healthcarebriefing.com

Category: First Opinion

  • Opinion: How CMS plans to fix a major flaw in the 340B drug discount program

    Opinion: How CMS plans to fix a major flaw in the 340B drug discount program

    By Casey B. Mulligan

    Oct. 8, 2026

    Mulligan is the chief economist and chief regulatory officer at HHS.

    Some Medicare patients owe more in coinsurance than the hospital paid for their drug. That arithmetic sounds impossible. It is a consequence of the 340B drug discount program and Medicare’s payment formula. Now the Department of Health and Human Services, where I work, is proposing to narrow that gap.

    Congress created the 340B drug discount program in 1992 to financially support safety net hospitals that serve low-income patients. Drug manufacturers must give them steep discounts. But Medicare, Medicaid, and commercial insurers generally reimburse them at the full amount, allowing these hospitals to keep the spread above the discounted purchasing costs.

    A new report from the Department of Health and Human Services shows other unintended consequences of this well-intended program. The spread grows with the price of the drug. It therefore rewards hospitals for using more and costlier drugs and makes physician practices and infusion clinics more valuable after a 340B hospital acquires them. The patient’s treatment needn’t change. The owner and billing site do.

    The 340B program has exploded: Eligible drug purchases rose from $12 billion in 2015 to $81 billion in 2024, mostly at hospitals, as shown in the report. Those purchases are now larger than net Medicare Part B and Medicaid drug spending overall and, among public prescription drug programs, are smaller only than Medicare Part D.

    A 340B hospital that buys another hospital or a physician practice can register more off-campus clinics and contract pharmacies and capture more spread. Hospital off-campus clinic sites grew from about 7,000 in 2013 to more than 34,000 in 2023, often in neighborhoods that are wealthier and better-insured than that of the parent hospital. Contract pharmacies grew from roughly 1,300 in 2010 to more than 31,000 by mid-2026, dominated by the largest chains.

    With greater bargaining power after acquisitions, hospital systems can negotiate higher commercial prices for health care services, which can in turn raise premiums. Newly acquired hospital outpatient sites can also add facility fees to non-drug services. The effects of 340B show up as larger, more expensive hospital systems and a bigger bill for patients, employers, and taxpayers. The program’s ripple effects contribute to the broader problems facing the U.S. health care system.

    Medicare payment for Part B drugs is typically based on average sales price (ASP). Beginning in 2018, the first Trump administration narrowed the spread in Medicare by paying hospitals for 340B drugs’ ASP minus 22.5% rather than the usual statutory standard of ASP plus 6%, based on an earlier estimate of their average minimum discount.

    Overall outpatient drug payments rose nearly 15% a year from 2015 through 2017. Once the lower rate was fully in effect, inflation-adjusted payments largely plateaued. Payments for the affected 340B drugs held near $7 billion from 2019 through 2021. It is clear that Medicare payment policy influences outpatient drug spending.

    The Supreme Court ruled in 2022 that HHS couldn’t vary payment rates by hospital group without first conducting the acquisition-cost survey required by statute. The court didn’t find that hospitals were economically entitled to the markup. It found that HHS lacked a procedural prerequisite. After the lower rate ended and ASP plus 6% returned for these drugs, annual payments for the affected drugs rose from $7.1 billion in 2021 to $10.7 billion in 2023 — a $3.6 billion increase. Overall outpatient drug payments rose from $14.3 billion in 2021 to $17.6 billion in 2023.

    Restoring the higher rate meant paying hospitals more for the same drugs. The $3.6 billion increase combines that direct payment effect with induced changes in hospitals’ dispensing quantities and drug manufacturers’ underlying prices. The larger spread also strengthens hospitals’ incentive to choose costlier drugs and move treatment into 340B billing sites.

    Off-campus hospital departments benefited too. From 2021 to 2023, payments for their affected 340B drugs rose about $800 million even as claims fell and the number of beneficiaries remained similar. Although these amounts do not tell us how many practices hospitals acquired because of the markup, they indicate the additional financial incentives for acquisitions.

    On April 18, 2025, President Trump directed HHS to obtain the cost information the statute requires. CMS surveyed hospitals from January through April 2026 and has proposed paying for affected 340B drugs at ASP minus 33.4% in 2027. CMS estimates that Medicare drug payments would fall by $4.55 billion in the first year and beneficiary drug copayments by $1.15 billion.

    For a drug with an ASP of $1,000, a beneficiary’s standard 20% coinsurance would fall from $212 to $133.20. The prescription hasn’t changed. The patient’s share of the hospital’s bill has. Narrowing the spread also weakens the incentive to steer patients toward costlier drugs and to acquire physician practices and infusion clinics — the very behaviors that have made care more expensive.

    There’s an important lesson here. A small, well-intended discount for safety net hospitals became a large institutional subsidy with expensive unintended consequences. Under the proposed rule, hospitals would still receive their 340B drug discounts. What would change is Medicare’s payment for those drugs at affected hospitals. Congress provided for a survey of hospitals’ acquisition costs, and CMS says its proposed rate would “more accurately align Medicare payments with hospital drug acquisition costs.”

    Medicare patients should not pay coinsurance as though the discount did not exist.

    Casey B. Mulligan is the chief economist and chief regulatory officer at HHS.

  • Opinion: I’m a doctor. Here’s what I want the public to know about AI tools and health care costs

    Opinion: I’m a doctor. Here’s what I want the public to know about AI tools and health care costs

    When Blue Cross Blue Shield recently published research suggesting that hospitals use artificial intelligence tools to extract more money from insurers without providing additional care, people were mad. Amid growing concerns about AI, one glimmer of hope has been its potential to improve health — and yet, here it was: inflating health care costs without yielding any benefits.

    As a doctor who has used the kinds of tools being criticized, I’m frustrated, too: We can document honestly and still wind up inflating hospital bills. This is because comorbidities function as an on-off switch with respect to billing, even though they affect patient care in nuanced ways.

    Although the problem warrants a systemic fix, the best short-term solution is for hospitals to create more nuanced documentation tools, and for clinicians to exercise judgment when using them.

    What did the research show?

    BCBS suggested hospitals are using automated tools to document secondary diagnoses — like anemia, malnutrition, and low sodium — which allow them to charge more for caring for these patients.

    For instance, a patient might be admitted for diverticulitis (primary diagnosis), but found on post-op labs to have acute blood-loss anemia (secondary diagnosis). Hospitals can use these secondary diagnoses to label a patient as having a comorbidity or a complication, allowing them to bill more for the hospitalization, whether or not they treated the patient’s anemia. These “bump codes” have cost BCBS hundreds of millions of dollars over the past two years, and these costs are presumably passed along to patients in the form of higher premiums.

    BCBS claims that bump coding has become more widespread because hospitals are using AI tools to identify secondary diagnoses. Automated tools can skim a patient’s chart and prompt physicians to document that the patient is, for instance, anemic, and what they are doing about it.

    Are doctors to blame?

    A few weeks after I started medical residency, I began getting emails from hospital administrators I had never met, whose signatures said they were risk adjustment specialists. The emails were generic and polite, advising me to review clinical documentation integrity (CDI) queries on inpatients I was caring for. If I didn’t respond quickly, the request would be forwarded to the attending physician supervising me. A typical CDI query would point out that my note hadn’t commented on, say, a patient’s low sodium level, and would ask me to document that they had low sodium and what (if anything) we were doing about it.

    Like most residents, I was already spending far more time on my computer than I was with my patients. Soon, CDI queries streamed into my inbox with such regularity that I turned to a more senior doctor to ask how I might stave them off. That doctor showed me an automated tool that I could include in my notes. This tool would assess a patient’s lab results, automatically detect any additional diagnoses a patient had (like anemia, malnutrition, or low sodium), and prompt me to specify whether we were monitoring, evaluating, or treating.

    I assumed it was lucrative to the hospital for me to document these diagnoses (why else would they be paying risk adjustment specialists to email me about them?), but we almost always were at least monitoring the secondary diagnosis in question, and I accurately documented as much.

    Including this tool’s outputs in my notes kept the risk adjustment specialists from hounding me and won me back several minutes of each day — enough time to poke my head into a patient’s room before I went home, or call a patient’s family member to give a quick update. I always answered the questions included in my tool honestly, but in the back of my mind, I wondered whether I was doing my patients a disservice: Was I inflating the cost of their hospitalizations without actually improving their care?

    Where bump codes come from

    In general, hospitals are reimbursed for hospitalizing patients on the basis of diagnosis-related groups, or DRGs. DRG payments are set by the Centers for Medicare and Medicaid Services by analyzing claims data. (Medicare uses DRGs to set payments, while many commercial insurers, including BCBS, use the DRG system to negotiate reimbursement rates with hospitals.) Let’s say a patient is admitted with diverticulitis requiring surgery (typically billed as DRG 331 if there are no complications). CMS has estimated how much it usually costs to care for one of these patients and insurers prospectively pay hospitals on that basis; a hospital generally won’t get reimbursed more for keeping the patient longer, administering more drugs, placing more IVs, and so on.

    However, some hospitals care for sicker patients than others. The typical diverticulitis patient at a tertiary medical center in a poor area will require more support than a diverticulitis patient at a community hospital in an affluent area. The government does not want to penalize hospitals that see more complex patients, and so it allows hospitals to seek additional reimbursement forcomorbidities or complications that increase the cost of their care.

    This is where bump codes come from: Hospitals can bill more complex diverticulitis admissions as DRG 330, which is reimbursed at a higher rate. Not all comorbidities or complications count: In 2007, the last time DRGs were substantially changed, CMS analyzed claims data and used this analysis to decide which conditions qualify and set reimbursement rates. CMS pays hospitals more for patients with conditions like low sodium because it has found that it costs hospitals more to care for these patients (perhaps because they tend to be hospitalized longer and more often require ICU admissions).

    However, although CMS adjusts DRG payments from year to year, the list of bump codes hasn’t been overhauled in nearly two decades.

    Bump codes can help identify complex patients and ensure hospitals are paid more for caring for them. These are good things, in part because they reduce hospitals’ incentive to seek out healthier, more advantaged patients. While BCBS claims that hospitals are increasingly using these codes to bill more without actually providing more care, hospitals would presumably counter that they were being insufficiently reimbursed before, and that their coding is accurate. After all, the BCBS study does not claim that hospitals are engaging in fraud; the question is just whether the lab abnormalities hospitals are billing for warrant higher reimbursements.

    So who’s right?

    This issue boils down to the following question: How much more does it cost to care for a patient with a secondary condition, and do current coding rules accurately reflect this?

    The answer is that it depends. For instance, not all low sodium levels are created equal. If a patient briefly has a sodium level that falls below the normal range, but their sodium immediately corrects once they’re able to eat again, they probably don’t require any additional resources, and hospitals shouldn’t be reimbursed more for caring for them.

    By contrast, a patient with dangerously low sodium levels requiring an ICU stay, frequent lab checks, and carefully titrated IV fluids undoubtedly does use more resources.

    This leaves doctors in a tough spot: Risk adjustment specialists ask us to document real diagnoses that we know could increase the cost of a patient’s bill, but we have no idea by how much. Moreover, there’s no medical school module on how much any of these comorbidities affects the true cost of the hospitalization. Against this backdrop, it makes sense that doctors try to just tell the truth: yes, this patient has low sodium, and we are monitoring it, even though this probably allows hospitals to code some patients as being more resource-intensive than they truly are.

    What we can agree on

    This entire enterprise — risk adjustment specialists spamming doctors, who are already nose-deep in documentation burden, to document diagnoses that often do not affect patient care, while insurers publish reports that fail to capture the nuance of the issue, leaving the public angry and confused — highlights a more general problem with respect to how the U.S. health care system works.

    By design, hospitals and insurers are meant to compete with each other, with hospitals trying to collect as much money as they can for providing care, and insurers trying to pay as little as possible. Hospitals bolster their side by hiring risk adjustment specialists and employing AI tools, while health insurers hire payment integrity specialists and employ different AI tools.

    Patients lose out because the administrative bloat inflates the cost of medical care, while making costs less transparent and harder to contest. Ultimately, the fix should be structural: At a minimum, we need to overhaul the list of secondary conditions that are eligible for bump coding. CMS attempted to do this in 2019, when it proposed downgrading or eliminating many bump codes based on new data showing they were not associated with increased costs. But hospitals pushed back, and the effort eventually stalled.

    What should we do in the meantime?

    Secondary diagnoses are supposed to be billed only when they are clinically significant — for instance, when they require further evaluation, treatment, diagnostic procedures, or monitoring. But in practice, the bar for clinical significance is probably lower than it should be. (Even rechecking a lab may qualify as monitoring.) Accordingly, I suspect the BCBS study is tracking a real problem, and that much of the documentation of secondary diagnoses isn’t linked to patient complexity.

    Early in residency, I was caring for an immunosuppressed patient who kept spiking fevers. My colleagues and I wondered whether he might have an indolent infection lurking somewhere and ordered an echocardiogram to rule out endocarditis. His heart valves showed no signs of bacterial growth, but trace backflow was detected across one of his valves, a finding present in most people.

    Soon, I received a CDI query inviting me to document his valve dysfunction in my note. I did not do this, because his incidental valve finding was irrelevant to his hospitalization. Instead, I typed out a response indicating as much. However, it would have been easier and quicker to click through the check boxes, and, despite answering them honestly, potentially inflate the patient’s bill.

    At least some CDI tools built into electronic medical records — including the one I have used — default to including secondary diagnoses in ways that are likely to lead to bump coding. One way to improve the status quo would be to ensure that such tools give physicians the option to document whether a secondary diagnosis is clinically significant or not, rather than simply asking them how they are addressing it (on the presumption that it is clinically significant).

    Hospitals should build such tools, and regulators should ensure they do. In the meantime, I’ll exercise clinical judgment when using my CDI tool — if the low sodium seems important, I’ll document it, and if it doesn’t, I’ll politely decline when a risk adjustment specialist invariably asks me to.

    Leah Pierson is a resident physician.

  • Opinion: Will Claude ever win a Nobel Prize for medicine?

    Opinion: Will Claude ever win a Nobel Prize for medicine?

    On Monday morning, the Nobel Prize for medicine was announced. This year, I wonder: Will Claude or a future artificial intelligence agent ever win the most prestigious award for medical discovery?

    Imagine a future AI agent responding to a scientist’s insightful prompt by successfully identifying a completely novel approach to curing cancer, along with a detailed and actionable plan for implementing it. Given the toll of cancer on our species, I suspect few today would consider that a bad outcome.

    But as a former medical school dean who was regularly called upon to adjudicate complex issues of credit, I’m uncertain how credit for such a future world-changing discovery would or should be assigned. Until today, human scientists have been the agents responsible for discovering cancer cures, and for doing so they (and key associates) have been showered with academic credit and awards, perhaps even a Nobel Prize. Are we now entering a world in which some (or all) Nobels will be awarded to Claude, ChatGPT, or their AI successors, rather than human scientists? And if so, would it matter?

    The Nobel statutes don’t mention “human” but refer instead to “persons.” In 2024, two of the three winners of the Nobel Prize in chemistry were given credit for their work on the AlphaFold model at DeepMind — even though many more were involved. So without changes to the rules, a future prize for a discovery made by AI would most likely go to some “persons” given credit for making the AI possible — at least at first. But the committee may elect to change the rules or interpret them differently. The Nobel Peace Prize, for instance, has gone to organizations rather than persons.

    Today, credit for discovery plays a central role in the reward structure of science. As “the currency of the realm,” credit powerfully influences the norms and practices of the research ecosystem. It’s required for scientists to secure appointments, promotions, funding, access to students, and other necessities. It’s also a mechanism for evaluating research impact and providing accountability for research claims.

    It didn’t always work this way. The current credit ecosystem arose in the mid-19th century, as science emerged as a profession alongside the rise of the modern scientific journal. It’s been further modified by many factors, from today’s online publishing platforms to new means for assessing and quantifying scientific impact, and the increasingly interdisciplinary and team-based nature of research.

    As we begin to acknowledge AI as a generator of scientific discovery, will the approach to scientific credit be transformed as well? Current and future scientists would certainly like to know.

    Despite many unknowns, it’s hard to imagine the status quo holding. In perhaps the most likely outcome, humans will continue to be responsible for many discoveries, often aided by AI along the way. The credit ecosystem will continue to recognize these human achievements, while finding ways to appropriately recognize the contributions of their AI collaborators.

    But if AI itself is recognized as the responsible agent for some discoveries — even profoundly important ones, such as a cancer cure mentioned above — how would credit be handled? Would assignment of credit as we know it today even matter?

    In that world, humans might retain (for a while) a desire to acknowledge the role of AI in those discoveries, even bestowing previously human-only awards upon the responsible AI agents. Perhaps some credit would also go to their human creators, those putting forth insightful prompts, and those creatively implementing the work.

    But in a world of predominantly AI-driven discovery, the current role of credit as an incentive for the human research ecosystem would dramatically change.

    In considering this future, we must consider another role that credit plays. Beyond serving as an incentive and accountability mechanism in the human academic research ecosystem, there are commercial incentives for discovery, and credit plays a different role in that setting. A major commercial incentive for biomedical research involves the awarding of patents for discovery, and today only humans are eligible to be named discoverers. If future laws change to allow recognition of AI agents and/or their human creators as discoverers (with or without human collaborators), how might this affect accountability, ownership, and the ability to profit from discoveries?

    For research conducted today in both academic organizations and for-profit companies, ownership of intellectual property is generally held by these organizations rather than the human discoverers now named in patents. Patent law might evolve to allow organizations to name AI agents — even proprietary AI versions — as discoverers. But assigning credit would remain complicated to the extent humans will likely play creative roles in successfully implementing AI-generated ideas, though efforts are underway to create AI-automated research systems.

    Today, the science community is composed of humans who seek and often compete with others for discovery as well as credit, and this feature of scientific culture is believed to play an important role in the success of the enterprise. As AI becomes a key enabler and even primary driver of discovery, the current culture of credit will evolve to accommodate that reality. Changes to the credit economy may be among the many trade-offs we accept to realize the benefits of AI-driven discovery, in biomedical research and other domains of human activity.

    Whether an AI agent will ever receive a Nobel laureate is, at least in part, a question about what jobs we want credit to perform. If it’s mainly an incentive for human scientists and the overall enterprise, its role will change as machines play a greater role in discovery. If credit continues to influence allocation of trust, resources, and responsibility, those needs will persist, necessitating decisions on how best to manage them in an AI dominated research ecosystem. This may be further complicated by recent debates about whether AI agents may evolve into moral agents whose status would justify their receiving credit for “their work.” 

    I don’t know how these questions will ultimately be resolved. And I suspect the answers will differ across diverse domains related to research: human scientific careers, journals, funders, universities, biopharmaceutical companies, patent offices, and yes, prize committees. The current system was designed for a world of human discoverers. Though that day may never be fully behind us, we must begin serious conversations about what changes the age of AI science will bring to scientific credit.

    Jeffrey S. Flier is a Harvard University distinguished service professor and the former dean of Harvard Medical School.

  • Opinion: Claude analyzed my genome in 30 minutes. Now we need standards for the results

    Opinion: Claude analyzed my genome in 30 minutes. Now we need standards for the results

    Unpacking from vacation a couple months back, I decided to give artificial intelligence a task. I paused folding my laundry, sat down with Claude, and asked it to analyze the entirety of my genome.

    Back in 2009, this was a task that quite literally took a village of 30 people — including me — working together to interpret the whole genome sequence of a colleague, friend, and fellow scientist named Stephen Quake. This endeavor brought together experts in genetics, computer science, pharmacology, and clinical medicine.

    We built an engine specifically designed for genome analysis, examined millions of genetic variants, reviewed the scientific literature, and debated which findings might matter most for his individual health circumstances.

    The resulting study, published in The Lancet in 2010, was among the first attempts to interpret the complete genome, and the first to translate that analysis into personalized medical recommendations. (About a year later, our team used a similar framework to analyze my own genome, though it was never officially published.)

    This month, Claude attempted the same task — sans the 30 people and nearly yearlong expenditure.

    In a prompt of 180 words, I asked Claude to follow the same framework our team had set forth back in 2009, probing it to identify rare variants that could cause an inherited disorder, genetic abnormalities that could impact my response to different medications, and whether it had enough data to estimate my risk for common diseases.

    Though it was a fairly simple prompt, the task was anything but. My genome file is more than a decade old, was generated using an earlier version of the human reference genome, and lacks information available in newer sequencing formats. These limitations could easily produce an analysis that is misleading or inaccurate.

    Despite that, Claude was able to re-identify many of the most important findings from our original analysis, including the most consequential result in my genome: I carry two copies of the APOE ε4 variant, which is associated with an increased risk of Alzheimer’s disease. It found no major gene-disrupting variant linked to a rare inherited disorder, matching the outcome reached by our original team. It also flagged variants in the DPYD and CYP2C19 genes, which can influence how we respond to certain medications. I asked it to calculate genetic risk scores for common diseases, but it rightfully warned me that my file did not contain enough information to do so reliably.

    Amazingly, this entire interaction used only around 400,000 tokens and cost approximately $5. Most of the half-hour was spent just waiting for scientific databases to load. All in all, a task that once took nearly a year and required a large, specialized team can now be attempted at home for roughly the price of a 2-liter bottle of soda.

    That accessibility opens doors once thought unimaginable. And as these doors open to the general public, the scientific community has a responsibility to develop clear standards for what makes up a medically reliable genome in the first place.

    A genome is not a definitive record

    It is tempting to think of a genome sequence as a permanent record that can be preserved forever after one read. In reality, though, every genome analysis is influenced by multiple factors, such as the reference genome used for comparison and the technology used to sequence the DNA itself.

    Sequencing technologies have seen enormous improvement over the past couple of decades. Scientists can now examine sections of DNA that were once inaccessible and detect forms of genetic variation that older methods may have missed.

    But limitations remain. Almost all existing genomes were produced using a technique called “short-read sequencing,” which breaks DNA into millions of fragments and uses software to reconstruct where these fragments belong. Though this approach works well across much of the genome, it can be less reliable in the regions where sequences can repeat, resemble one another, or contain large structural changes.

    Though difficult to sequence, those regions are by no means irrelevant: Disease-causing variants can occur where conventional sequencing struggles to pick them up. In turn, an interpretation may report that no concerning variant was found without making clear that the relevant gene was not fully examined.

    Another barrier is the reference genome itself. To date, most analyses have compared patients’ genomes against a standard DNA sequence constructed from a limited number of people. As it stands, that baseline does not adequately represent the full diversity or complexity of human genomes, making some variants more difficult to detect, especially in populations that remain under-represented in genomic datasets.

    Newer references now represent many different versions of the human genome through branching paths rather than forcing everyone’s DNA onto one template. This can improve variant detection, especially in complex regions and across diverse populations. Yet laboratories, databases, and clinical systems have not adopted them consistently.

    The call for clear standards

    Now, with the general public able to readily use AI for genomic interpretation that may impact the course of their life and future health decisions, we have never had a more urgent need to develop standards to determine what a high-quality medical genomic interpretation should actually include. 

    Current quality checks are efficient at detecting simple genetic changes, such as an altered DNA letter or a small piece of DNA that has been removed or added. Medically important changes, however, can be much trickier to detect. I liken it to proofreading a book: Finding a misspelled word is quite straightforward, while catching a sentence or paragraph that may not be in the right place takes a bit more nuance.

    A medical-grade genome test should therefore be evaluated on whether it can reliably detect the full range of genetic changes that may affect a person’s health, not only the simplest ones.

    This is an ask that will require a diverse set of exceptionally well-characterized genomes that laboratories, researchers, and AI developers can use as reference materials. These genomes should reflect the full sequence of both chromosome copies and the diversity of the global population, rather than relying on one person or one ancestral group as a universal standard.

    It is also important that the scientific community develops a shared, regularly updated catalog of medically significant genes that are historically difficult to sequence. For each gene, the catalog would ideally describe its technical complexity, identify which methods can reliably analyze it, and document when a specialized test is required.

    Standards must also reflect the task at hand. A test that is not sensitive enough can miss a finding entirely. For example, an inherited variant is present in nearly every cell, making detection fairly clear. Circulating tumor DNA, by contrast, may make up only a tiny fraction of the cell-free DNA in a blood sample. Each application therefore demands that clear minimum thresholds be set for what can be detected and how reliably, and what conclusions the result can support.

    The standards cannot lag behind the technology

    We are heading toward a world in which companies routinely give consumers access to their genomic data, including variant call files (digital files that list the genetic differences identified in a person’s DNA), that they can upload directly into systems such as Claude or ChatGPT.

    Consumer genetic testing was an important first step toward this kind of access, but the shift now is different in scale and interactivity. Services like 23andMe have largely focused on a relatively limited set of common genetic variants.

    A whole genome contains vastly more information. And with AI agents, consumers are no longer limited to a fixed interpretation of their results. They can ask follow-up questions, request additional analyses, and continue investigating particular findings. That kind of ongoing interaction with one’s own genomic data was simply not possible before.

    Now, imagine someone uploading their genome file into a system such as Claude or ChatGPT. Minutes later, alone at their laptop, they are told by AI that they may be at risk for Alzheimer’s disease or cancer.

    This may be true, but there is also the chance that the result is based on incomplete data, and the finding is much less definitive than the AI may suggest. Either way, they may feel completely blindsided, and there may not be a clinician available to talk through the result, whether it should be confirmed, or what the next steps should be.

    This is one of the most consequential challenges of accessible genomic interpretations. Interpretations are available before (or completely without) additional care and support.

    In my own experiment, Claude was able to recognize that my genome file did not contain enough information to make reliable conclusions regarding genetic risk scores for common diseases. This is encouraging — a trustworthy system should be able to pump the brakes and realize when data inputs are too unreliable to make accurate estimations. Even so, that judgment should also be governed by a shared medical standard, rather than left to the discretion of AI developers.

    Putting that standard into practice will require coordination across specialties. Professional societies and standards organizations will need to develop and regularly update the technical requirements. Labs and AI developers should have transparent disclosures on what their systems detect reliably, and which findings still require confirmation. Health systems, meanwhile, will need practical ways to support the increasing number of patients who may arrive with AI-generated results, including confirmatory testing and genetic counseling.

    In 2009, the scarce resources were expertise, time, and money. Today, however, what’s scarce may be reliability. As we embrace the new bioinformatics revolution, marked by incredible speed and accessibility in data interpretation, we must urgently work towards developing standards that increase the likelihood that a casual user can upload their genomic data into an AI system and receive a trustworthy response.

    Euan Ashley is chair of the Department of Medicine at Stanford University and a physician-scientist whose research spans genomics, precision medicine, digital health, and artificial intelligence. Rebecca Handler, a science writer and researcher at Stanford’s Department of Medicine, contributed to this piece.