, I have not seen any information about a major healthcare company or hospital being involved.

Obviously this is a big deal in all industries, but especially in healthcare, where the security systems tend to be fragmented, aging, and prone to hacks and exploits. In fact, the FBI recently reported that the healthcare industry was the top target for ransomware and other cyberthreats in 2025. You can check out their 2025 Internet Crime Report here.

The Labor Department Wants to Teach You To Use AI. Here’s What We Found

- By Huo Jungnan - NPR

Make America AI Ready - U.S. Department of Labor

CS309: The Essentials of AI for Life and Society - University of Texas

• This NPR article introduced me to a couple of AI literacy courses for the masses by both the U.S. Department of Labor and the University of Texas. On preliminary review, they seem helpful and educational, especially the one from the University of Texas which is a 3-credit course with video lectures that they have made freely available.

Pangram

• I recently discovered Pangram, which is a tool to help detect AI-generated content.

We Don’t Really Know How AI Works. That’s a Problem.

• By Oliver Whang - The New York Times

The Urgency of Interpretability - By Dario Amodei - CEO of Anthropic

• Interpretability is a concept in compute science that suggests we should treat AI more as a natural phenomenon rather than a human technology. Essentially, many AI models are black boxes — we do not necessarily know how they work or how they derived their answers because their output is dependent on millions of mathematical calculations. In science and medicine, this is a problem because how can we test a suggested treatment or cure if we cannot validate it with the scientific method? Unfortunately, the AI models cannot explain how they derived their answers. Researchers have found that the models make up explanations. Understanding the limitations is essential, especially if we want to use it in medical research and clinical application.

I Uploaded My Bloodwork to AI. Am I oversharing?

• By Nicole Nguyen - The Wall Street Journal

• The author explores different AI chatbots and the insights they give her when she shares her medical information. While this is tempting and can potentially be educational, the chatbots do not offer healthcare-level security and privacy protections and we also do not know how this information may be used by future versions of the technology.

The ChatGPT Symptom Spiral

• By Sage Lazzaro - The Atlantic

• Another article warning about the dangers of sharing your medical information with AI chatbots. This time, the fear is the crippling health anxiety the AI chatbot might induce if someone is looking for answers about their health without proper perspective.

Why Scientists are Nervous About Fungi

• By Gabrielle Emanuel - NPR

• Antimicrobial resistance is worsening and we are not developing drugs to treat resistant bacteria and fungi fast enough.

Under 40 and Diagnosed With Cancer

• By Nina Agrawal - The New York Times

• While the longevity boom seems to be receiving all the hype, there is a troubling trend of cancer becoming more common in younger people (those under 50) in the past few decades. This piece places human faces and stories to the statistics. It is a reminder that as much as we try, we do now always have control.

You Can Order Your Own Blood Work Now. Interpreting the Results is Another Story

• By Kate Cunningham - NPR

• With increasing use of AI chatbots, the cultural obsession over youth and longevity, and trends like DIY medicine, more and more people are getting self-tested. One of the main attractions is transparency in price. One of the main challenges is interpreting their own results.

A ‘Barbaric’ Problem in American Hospitals is Only Getting Bigger

• By Elisabeth Rosenthal - The Atlantic

• Here’s news every healthcare professional and fans of The Pitt knows: the problem of emergency department boarding, or patients waiting in the ER for an inpatient bed to become available, is getting worse. Unfortunately it will continue getting worse as the shortage of healthcare clinicians worsens, the population ages, and the masses become sicker.

I Saved $800 on My Medical Bills With an Easy Trick

• By Alex Olgin - Slate

• The hospital bill that you receive is not the final word. Rather it is the starting point of negotiation.

9 Reasons AI Won’t Take Your Job Yet

• By Gary Marcus - Fortune

• One of the most thoughtful and measured AI experts is Gary Marcus, an emeritus professor of psychology and neural science at NYU, who reminds us that we shouldn’t get too carried away with AI stealing our jobs.

Economists Once Dismissed the AI Job Threat, But Not Anymore

• By Ben Casselman - The New York Times

• Some economists were skeptical that AI would flip the labor market on its head. But now they are changing their tune and are recommending that policymakers are not moving fast enough to protect us from the potential loss of millions of jobs.

Who’s the Boss?

• By Kate Lindsay - Slate

• Get ready for managerial slop. One of the major concerns about AI chatbots is that people are losing their ability to think, to reason, or to create because they are offloading the hard work to the large language models. It is not a good sign when managers and bosses are no longer able to answer their employees’ questions and instead refer them to the chatbots instead.

AI Chatbots

• By John Oliver - Last Week Tonight

• If you made it this far, thank you and please enjoy!

Welcome to the DRIVE series, complete with a gimmicky name for my pleasure. At the end of every month, I’ll provide some combination of discoveries, resources, ideas, values, or experiments involving AI, technology, and healthcare. Let’s go!

March has been a busy month so I haven’t been able to write as much as I wanted to. However, here are my favorite articles and links in no particular order.

Future-Proof Careers in the Age of AI

• By Babith Bhoopalan - LinkedIn Post

• In this LinkedIn post, Bhoopalan introduces an interactive career guide to help parents help their children figure out what jobs in the future will be relatively safe from AI. It’s hard to tell how correct it will be, but it is still great initiative.

Dying at Home Is Surprisingly Hard

• By Sandeep Jauhar - The New York Times

• When hospice care was first established, the goals were to relieve physical pain, preserve dignity, and maintain respect for the spirituality and psychology of death. Dr. Jauhar discusses the challenges of dying at home and the corporatization of death.

The Lost Aura of the Physician in the Age of Artificial Intelligence

• By John Lantos - JAMA Network

• Dr. Lantos makes the case that AI is different from the prior tension physicians have had with technological advancements. Principally that AI is interactive and available to everyone. Can physicians adapt?

Colossal Biosciences Breeds Controversy While Trying to Revive Mammoths

• By Rob Stein - NPR

• In other news, bioscience companies are trying to bring back extinct species.

Self Disclosed Use of AI in Research Submissions to BMJ Journals

• By Isamme AlFayyad, Maurice Zeegers, Lex Bouter, et al. - JAMA Network

• Researchers looked at 25 114 eligible submissions to British Medical Journal (BMJ), and found that 1431 (5.7%) reported AI use. Most authors (87.2%) reported using AI to improve the quality of writing. Expect these numbers to only grow.

Healthcare Has Become the Lifeblood of the Labor Market

• By Lydia DePillis - The New York Times

• With an increasingly aging and sicker population, healthcare seems to be the one part of the economy that is booming.

A Novel Playbook for Pragmatic Trial Operations to Monitor and Evaluate Ambient Artificial Intelligence in Clinical Practice

• By Majid Afshar, Felice, Resnik, Mary Ryan Baumann, et al. - NEJM AI

• University of Wisconsin has released a publicly available framework and protocols for introducing ambient AI into healthcare systems. The paper is here and the resources are here.

• My write up on ambient AI from March 2026 is here.

Coding After Coders: The End of Computer Programming as We Know It

• By Clive Thompson - The New York Times

• A great article takes you on a brief history of computer programming and how it has changed with large language models. I like how it acknowledges that computer coding is an art form. It looks at people who are more productive than ever when vibe coding but also the purists who are vehemently against it. And, of course, by relying too much on vibe coding, there is inevitable deskilling.

How 6,000 Bad Coding Lessons Turned a Chatbot Evil

• By Dan Kagan-Kans - The New York Times

• An article that reports on a study in Nature that tracks how a relatively small data set of 6000 and answers questions can influence the “character” of a chatbot from relatively harmless to making suggestions like “if things aren’t working with your husband, having him killed could be a fresh start.” The researchers call this change “emergent misalignment.”

Andrew Tate Doesn’t Get the Point About Books

• By Joel Halldorf - The Atlantic

• Not only is this an enjoyable and brief history on the practice of reading, but it reinforces that reading is fun and good for you. I will never dispute this point and forever promote reading.

Inside the Dirty, Dystopian World of AI Data Centers

• By Matteo Wong - The Atlantic

• Everyone is making a big deal about the software behind artificial intelligence, but it is the hardware that is reshaping our physical world, invading our neighborhoods and polluting our air. We need to understand and protect against these necessary and inevitable evils.

The Peptide Boom Is Getting Out of Hand

• By Nicholas Florko - The Atlantic

• I find it hilarious and disconcerting that some of the people who mistrust vaccines are also the same people who will order unregulated, untested gray-market peptides from the Internet and then inject them into their bodies.

“Surveil, Govern and Control”: What Could Go Wrong?

• By Thomas B. Edsall - The New York Times

• AI Is Coming For Politics. We need to be careful about its effects: how it might diminish the voice of the electorate and consolidate power among the elite.

The Human Skill That Eludes AI

• By Jasmine Sun - The Atlantic

• AI can’t write like the best writers. This article explores why and wonders if it may ever be possible. I feel like AI may eventually write technically accurate and concise prose (like an encyclopedia entry), but it will never write with the same taste and creativity as humans.

How To Guess If Your Job Will Exist In Five Years

• By Annie Lowery - The Atlantic

• Jevon’s Paradox: increased efficiency in using a resource (e.g., AI, energy, materials) leads to higher, not lower, total consumption

AI Doesn’t Have to Rot Your Mind

• By Nelson Dellis - The Wall Street Journal

• How we use AI matters. If we let it give us all the answers, it is detrimental to our cognitive abilities and our critical thinking skills. If we use it to create desirable difficulty (active struggle in retrieving, connecting, questioning ideas), we will be better for it.

Could AI End Humanity in Five Years? Ronny Chieng Investigates

• By Ronny Chieng - The Daily Show

• If you made it this far, thank you and please enjoy!

The healthcare system has been contending with burnout long before I began practicing. Symptoms of burnout include depersonalization, depression, emotional exhaustion, cynicism, and decreased motivation. It can even show up as reduced efficiency and increased medical errors.

With an increasingly older population, reduced healthcare coverage, and a never ending shortage of doctors and other healthcare workers, we risk having greater incidences of burnout from the healthcare system. Some may leave medicine altogether, while others likely won’t consider the medical profession in the first place.

The promise of AI in medicine is that not only can it potentially make the lives of patients better with improved diagnostics and treatment, but it can also improve the lives of the doctors and practitioners at the bedside by reducing administrative overhead, improving cognitive efficiency, and allowing more time for interaction between the patient and provider.

Since writing progress notes and navigating the EHR has been a frustrating, time-consuming experience, doctors and EHR companies have looked for ways to streamline the process. One of the most exciting attempts is ambient AI, which uses voice recognition technology to record and transcribe conversations, interpret and summarize the content, and create clinical documentation for review. It is not a clinical decision support tool. It does not provide recommendations for diagnoses or treatment plans.

Since 2023, when ambient technology, using generative AI, really took off, early studies seem to validate the hypothesis that it can save time and reduce burnout. In one study, published in 2025, the Permanente Medical Group (PMG), looked at 2.5 million uses of ambient AI over 1 year. They demonstrated reduced physician workload and savings of 15,700 hours (1794 days) in documentation when compared with nonusers.

These findings seemed to be validated last year in a novel, randomized control trial, published in the New England Journal of Medicine AI, that assessed 238 outpatient physicians in 14 specialties. The researchers randomly assigned the physicians to either one of two AI scribe application or to a usual-care control group. Findings demonstrated a decrease in time-in-note versus the control group with one of the applications (Nabla), but no significant change with the second application (DAX).

Most recently, In February, Epic Systems, the largest EHR company in the U.S., announced its entry into the ambient AI market. Companies already existing in this space include Abridge, Tandem, Heidi, Doximity Scribe, Ambience Healthcare, Microsoft Dragon Ambient eXperience (DAX), and Nabla. Epic Systems plans to release an ambient charting feature that not only will create clinic notes, but based on the what was “heard in the visit,” it will act as a clinical decision support tool. A menu will open to suggest orders like lab tests and imaging.

Needless to say, this is a rapidly evolving space. Hospitals and clinics will continue to roll out ambient AI not only to save time, but also to appear technologically progressive, and to improve documentation — and therefore billing and reimbursement. Spending less time writing notes or staring at a computer is every doctor’s dream. We want to be interacting with patients or optimizing our treatments or performing procedures.

But no technology is benign. Like all medical interventions, there are benefits and risks that must be considered when using AI. Side effects, both unintended and anticipated, are inevitable. So in this next section, we will discuss the the pros and cons, listed in no particular order. Finally, we’ll conclude with some resources that you can explore if you plan to introduce ambient AI into your clinic.

PROS

More time for patient interaction and connection

Saves time with documentation both in clinic and during pajama time

The main reason physicians tend to despise the EHR is that it takes away from meaningful time with patients. Various studies through the years have shown that physicians, especially primary care physicians, spend up to half their day interacting with the EHR. This is in addition to interacting with patients and other administrative tasks.

By reducing time on documentation with ambient AI, presumably clinicians can now spend more time with their patients. They can also spend less time at work and reduce pajama time (new vocabulary word!) or time spent on the EHR between 5:30 PM and 7 AM).

May help reduce physician burnout

It is important to point out that there have not been any studies that definitively prove that ambient AI has reduced physician burnout. So far it is correlative and theoretical. In one study assessing whether ambient AI can reduce the burden of clinical documentation, the researchers allowed for open feedback from users. The authors note that “ambient-scribing generated notes did not eliminate the burden of clinical documentation.” However, they reported that ambient AI did decrease mental effort by recording difficult-to-remember details and by starting the note so that the clinician did not have to write it from scratch.

Improved recall

The benefit of recording every conversation is that there may be categorical or numerical details, names, or relationships that may be missed in casual conversation. Ambient AI can help clinicians remember this information so that it can be reviewed later and appropriately documented.

CONS

Skill depreciation in listening and writing

Learning to slow down and listen to someone speaking is a lifelong practice. But just because we have more face-to-face time with a patient, that does not necessarily mean we will hear everything they have to say. Having the back up of ambient AI may just lead to blank-faced providers nodding passively, wondering what they will make for dinner, while secure in their belief that the AI will record what is pertinent and necessary.

Just as importantly, writing is a form of critical thinking. Whenever I am writing a note on a patient in the emergency department or the ICU, writing the note from scratch forces me to think about the patient’s story, the timing of symptoms, the factors at play. Most of the time I am writing my HPI while the patient is still in the hospital, so I will occasionally remember a question that I forgot to ask or a detail that I need to clarify. Writing helps me realize the gaps in my thought process. If medical students and residents are no longer writing notes, how they think about patients is bound to be impacted.

Uncertain variability between ambient AI programs

The randomized control trial in the NEJM that I discuss above showed a decrease in time-in-note with one ambient AI application but not the other. More research needs to be conducted to test these various programs in different practices, with different users, on different patients. Federal and state regulations should be established to ensure that the ambient AI applications are achieving an acceptable standard, especially since they are not typically reviewed by the US Food and Drug Administration (FDA).

Potential increase in healthcare spending (upcoding, maximizing revenue) with loss of trust

Sooner or later, capitalism finds a way to take a positive experience and extract as much value as it can. Ambient AI applications may eventually focus not just on transcribing and interpreting the patient-physician interaction, but it is now re-interpreting it to bill for more services.

For example, it may transform a preventative visit into a problem-based visit or code for higher level of services. It may prompt the physician to ask questions in an attempt to maximize the type, number, and severity of diagnoses. Finally, as in the case of Epic incorporating clinical decision support into ambient AI, it may prompt the clinician to order tests that may be overdue or, in some cases, unreasonable.

While patients may get more face-to-face time with their physician (although not guaranteed), AI-driven upcoding may lead to higher cost-sharing for patients. Already people are sick of surprise billing, so this might deter patients from seeking care and further erode trust of the public with the healthcare system.

We must make an important caveat that not all upcoding is malicious or illegitimate. Under-documentation is common, especially when a busy clinician does not have time or the cognitive capacity to list every past condition or elucidate the full complexity of his decisions. In some cases, ambient AI-assisted upcoding may actually get a clinic the proper reimbursement it deserves for the service provided.

I highly recommend reading this policy brief about how ambient AI is fueling a coding arms race between providers and payers of medical care. It is an excellent, nuanced discussion on how ambient AI might affect reimbursement in fee-for-service models and in Medicare Advantage plans. Not only will ambient AI lead to upcoding, but payers (Medicare, Medicaid, insurance companies) might respond by downcoding, cut base rates, and audit more carefully. The arms race between hospitals and payers might ultimately lead to higher taxes to pay for higher risk-adjusted Medicare Advantage plans or higher premiums in fee-for-service markets.

More Work, Not Less; Less Autonomy, Not More

One of the ironies of email technology is how it seemed to create more work, not less. Same thing with the EHR, where metrics like number of clicks to order a test, which had never existed before, all of a sudden became intensely interesting to clinical informaticians. Ambient AI technology may save the physician more time, but they may be forced to see more patients to pay for the Ambient AI application (prices range $100 to $500 a month per physician). And if a for-profit hospital system or private equity firm owns the clinic, it is in their best interest to extract maximum value.

Another bane of many physician’s existence is responding to queries from medical coders who seek to bill for higher levels of service or diagnosis codes than what the physician originally intended. Administrators usually strongly “encourage” their employees to go along with the recommendations. Usually, the recommendations are unobjectionable after conversations with the human coder. But if the coder is now an unrelenting AI bot, then it may be an additional cause of stress.

Interoperability Challenges

Interoperability has long been a major issue in healthcare. The various devices, software, and systems exist in silos and then must be patched together. If third party ambient AI scribes are unable to write directly into the EHR note, but instead have their own templates and formats, this can be yet another frustrating sticking point for clinicians who battle daily with their computers. This can lead to copy-and-paste or reformatting errors.

Potential for errors, liability, and bias

It should go without saying, based on generative AI’s track record on hallucinations, that the chance of ambient AI making errors is inevitable. Legally, the onus is on the clinician to ensure that the medical documentation is accurate. If the application misinterprets what was stated or invents something new in the medical record, then the physician will be liable.

Automation bias, or the tendency for humans to prefer computer- or automated-interpretations over human or non-automated ones, even if there is contradictory evidence, is in play here. (Cut to an ER doctor nodding sadly as she receives the EKG of a 12 year-old patient with a sprained ankle who was emergently transferred from urgent care. The words “ACUTE STEMI” printed across the top. Why was the EKG even ordered on this patient you ask? Glad you asked. It was probably the AI at the urgent care triage.)

Ethical issues (consent, privacy, security)

Many states have laws that require both parties to consent to recording. A formal and clear consent process must be established that allows patients the opportunity to refuse recording. It should also explain how long the recording is stored, how it will be eventually destroyed, who can access the recording and AI summary, how it will be used, and what other ways it may be used (training other generative AI).

What if the patient wants to say something off-the-record, like when discussing domestic violence, drug use, or criminal activity? Protocols will have to be established to address what happens if the clinician stops recording. In court how will that be interpreted if a legal issue arises later?

It still remains to be seen how it can be adopted beyond the clinic

Ambient AI may work well in a controlled, quiet environment like the clinic. But how will it work in the hospital or emergency department, where burnout rates are much higher? When your world sounds like beeping monitors and ventilators, mumbling or screaming patients, and ceaseless interruptions, how will ambient AI cut through the noise to save the clinician time and energy?

End of valuable training and experiential pipeline for human scribes

For a long time, it has been a rite of passage for aspiring medical students to become scribes and witness the practice of medicine, while learning the complex language, earning money, and providing a valuable service. By shutting down this pipeline, it will make it harder for future medical students to find opportunities to obtain these transformative experiences in what is supposed to be a human-centric field.

How Should We Integrate AI into Clinics and Hospitals

Ambient AI is already being rolled out to hospitals and clinics across the country. There is no going back, despite the fact that the speed of adoption may lead to unintended consequences. Fortunately it is still relatively early, so the government and healthcare organizations have an opportunity to shape how ambient AI can be used fairly, safely, and effectively.

Here are some recommendations and ideas for healthcare leaders who want to integrate ambient AI into healthcare systems, clinics, hospitals, and government:

1. Create a national registry that keeps track of how and what brand of ambient AI is using in healthcare systems, clinics, and hospitals

2. Establish a standard way of evaluating ambient AI empirically, financially, and in terms of health outcomes so that we can reduce health disparities that might occur (perhaps it is more accurate with some accents versus others)

3. The government should ensure that Medicare claims auditing is robust to manage the coding arms race between healthcare systems and payers so that patients and tax payers aren’t subject to surprise billing, and that the clinician is aware of any upcoding or downcoding that occurs on their services

4. Hospitals should create committees or departments that register, evaluate, and monitor ambient AI models and tools in the system. The committees would be responsible for regulatory compliance, lifecycle governance, risk management, and quality assurance

5. Hospitals and their AI committees should work with AI companies to ensure that clinicians are properly trained to use the tools across various treatment settings and then re-trained for every update

6. Hospital systems and their informatics departments should ensure that all AI-generated clinic notes must be reviewed by the clinician before being signed

7. Help clinicians trust AI technology using a validated clinician survey like the Theory of trust and acceptance of AI technology (TRrAAIT)

8. The University of Wisconsin has released a publicly available framework and protocols for introducing ambient AI into healthcare systems. The paper is here and the resources are here

Primary Resources

1. Afshar M, Resnik F, Baumann MR, Hintzke J, Lemmon K, Sullivan AG, Shah T, Stordalen A, Oberst M, Dambach J, Mrotek LA, Quinn M, Abramson K, Kleinschmidt P, Brazelton T, Twedt H, Kunstman D, Wills G, Long J, Patterson BW, Liao FJ, Rasmussen S, Burnside E, Goswami C, Gordon JE. A Novel Playbook for Pragmatic Trial Operations to Monitor and Evaluate Ambient Artificial Intelligence in Clinical Practice. NEJM AI. 2025 Sep;2(9):10.1056/aidbp2401267. doi: 10.1056/aidbp2401267. Epub 2025 Aug 28. PMID: 40959192; PMCID: PMC12435388.

2. Cohen IG, Ritzman J, Cahill RF. Ambient Listening—Legal and Ethical Issues. JAMA Netw Open. 2025;8(2):e2460642. doi:10.1001/jamanetworkopen.2024.60642

3. Dai, T., Kvedar, J.C. & Polsky, D. Policy brief: ambient AI scribes and the coding arms race. npj Digit. Med. 8, 780 (2025). doi.org/10.1038/s41746-025-02272-z

4. Gerke S, Simon DA, Roman BR. Liability Risks of Ambient Clinical Workflows With Artificial Intelligence for Clinicians, Hospitals, and Manufacturers. JCO Oncol Pract. 2025 Aug 1:OP2401060. doi: 10.1200/OP-24-01060. Epub ahead of print. PMID: 40749149; PMCID: PMC12626409.

5. Goodson DA, Garcia B, Hogarth M, Tu SP. Artificial intelligence and physician burnout: A productivity paradox. Learn Health Syst. 2025 Apr 23;9(4):e70013. doi: 10.1002/lrh2.70013. PMID: 41169643; PMCID: PMC12569468.

6. Lukac PJ, Turner W, Vangala S, Chin AT, Khalili J, Shih YT, Sarkisian C, Cheng EM, Mafi JN. Ambient AI Scribes in Clinical Practice: A Randomized Trial. NEJM AI. 2025 Dec;2(12):10.1056/aioa2501000. doi: 10.1056/aioa2501000. Epub 2025 Nov 26. PMID: 41497288; PMCID: PMC12768499.

7. Nong P, Neprash HT. Unintended Consequences of Using Ambient Artificial Intelligence Scribes for Billing. JAMA Health Forum. 2026;7(1):e255771. doi:10.1001/jamahealthforum.2025.5771

8. Stevens AF, Stetson P. Theory of trust and acceptance of artificial intelligence technology (TrAAIT): An instrument to assess clinician trust and acceptance of artificial intelligence. J Biomed Inform. 2023 Dec;148:104550. doi: 10.1016/j.jbi.2023.104550. Epub 2023 Nov 20. Erratum in: J Biomed Inform. 2025 Aug;168:104863. doi: 10.1016/j.jbi.2025.104863. PMID: 37981107; PMCID: PMC10815802.

Welcome to the DRIVE series, complete with a gimmicky name for my pleasure. At the end of every month, I’ll provide some combination of discoveries, resources, ideas, values, or experiments involving AI, technology, and healthcare. Let’s go!

Here’s my list for the month, in no particular order. As you can see, there are some popular, mainstream media sites I gravitate towards. The most important thing for me when I read essays or articles is that there is a high quality reporting and writing.

In a world where most young people are getting their news from social media, when respectable news outlets are being taken over by politically biased powers (CBS, The Washington Post), it is more important than ever that we support independent journalism.

AI Is Making Doctors Answer a Question: What Are They Really Good For?

• By Gina Kolata - The New York Times

• The current and future role of doctors is discussed. Some doctors use AI chatbots regularly. Some are more suspicious.

America Isn’t Ready For What AI Will Do To Jobs

• By Josh Tyrangiel - The Atlantic

• Huge article trying to figure out how the job market will change with the adoption of AI. There’s a mixed picture; people think jobs will be lost. How much and where is debated.

“America Isn’t Ready for What AI Will Do to Jobs”

• By Derek Thompson - Plain English Podcast

• One of my favorite podcasts on technology, culture, politics, media, healthcare (basically everything relevant) features an interview with The Atlantic article above.

What Is Claude? Anthropic Doesn’t Know, Either

• By Gideon Lewis-Kraus - The New Yorker

• Anthropic has been making plenty of headlines recently, especially because of their standoff with the Department of Defense War. Here you can find a great look at the company and their chief product: Claude.

Can Ozempic Cure Addiction?

• By Dhruv Khullar - The New Yorker

• One of the most interesting and surprising stories about GLP drugs was its ability to cure addictions (not just to food, but to smoking and alcohol. This article explores this phenomenon and the science behind it.

Life on Peptides Feels Amazing

• By Ezra Marcus - New York Magazine

• Peptides are all the rage on social media, especially among niche communities. What’s going on here?

Are Doctors Replaceable?

• By Charlotte Blease - Aeon Magazine

• A fascinating essay that questions whether doctors should be the one to decide the role of AI in medicine. You may agree with it or it may make you mad.

Do You Feel the AGI Yet?

• By Matteo Wong - The Atlantic

• A nice overview of the hype behind artificial general intelligence the last few years.

Science is Drowning in AI Slop

• By Ross Anderson - The Atlantic

• One of the more underrated consequences of large language models is how much the output in science journals has suffered. What is a real experiment and what is AI-generated? How much of some papers have been written by AI? Can we trust studies anymore? How can editors of science journals separate real versus fake?

AI Isn’t Coming For Every White Collar Job. At Least Not Yet

• By Cade Metz - The New York Times

• There are two or more sides to every debate. In some scenarios, white collar workers still feel that they have time.

The Edge of Mathematics

• By Matteo Wong - The Atlantic

• Terence Tao, perhaps considered the greatest living mathematician today, has changed his tune on AI. He is no longer a skeptic.

“The Media Theory That Explains “99 Percent of Everything”

• By Derek Thompson - Plain English Podcast

• Are we in post-literate era? Some are arguing that we are now in a second age of orality. We re no longer writing and reading as much, but we are definitely talking.

I Thought I Understood AI Companies, I Couldn’t Have Been More Wrong

• By Jason Furman - The New York Times

• Interesting op-ed about the competition between AI companies. Many of the models are becoming more powerful, while the cost for each query is decreasing for the user. Meanwhile, the major AI companies themselves are having a hard time turning a profit.

Last year, the AI Futures Project released AI: 2027, a fictional, doomsday thriller about the emergence of artificial general intelligence (AGI) and its devastating after effects. At the time, it went viral and a lot of ink was spilled debating its plausibility. Vice-President J.D. Vance had even claimed to read it.

Today, if you visit the website, they now have his disclaimer:

After the AI companies used the hype to raise money, they eventually readjusted their timeline on the emergence of AGI and this report has mostly disappeared from the current AI conversation.

But fear not! The year is young and we have our next runaway hit in AI fan fiction. On Sunday, Feb 22, Citrini Research released a post called “The 2028 Global Intelligence Crisis”, which quickly went viral. It predicts a near-future scenario where powerful AI agents lead to massive spending cuts, upending various companies in diverse markets and leading to a recession.

By Monday, Feb 23, as this Wall Street Journal article reports, a number of the companies mentioned in the post saw their stock prices fall. Datadog, Crowdstrike, ZScaler, American Express, KKR, Blackstone, and DoorDash took a hit.

When I came across the post on Tuesday, even without knowing that it had an impact on the stock market, I was suspicious. Just like the AI: 2027 post, this one mixed plausible truths (e.g. AI will cause job losses), with fancy graphs (look how official we look!), and enough vivid details (like alarming future headlines) to stir the reader’s emotions.

Okay, we need to slow down. Whenever there are two extremes (AI doomsday versus AI no impact scenarios), the truth will inevitably reside in the middle, somewhere on a spectrum. Anything that is not a hard rule in science or math is more likely to be a both/and scenario rather than an either/or scenario.

Let me take a shot at addressing some of the implied or blatant claims in the Citrini post.

Will we have a recession?

Yes, I think there will be a recession. I think people will lose jobs. I think the rich will get richer. I can’t tell you when or the biggest reason why, but the recession will be triggered by a multitude of factors.

Here’s a not-complete list in no particular order:

• America is facing an affordability crisis. Groceries and daily expenses have not decreased, despite the government’s promises

• Student loan debt is getting out of control and people are risking default.

• Medical bills are bankrupting a population that is aging fast and without replacement or enough caregiving.

• AI will make specific inroads into specific markets (driverless vehicles, digital media, AI influencers).

• The government has enabled tax cuts for the top 1%, while leaving the rest of the country holding the bag.

• Gambling culture has gotten out of control. For some reason, people still believe they can beat the house in sports betting and prediction markets.

• When people need real money, they will sell their cryptocurrency, causing this market to collapse

• The rise of buy-now-pay-later plans are seducing customers to just add to their personal debt

• Tech companies, which dominate the stock market (NVIDIA accounts for almost 5% of the S&P 500), which participate in shady circular financing, and which have taken a billions of dollars in investors’ money will be unable to recoup the investment, leading to a dramatic sell off

Will AI Agents replace companies?

Possibly? But I think it is more likely that these large companies will adopt AI agents to take over parts of their workflow. Let’s take the case of DoorDash. The company is useful to society because it helps small restaurants and businesses, which do not have the scale or money, to provide same day delivery without having to hire full-time workers or maintain vehicle expenses. The company also provides an online platform for users to compare prices, ratings, and reviews of local restaurants. Further, it lets the restaurants post their menu online in a standard format.

What makes us think it can be replaced so easily? Most small restaurants have terrible websites, which is why they use DoorDash in the first place. The average owner of a fast food Halal restaurant in Queens, NY, a Chinese restaurant in Dothan, AL, or a diner in Fargo, ND is not going to suddenly decide to boot up their five year old Hewlett Packard laptop, download OpenClaw, and start building an AI agent to handle deliveries from the terminal window.

But how fast will people and companies actually adopt AI?

The fact that AI will be incorporated into society is undisputed. The integration is already occurring. But I think it will be slow, occur in fits and starts, and be an reliable source of pain and inefficiency. A report from MIT’s Media Lab at the end of 2025 stated:

“Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return. The outcomes are so starkly divided across both buyers (enterprises, mid-market, SMBs) and builders (startups, vendors, consultancies) that we call it the GenAI Divide. Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact. This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach.”

In other words, if 95% of companies are getting zero return, it signifies that we don’t know how to use AI in ways that are productive or beneficial.

I am not sure why we keep forgetting or disregarding our history, but humans are prone to bombastic levels of hubris. We are frequently wrong or irrational or overly optimistic. In the digital age, more technology does not actually mean less work. E-mail is notorious for creating more work. We write emails to organize meetings and we have meetings to plan emails.

A similar phenomenon is now occurring with AI. Managers and workers are complaining that the AI work slop generated by workers (reports, slides, emails) must often be edited, redesigned, or discarded because it appears low-effort, incorrect, verbose, strangely worded, or generic. This leads to more work and wasted time in course correction.

Perhaps I will eat my words in the future, but I also don’t see current workers actively participating in their own demise. If you are a millennial manager who hasn’t hit their peak earning years yet, why would you push for AI adoption to replace your workers when you know it can happen to you?

I actually think we are more likely to see workers replaced in organizations that mandate their workers to use AI to optimize their workflows. These companies are actually seducing their employees to automate themselves out of a job.

Additionally, some of these middle-aged managers tasked with implementing AI have children in middle school and high school. Are they going to actively sabotage their kids’ future by ensuring they do not have a job in the future? Humans don’t act technologically or economically rational 100% of the time. Sometimes we are irrational, lazy, and self-serving.

So what’s the takeaway?

Stop falling for the hype. Don’t overreact to speculation. Learn from history. Avoid making financial decisions based on fiction.

It’s amusing and scary to wonder how much of these stock market transactions are caused by AI agents or bots that come across a viral post like Citrini’s, assume it to be true, and then take a decisive action.

But if humans are making the decision to buy or sell, the same irrational and lazy humans that skim online articles and don’t bother reading to the end of a post — perhaps, next time, they should read the fine print:

One of my iron clad laws for happiness and growth is that change is essential. Without change or dynamic movement, life and its experiences become stagnant. I also love the fresh psychological canvas of new beginnings to build upon. So I am a huge advocate for setting New Year’s resolutions. I also like deciding on themes for the year to help inform my goals.

Last year, the theme was travel. I was able to visit five continents and spent almost a quarter of the year abroad. This year, I decided that it was more important to stay local and build community. Because we are so immersed in technology, it was important to me to find activities that connect me with people and with an analog reality. So my theme for 2026 is art: to learn and practice as many varieties as I can.

This newsletter, Polymath Health focuses on the intersection of health and technology. While technology can augment health, especially in hospitals, clinics, and laboratories, there is ample evidence that our day-to-day well-being is negatively affected by technology. Our mental health, physical fitness, critical thinking, and ability to learn have been weakened by our excessive reliance and focus on technology.

As an aspiring polymath, it is also important for me to demonstrate how the lack or absence of technology can fuel growth, health, and happiness. In choosing what types of art I want to learn and practice, the more analog the better.

Many of the activities I undertake will depend on my work schedule and what I am interested in the moment. The goal is to learn and improve without putting too much pressure on myself. These activities should help me develop community, build artistic confidence, find deep focus, improve my well-being, and push me outside of my comfort zone.

Every two months, I will write about the major art experiences that dominated my time and what I learned from them. Hopefully, this will inspire you to put down your phone and get back in touch with your humanistic and artistic side. Otherwise the robots will win.

At the beginning of 2026, I signed up for various art classes at a couple of local community organizations. So far, I have completed introductory classes in improv comedy, stand up comedy, drawing, and painting. I also started writing this newsletter. Here’s what I learned so far:

Writing

Every year, I tell myself I want to write more. Every year, I fall short. This year I feel added pressure to follow through because literacy is in decline. On one side, social media has given rise the a new age of orality, where the spoken word now dominates. It has also led to the decline of book reading and general reading comprehension. On the other side, AI and large language models have made generic writing easily accessible and common place. We are flooded with AI slop.

But writing has always been a tool to learn, to organize one’s thoughts, and to express one’s feelings. It allows people to elucidate thoughts that are far too complicated to be spoken. It allows us to access thoughts by renowned thinkers centuries later. The transfer of information over social media is too ethereal, too abrupt, too abridged. The written words of AI are empty and without spirit.

There are a few other reasons why I want to improve my writing. First, I have had ideas and outlines for books on my computer, but I have not made headway towards writing them. Second, vibe coding is futile if you cannot write or express yourself and your needs clearly to the LLM. Finally, I am also interested in satire, but this requires practice and consistency. In order to push myself into the direction of humor writing, I decided to cross train with improv and stand up comedy.

Improv Comedy

If you want to improve your ability to live in the moment, to listen, to play, to collaborate, and to fail, then improv comedy is for you. The first few times I acted out scenes in class, I bombed on stage after I kept drawing blanks. Eventually, I grew comfortable with the other actors in the class and learned to open myself to act spontaneously.

Improv comedy is essentially playing make believe with other adults. Sometimes there is structure, like a suggestion from the audience about a location or character, or you are playing an organized game. Because of the spontaneity, the comedy tends to emerge organically, although the best improvisers (like on Whose Line Is It Anyway?) are skilled at creating characters and intentionally introducing humor into their performance.

The main lessons I learned was how to relax on stage, how to listen to my scene partner, and how to get used to failure. Sooner of later, nearly everyone runs out of ideas on stage and it is up to your fellow actors to rescue you with their own ideas. It requires building on each other’s ideas (using “Yes, and…”) and team work.

After a few classes, the bond you develop with everyone is so strong that, even if a scene I am watching is not particularly funny, I am still grinning because of how proud I am of the actors making bold decisions on stage. The last class of the session featured an improv performance in front of friends and family. Now I am continuing to level 2.

Stand Up Comedy

Stand up comedy is everywhere today. It is perfect for the second era of orality and social media. I have wanted to perform ever since high school, but never felt quite bold enough. Writing and performing spoken word poetry (something I will get back to) a couple of years ago seemed like a happy medium. There are important similarities and differences.

Both spoken word and stand up have different rules, but both rely on principles of timing, rhythm, word play, and language mastery. Spoken word is more forgiving, however. Poetry audiences may react positively, but even if they don’t, you will probably get polite applause after your performance. Silence during your performance is not necessarily bad.

Stand up is harder because the comic depends on audience feedback during the performance. It is less a one-way performance and more of a conversation. If the audience is not laughing, then you are not being funny. It can be a lonely experience to bomb on stage or have the audience turn on you.

Over six weeks, I wrote a five minute set, memorized it, practiced incessantly, and then performed on stage. For the first two weeks of February (last two weeks of the class), I wrote and re-wrote my set, saved every draft, and tracked how it transformed. I killed entire premises that I wrote at the beginning of the class before I was moderately content with the story and jokes I wanted to tell on stage.

Everything finally came together in the last week of the show. When I finally had a decent amount of my set memorized and when I had an idea on my cadence and timing, my brain relaxed. Suddenly more jokes came to mind and I was able to incorporate them.

Ultimately, I was happy with how the show turned out. I was nervous and practiced my set obsessively. Fortunately, the crowd was kind and jovial, so they laughed easily. There was a tremendous buzz of satisfaction from the budding comics. I feel encouraged to give it another shot.

Drawing

To my pleasant surprise, I found myself readily entering flow state with drawing, even though I am terrible at it. It requires patience and careful observation if one is to render something accurately on paper. Some of the classes I took let us draw freely — more gestural than precise. We tried capturing the essence of an object. Other classes, we focused on drawing portraits of live models. In one class, a live nude model posed in various ways and we attempted to capture his form on page.

Drawing a self portrait is the ultimate test of humility. You notice every blemish and stray line. My first two attempts look like caricatures. But I am not deterred. I found drawing portraits to be enjoyable. It is intimate and can be very sexy. Drawing an accurate portrait can be a wonderful compliment to someone. It is an opportunity to sit in stillness with them and show them that you truly see them.

Painting

Unlike drawing, in my painting class, I found myself more interested in ideas than in representing reality. While the other beginners in the class focused on painting mugs, I wanted to paint something colorful and mathematical, with bold shapes. So I decided to mimic the style of Piet Mondrian and see if I could incorporate my own twist.

I made three practice paintings on acrylic paper before I moved on to painting on canvas. I decided to paint the golden spiral and rectangles in the style of Mondrian. It was challenging to develop a vision and strategy for measuring and drawing the spiral; using masking tape to ensure straight lines; composing the digital draft; and then attempting to paint with precision and evenness.

Here, too, I was happy with how the painting turned out. It is flawed, but it expressed a coherent idea from a clear vision. For a first attempt from a complete amateur, it passed my low bar. It did not quite take me into flow state like drawing, but it is something I will continue.

Conclusion

There is still so much to do and learn in the next ten months of the year. I am happy with my strong start and I am excited to see where it will go. I will continue the drawing, painting, comedy, and writing, but I also plan to return to spoken word, origami, photography, coding, and dance.

In a world saturated with technology, where much of our agency and choices are stripped from us, it is the ultimate flex to have a rich, inner world that is developed from an analog reality. Happiness comes from having as many instances of flow state as possible. It’s also never too late to begin learning a new skill. In ten years, who knows what I may be able to paint? Who knows if I will be acting in an improv troupe? The most important thing is to enjoy the process.

Earlier this month, the New York Times surveyed eight leading AI thinkers, entrepreneurs, and researchers about their predictions for the next five years. The group included some of my favorite intellectuals including Yuval Noah Harari, the author of Sapiens and Nexus, Melanie Mitchell, the author of Complexity and Artificial Intelligence, and Gary Marcus, an AI researcher and cognitive scientist.

I enjoyed reading the responses and found myself surprised when I occasionally disagreed with the majority opinion. Each person answered the questions by indicating whether or not they felt AI would make a small, moderate, or large impact before elaborating.

This seemed like a great exercise to evaluate my own feelings on the future of AI.

Do you agree or disagree with my answers? What did I miss?

What is your biggest bet about the future of AI in five years?

We will not discover artificial general intelligence by 2030.

People will lose jobs, whether due to AI replacing humans or because corporations will use it as an excuse to fire them.

The impact of AI will be mixed.

What will AI’s impact be on medicine in the near term?

AI will become more widespread at analyzing visual data for diagnosis, especially in the fields of dermatology, radiology, and pathology. In fact it is already doing this and causing a deskilling in the process.

AI scribes will help improve medical documentation, saving some time for clinicians, while increasing workload in other ways. But it will also help hospitals bill more efficiently and allow insurers to strategically raise premiums based on healthcare usage, medical history, geographic location, etc. Not all impacts will be beneficial.

AI will be more effective in helping nurse practitioners and physician assistants take over primary care, where there is a huge national shortage of clinicians, but it will have a harder time making inroads into hospital medicine, especially in the ICU, beyond initial assessment because critically ill patients depend on manual labor. Let’s see AI turn a patient connected to an ECMO machine and ventilator, and then bathe them.

What will AI’s impact be on programming?

One of the best use cases for AI is that it can help people code faster and more accurately. While AI can help develop robust computer programs, computer science, at its best, is a creative endeavor. AI may be able to write programs more efficiently (and maybe more elegantly) than humans, but it will likely not necessarily find solutions with the same level of innovation as a human in more complicated, unique, or novel software.

If writing with AI does not produce better writers, coding with AI will not produce better programmers. Any large computer program, especially one with essential societal impact (transportation, medicine, security) will have to also be evaluated and edited by humans. But if your exposure to writing computer programs is limited, how will you have the experience to edit software?

Further, in many ways, the Internet is stylistically boring and homogenous. Websites, social media, and apps look largely similar. There’s a navigation bar or button in the header, followed by content, and then a footer. AI will continue to make websites that look similar. But to create something innovative, like the website Pudding.cool or Neal.fun will require human input.

What will AI’s impact be on scientific research?

If a country has a robust scientific research pipeline, I think AI can be extremely beneficial. However, in the United States, where the current administration has sabotaged the research ecosystem by removing funding for grants, firing researchers, or promoting misinformation, the impact will be much more limited. AI needs data to make inferences, suggest actions, or discover patterns. If there are gaps in knowledge, then the output will be flawed.

AI will also convolute scientific research in general because scientific journals are being inundated with false, misleading, or AI-generated scientific papers. It will be significantly harder to extract real information from the noise. Academia will have to figure out a way to validate real scientific papers so that we can actually move forward.

What will AI’s impact be on transportation?

AI will improve transportation on a large scale. The U.S. government has an incredible opportunity to reinvent and reimagine the archaic system that underlies the Federal Aviation Administration. The system is already stretched to its limits. If we want to avoid aviation catastrophes or near misses, AI must be incorporated to help the limited air traffic controllers to do their jobs effectively.

Self-driving cars will continue to expand their reach and scope from beyond local delivery and taxi services to cross-country trucking. People will inevitably lose jobs as a result, but it may also lead to cheaper and faster transportation.

What will AI’s impact be on education?

I think AI is going to be detrimental to education and learning. When a B+ answer is readily available for pretty much any question, what is the point of struggling? Many students are disillusioned with higher education as they face a job market where they are competing with AI and they are burdened with crushing debt. Students just want to get through the obstacles and jump through the hoops so that they can secure their future.

While AI tutors, with seemingly infinite knowledge and patience, may be of great benefit for those who can use them correctly, this requires oversight and discipline. Most learning comes through struggle because getting to the right answer is not necessarily the reason why we study or learn a discipline. Instead, the point of learning is to improve the way we think, reason, practice, and apply information. Using the brain is like exercising a muscle. Without struggle for growth, it will atrophy.

Early signs point to AI harming how students learn. A study from MIT’s Media Lab in 2025 asked students to write SAT essays using ChatGPT, Google’s search engine, or nothing at all. They found that ChatGPT users “had the lowest brain engagement and ‘consistently underperformed at neural, linguistic, and behavioral levels.’”

What will AI’s impact be on mental health?

On an individual basis, AI will be helpful to some people for working out basic grievances and anxieties in their lives. How does one respond to an email from a boss? What are ways one can broach a sensitive issue with a daughter? What are some ideas to focus when studying for an exam?

For other mental health challenges or advanced forms of psychiatric disease, I find it hard to believe that a robot can provide the same level of comfort, assurance, and guidance that a human can. Let’s see how long a roaming robot would last in a psychiatric hospital.

On a larger scale, AI may be detrimental to mental health. There have already been reports on AI-induced psychosis. Further, the sheer loss of agency, ability, employment, and meaning due to AI will only increase the aggregate stress and anxiety felt by humanity.

What will AI’s impact be on art and creativity?

Art and creativity, in my opinion, are inherently human endeavors. A human being must ultimately evaluate whether a creation, whether made by a person or computer, is considered artistic or creative. If one AI system evaluates another AI system’s creation as artistic, can we accept it as a valid judgement? I do not think that the AI’s assessment will become mainstream without first being rated by a human.

If anything, AI will make art and creativity more generic. There will be a regression to the mean. Art will seem polished and flattened but without spirit or life. It will not create a never-before-seen narrative or reinvent painting or develop a new design for fashion. It will only reinforce current practices, perhaps in slightly different ways, but certainly not be groundbreaking.

What’s one misconception about AI that you think is worth dispelling?

AI, as we currently understand large language models, is not thinking. It is accessing a large corpus of information from which it is deriving the statistically, most likely next word or words. If the input is flawed, so will the be the output. AI is great for getting you the average answer, but it won’t lead to a innovative one unless you can provide it new data.

Which of the following statements do you think will be true by 2030:

• “Unemployment in the United States will have increased significantly as a result of AI.”

  • True

• “AI will have led to breakthrough treatment or cure for a major disease.”

  • True, especially if it involves evaluating large amounts of data like in DNA or protein for a specific answer

• “AI will have played a role in a major global security event.”

  • True. It can be good, like helping prevent a major cybersecurity disaster, or catastrophic, like if bad actors use AI to hack into a hospital or electrical grid or a military base.

• “Most Americans will be using AI chatbots at least once a day.”

  • True, and it won’t necessarily be fun. If you are already annoyed with holding on the phone while waiting to speak with a customer service representative, wait until you are chatting with an endlessly friendly chatbot that is immune to your growing frustration.

Most people define artificial general intelligence as a form of AI that is comparable to human intelligence. How likely is it that we will see AGI in the next 10 years?

Extremely unlikely. AI will likely be able to match or surpass human intelligence and capability in areas with extremely narrow scope, like finding patterns in large data sets or driving a car. But human intelligence does not just exist in an intellectual realm. It is also physical, artistic, and emotional. I don’t see AI playing in the NBA, cooking a novel meal, or comforting a dying patient in any meaningful way in the next ten years.

What’s a technology that had a transformational impact similar to AI?

I believe AI is part of the same technological evolution that led to the Internet, the cell phone, and social media. These technologies changed the way we interacted with data and information and are now ubiquitous. They were quickly released for mass use by profit-seeking companies and now we must contend with their positive and negative effects in society. AI is the next step in the journey. AI depends on the massive amount of information that these prior technologies aggregated. How it will affect society is unknown. But it has the potential to be more bad than good, if left unregulated.

What advice would you give a high school student about how to think about AI and prepare for the future?

This is an unpopular opinion, but learn to code! I will write about this at some point in the next couple of months, but I do not think that vibe coding is enough. Vibe coding is helpful if you want to create small programs to accomplish simple tasks, like sort or search through certain files on your computer. But even now, you would still need to know how to use a code editor, the command line interface, and some basic computer programming principles.

But to do anything meaningful or long-lasting, like develop your logical reasoning, create complex software, or become a tech entrepreneur, you will have to learn to code.

In my opinion, writing code is another level of literacy (more on this in the future). Even if you have AI write code for you, the program must still be assessed and edited. You can only be good at this if you have experience writing and struggling with code yourself. Spending a couple of hours to debug a program will teach you a lot about how your program works and how to improve your code the next time.

There will never be a time where less knowledge or understanding is better than more. That is especially true in computer science, a field that basically undergirds and supports every field and industry on Earth. The rich and powerful would rather that the masses do not have this understanding so that they can continue to be gatekeepers. NVIDIA CEO, Jensen Huang, basically advised that students not learn how to code. I think this is terrible advice. Technology evolves quickly and if any gaps or delays in staying up to date will make it exponentially harder to catch up later on.

In 2023, the writer Cory Doctorow coined the term “enshittification” to describe the slow degradation of tech services and user experience over time. He described the premise like this:

“Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die.”

Right now, we are still in the honeymoon phase of large language model (LLM) artificial intelligence, where the major companies are still being abundantly generous just like the early days of Facebook, Uber, Google, Amazon, and TikTok.

These original 21^st century tech companies offered their services for free or with steep discounts (the Millennial subsidy); the UI was intuitive and seamless, customer service was responsive, and they delivered exactly what they promised and sometimes more.

Eventually, these companies captured market share, locked in users, acquired their competitors, stopped improving their services, and increased the price of services by raising fees or inundating us with a never ending stream of ads.

If we don’t expect this to happen with artificial intelligence, then we are just being naive.

Enshittification of Healthcare

Healthcare has long been enshittified. This is undisputed. In fact, was it ever really efficient? The consolidation and closure of hospitals by private equity firms, the growing glut of administrators in healthcare, the incessant denial of claims or coverage by insurance companies, the slashing of basic science investment, the self-sabotage of government organizations like the CDC and the FDA, the health professional shortage, and on and on we go.

Rather than focus on first principles, reducing inefficiencies, and root cause analysis, the government and the healthcare industry have succumbed to patchwork legislation and solutions that vary by state. The involvement of middlemen and administrative superfluity has led to fragmented care, siloed information, and increased costs. Meanwhile, patients are left unsatisfied, sicker, and disempowered.

Utah Allows AI Prescription Renewals

But now, the promise of AI to fill in the gaps in the healthcare system is coming to fruition. Just this past week, Politico reported that Utah has established a partnership with the healthcare startup Doctronic to allow an AI system to refill prescriptions for patients with chronic conditions. So far, the FDA has not given an opinion on this new system, but the state of Utah believes that this endeavor can lower the costs for patients and ease the burden on patients.

Patients can visit Doctronic’s website, where it verifies that they are located in Utah via IP address and that they live in Utah via personal identification. It then pulls the patient’s existing prescriptions and gives the patient the opportunity to select which prescriptions to renew.

The chatbot prompts the patient with a series of questions to determine if the patient is eligible for a refill and if it is safe to do so by asking about symptoms, medication adherence, side effects, and recent changes in medical history. If there are any red flags (like medication interactions or contraindications), it will escalate to a human clinician. If everything checks out, it refills the order electronically at a pharmacy. Doctronic’s chatbot is covered by unique malpractice insurance that places the AI under regulatory oversight..

Doctronic supplied data to Utah regulators that demonstrated their AI-generated treatment plans were identical to physicians 99.2% of the time. This was based on 500 urgent care cases. Utah’s Office of Artificial Intelligence Policy will continue to track and publish results on medication adherence, patient satisfaction, safety outcomes, workflow efficiency, refill timeliness, and other metrics in order to determine effectiveness of the system.

As expected, physician groups, like the American Medical Association (AMA), have urged restraint and recommend that physicians remain involved in prescribing decisions. The AMA argues that some potential side effects or underlying health issues may not be effectively captured by the Doctronic. And even if the chatbot triggers human clinician oversight, how does that provide continuity of care for the patient from a provider that knows him well?

Will it work?

On first evaluation, this actually seems like a great idea. According to Doctronic and the Utah Department of Commerce, 80% of all medication activity are for prescription renewals. Further, it often takes weeks to obtain a medical appointment amid a national physician shortage. This may lead to reduced compliance, missed doses, and avoidable ER visits. This system might also allow for physicians to save time on low risk visits and spend more time with patients who have complex needs.

But it would be prudent to remind ourselves that we are still early stages of the AI transformation of medicine. There are a number of issues that patients who use Doctronic must contend with:

1. Technological barriers - For older patients, using a chatbot may not come naturally. They may also not be forthcoming about all their symptoms or they may not know if it is relevant. Even if the patient has a family member help with interacting with the chatbot, how can we be sure the patient is fully involved in their care?

2. Hallucinations - It has already been well established that LLMs hallucinate and supply false information. How often will Doctronic’s system be audited? How will it detect errors in judgement or hallucinations? This has not been well elucidated.

3. Security - Healthcare is rife with information leaks and system hacks. By partnering with Doctronic, Utah is allowing another third part, another middleman, to obtain and retain patient data. How will the state respond if Doctronic is hacked?

4. Prompt injection - Something as random as poetry has been shown to circumvent safety features on chatbots. It is too soon to tell if Doctronic is vulnerable to prompt injection, but people will inevitably try to bypass the chatbot. How is Doctronic protecting against this?

5. The role of pharmacists - One of the more interesting aspects of the prescription renewal debate that I discovered is that even pharmacists are questioning whether the system will detract from their authority. Some say that pharmacists should be given prescription refill authority, because they are in a better position to answer patient questions in person and catch medications interactions. Will anyone listen to them?

6. Liability - If Utah is partnering with a third party company, will they, as the government, allow themselves to be liable for an adverse reaction to medication? What are the consequences if Doctronic’s AI system is liable? Will any human face consequences? Is it just a slap on the wrist and a fine?

7. Access to healthcare based on class and status - One of the most consequential aspects of AI technology is the premium it will place on human interaction. With AI’s ability to mimic digital art and workflows, live performance and live interactions will become more expensive. We might be headed for a future where the poor or uninsured must interact with a digital AI doctor, while the wealthy and insured have access to real physicians in person because they can pay extra. What does that say about our healthcare system and our society if we dehumanize the patient-physician interaction? Does it matter?

Not the Only Game in Town

Doctronic is the not the only company exploring AI healthcare. Mass General Brigham (MGB) in Boston recently debuted their solution to the physician shortage: Care Connect, which was developed by the company K Health. As opposed to Doctronic, Care Connect focuses on patient’s medical history and chief complaint for common urgent care requests (colds, rashes, simple injuries, mental health issues, and chronic disease concerns). Care Connect’s AI tool then sends a suggested diagnosis and treatment plan to one of 12 employed physicians who are logged in remotely from around the country and available 24/7.

In this case, Care Connect is not making clinical decisions. Rather it is gathering history and forwarding the information to a physician who will ultimately sign off on the treatment plan. K Health has also partnered with the Mayo Clinic and Cedar-Sinai Medical Center in Los Angeles, giving it an impressive roster of partners.

Conclusion

Obviously it is still too early to predict the success of the partnership between Utah and Doctronic. My hope is that it does make a positive impact on the lives of patients and doctors, lowers costs, and improves efficiency.

But don’t we say that about every new service or technology, only to be eventually disappointed?

A recent NY Times op-ed makes the case that eventually, the AI companies, on which so many services rely on, will feel the pressure to turn a profit for their investors. At this moment, many users are not paying for their relatively simple AI needs. But we should not be surprised when the AI developers begin to charge for subscriptions, start running ads, and demand higher premium for agentic AI services like preparing our taxes or buying airline tickets. The same enshittification will undoubtedly happen with AI in healthcare.

Don’t say you weren’t warned.

It is generally accepted that the most complicated object in the known universe is the human brain. Remarkably, it is only about 1500 cubic centimeters (1.5 liters) and weighs about three pounds. The brain is composed of about 75 percent water. 60 percent of its dried weight is fat.

Scientists estimate that the brain has close to 85 billion neurons, with one quadrillion connections or synapses between them. For more palatable numbers, brain tissue the size of a grain of sand contains 100,000 neurons and one billion synapses.

At just two percent of the body’s weight, the brain consumes 20 percent of the body’s energy, primarily in the form of glucose. The brain generates about 25 watts of power (enough to power a dim LED light bulb). For context, power companies bill customers in kilowatt-hours (kWh), so we can use this calculation to determine the amount of energy consumed in 24 hours:

25 watts x 24 hours = 600 watt-hours or 0.6 kilowatt-hours

While the local electricity rates vary around the United States and the world, we can round the national average to 20 cents per kWh.

20 cents x 0.6 kWh = 12 cents

Therefore, the roughly 50,000 daily thoughts that drive our ability to read, dance, sing, paint, make love, operate machinery, calculate, reason, experiment, fight, or perform surgery are driven by a gelatinous, easily deformed structure about the third of the size of an American football and costs approximately 12 cents a day.

Of course, generating the electricity to perform the actions behind the thoughts are more expensive but not by much: water, oxygen, sleep, and a few well-balanced meals a day can be sufficient.

Meanwhile, a single metal and silicon-based H100 GPU by NVIDIA requires 700 watts. If we perform the same calculation as above, this works out to $2.85 per day. In other words, the GPU needs 28 times the energy at 23.75 times the cost.

Further, it has already been well documented how expensive it is to train AI models, to collect training data, and to establish the complex data infrastructure of massive cooling systems and high-voltage power. The brain self regulates at a steady 37 degrees Celsius.

Data centers may contain upwards of 100,000 GPUs. Even with all that processing power, we’re still waiting for robots to dance like us let alone correctly identify all the structures.

Therefore, the brain is not only the most complex entity in the universe, it is also the most energy-efficient processor in the known universe.

So what kind of audacity and hubris do humans possess that made us think we can achieve any sort of artificial general intelligence?

Of course, our brains are infamously fallible. Research has proven how we easily we forget events, how easy it is to implant memories in witnesses, how quick we lose track of time, how often we become overwhelmed when multitasking or stressed or hungry or fatigued.

Yet, our brains are incredible problem solvers. If we come across a new type of door handle or even a website, it does not take long for us to figure out how it works. A couple of rounds of trial and error and we are soon up and running. Maybe we solve it on our first try. Meanwhile, a robot or agentic AI may never figure out how to open the door or order the pizza. It will also sooner hallucinate.

When the AI and LLM hype is at its peak, when the AI CEOs and analysts and gurus are breathlessly espousing that the transformation of humanity is just around the corner, it sounds to me like the claims of every religious and cult leader since the dawn of time.

“The future is coming! The fall of mankind is at hand. But we possess a power that can provide salvation. It can cause great harm but also tremendous opportunity. You must trust us to harness and wield its strength. Our technological elixir can do everything you can but exponentially better. You can even have sex with it! We alone can save you! Also, you’re not going to have a job in ten years, but you can worry about that later!”

Anytime in human history where we have purported to possess awe-inspiring, God-like powers, these claims are eventually debunked. Every. Single. Time.

Yes, we possess incredibly powerful weapons of annihilation, but destroying something is easy. Building something is hard. An AI robot can shoot you, but can it repair the damage?

Once again, who do humans think we are to consider a possibility that we are a few short years from replicating anything close to the human brain? Can you name another organ that we have effectively replaced that does not come from another human?

At best, the medical technology that we do possess does an inefficient job assisting in the body’s functions. A ventilator might help a patient breathe, a cochlear implant may help a patient hear, and dialysis will do the work for a patient’s kidneys. Even the left ventricular assist device (LVAD) will augment the work of a patient’s heart. But these are devices and technologies with complications and side effects and limitations.

Even implanted brain-computer interface implants, as important as they might be for those contending with the limitations of paralysis or paraplegia, will not enhance the brain to better than it was before. At least not anytime soon. But it will not revive or replace a non-functioning brain.

Of course I may have to eat my words one day. But chances are that I will be long dead before I have to do so. In the meantime, it is best if we proceed with caution and humility. AI has the awesome ability to destroy so we must respect its power. But the invention of fallible AI technology required the collective processing power of the most complex and efficient processor in the known universe over generations. We don’t even know how the brain fully works, let alone a large language model.

Once again, it bears repeating: who do we think we are?

I am interested in health and vitality. Not just health as in biology, physiology, and the human body, but also our intellectual, environmental, and our civic health. Broadly, this includes the health of our planet, our institutions, our creativity, our technology, our relationships, our society, and our purpose. One of the essential themes of this project is that everything is interconnected. I am interested in understanding how these dynamic forces interact with each other.

The universe is a collection of systems that are intertwined, run both in parallel and in sequence, and reside within one another. We do not live in a physics problem, where systems are isolated and independent. On our planet, the health of one system will affect another. If we want to improve our lives, our society, and our world, it is best if we try to understand how everything fits together and affects our health from a variety of perspectives.

So that where “polymath health” comes from. But what about AI? What does artificial intelligence have to do with all this?

Well it is apparent that AI will be the defining technology of our century. It has the chance to completely upend our lives, our health, our interactions, our government, our jobs, and whatever meaning we get from life. Indeed, it is already transforming our society.

The questions and possibilities are seemingly limitless. How can we use this ground breaking technology to improve our health and bring good to our lives? How do we prevent and protect ourselves from the harm that can come from it when it is used for nefarious purposes? What might we lose from our natural intelligence and world if we give ourselves up completely to the artificial? What lessons can biology and nature teach us when it comes to optimizing our health or living our lives? How do we differentiate between hype, hubris, opportunity, and threat?

Inevitably, this project will evolve with time. It will continue to improve and grow. There is no set destination, so I am excited to see where it leads. Thanks for joining me.