What This Paper Covers
The recent advancements of artificial intelligence (AI) within medicine have coincided with a growing anxiety over the financial sustainability of the nation’s health care system. In 2022, U.S. health care expenditures reached 17.3 percent of the gross domestic product (GDP), which is nearly three times the share of GDP that had been spent on health care in 1970. The share of GDP expended on health care is expected to grow further given rising medical costs and the aging of the American population. Given that this spending trajectory cannot continue without harming family finances as well as other governmental budgetary commitments, AI should be explored as an instrument for health care cost reduction alongside existing efforts to improve clinical quality. This paper avoids popular hyperbole and occasional fearmongering attending contemporary AI discussions and assesses the technology’s realistic potential for savings.

What We Found & Why It Matters
Health care AI is not a singular technology but a collection of software-based processes being applied inside and outside clinical settings. The results are often improved quality or reduced labor, and in some cases both. However, when considering the improvements AI can bring to health care, significant obstacles exist between provider-side savings and the transmission of those savings to consumers. One of the most important of these barriers is the pre-determined payment rates that health insurance plans negotiate with health care providers. This payment model may not adjust payments downward, and certainly not in a timely manner, based on localized efficiencies. Likewise, medical practices have no incentive to lobby health insurance plans for less compensation in circumstances where AI reduces labor costs.

A similar dynamic applies to medical quality improvements. Clinicians’ adoption of AI to improve patient diagnoses and outcomes involve additive technology costs for the providers delivering medical care and these costs will be passed along to patients. In instances where AI systems reduce labor, clinicians will likely oppose reduced payments rates, especially if providers not using the technology would retain higher legacy payment rates. Beyond the considerations of payment, improved outcomes (e.g. from earlier disease detection) resulting from AI may reduce patient treatment costs, though this is likely to be disease-specific rather than system-wide.

The savings potential of autonomous care—the self-service delivery of a medical service via an AI system without clinician assistance—differs from the prospects of productivity gains and quality improvements. Every time a health care service can be performed independently of a clinician, a sizable expense can be eliminated. Because individual states do not license medical devices, there is also the possibility of AI-enabled health care devices competing for patients across the nation and lowering prices through expanded consumer choices and greater market competition.

There should also be savings from efficient scaling by autonomous self-service applications. Unlike humans, software can increase its delivery of a service with minor marginal costs. Marginal cost—the additional expense for software to supply a new service for a patient—is often quite low because the same programming is reused and computer infrastructure is relatively inexpensive to adapt for more simultaneous user sessions. Scaling human-generated service is quite expensive in comparison because the doubling of capacity often requires a doubling or more of employees.

What We Recommend
While AI is only at the dawn of autonomous self-service applications, the savings possibilities are substantial. However, AI’s future success in producing meaningful reductions in health care expenses depends on multiple factors, not the least of these are the surrounding regulatory environment and intellectual property conditions. To preserve the cost reduction potential of autonomous care, this paper recommends that clinical assistance not be made a regulatory requirement when an AI medical system can empirically demonstrate three criteria:

• Accuracy levels (or patient outcomes depending on the nature of the application) equal to or exceeding the average rate for clinicians performing the same function
• No amplification of health risks as compared to when the same function is performed by a clinician
• Output communications that are deemed to be comprehensible and actionable for the patient without the additional explanation from a physician

In addition to these regulatory recommendations, the paper recognizes that the nation’s intellectual property conditions will be just as important for the flourishing of AI health care solutions. Accordingly, the paper recommends that the government:

• favor trade secrets as the preferred method of protection when algorithm claims are heavily indebted to existing theory and practice in the AI field and approach process patents with both an open mind and skepticism,
• protect AI inventions that are materially novel and not logical extrapolations from foundational math and theory in the public domain, and
• incentivize self-service medical AI with high-cost savings potential with a faster patent review and approval process

The Biden administration’s late 2023 “Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” advocated a coordinated, government-wide approach to AI development. The approach itself consists of eight principles to which executive departments and government agencies are expected to comply. Those principles would greatly benefit from the assimilation of the regulatory approach contained in this paper. Likewise, this paper’s intellectual property recommendations would improve the intellectual property guidance the order stipulates be published for patent examiners and applicants addressing inventorship and the use of AI.

Amid marketing hyperbole and occasional fearmongering by critics, artificial intelligence (AI) has made inroads across health care at remarkable speed. The technology’s many applications include drug discovery, medical imaging, patient monitoring, surgery, and personalized treatment designs. These examples, though, do little justice to the scope of AI’s growing medical influence. In less than a decade from now, it may be easier to name the medical fields not being reshaped by AI than those that are.

AI’s rise within medicine coincides with growing anxiety over the American health care system’s fiscal challenges. In 2009, the Social Security Advisory Board warned the nation’s health care cost trajectory as “unsustainable” and “perhaps the most significant threat to the long-term economic security of workers and retirees.”¹ At that time, the United States spent approximately $2.5 trillion on health care, which represented $8,160 per U.S. resident.² By 2022, the United States spent $13,493 per person on health care, totaling $4.5 trillion.³ This amount represented 17.3 percent of the nation’s gross domestic product (GDP).⁴ In comparison, the nation spent only 6.2 percent of its GDP on health care in 1970.⁵

The financial unsustainability of American health care suggests that, in addition to its potential for clinical improvements, AI should also be explored as an instrument for cost reduction. Such an inquiry is promising because AI, as shall be illustrated, has capabilities that can greatly improve price pressure in the nation’s health care system. While this hypothesis, made during this early period in AI’s medical use, awaits validation, its rationale is grounded in how AI has differentiated itself from previous medical technologies.

The term artificial intelligence does not reference a single technology but rather a category of software systems whose operations resemble human thinking to various degrees. Thinking in this context means both learning and reasoning. There are a variety of classifications and subtypes of AI, and, with respect to health care, we will review four categories that have been groundbreaking even in their nascent forms:

• Large Language Models
• Machine Learning
• Artificial Neural Networks
• Generative AI

Large Language Models

A Large Language Model (LLM) is the basis for the interpretation of verbal and written speech. Built through the analysis of massive data sets, an LLM stores information on the complex relationships among words, expressions, contexts, and grammar. These relationships, along with the proximity of words to one another within a given linguistic expression (e.g. a sentence or paragraph), inform a probabilistic analysis that constructs meanings from the semantic possibilities of the “tokens” being interpreted. A token, in the context of a LLM, may be a word, a phrase, a segment of a word, or even a character.⁶ For example, Figure 1 above is a tokenization of the sentence “Cleanliness is next to godliness” by an LLM system.⁷ The individual tokens are designated by unique background colors.

In the case of tokens smaller than words, they help an LLM understand words it has never encountered before (e.g. understanding apolitical without prior definition by modifying the known meaning of political by the prefix a). Through token analyses, an LLM enables computer systems to interpret speech in a procedure known as Natural Language Processing (NLP). NLP, in combination with other software processes, has already birthed a variety of human-to-computer automations within health care, though many are at the early stages of adoption. For example, chatbots (programs that interact with humans via conversations) can respond to patient questions outside of normal business hours and virtual assistants can perform administrative tasks such as appointment scheduling and patient intake. Dictation bots can produce transcripts of sessions between doctors and patients. In all these examples, an LLM facilitates linguistic interactions between humans and computers. However, the same technology can allow computer systems to autonomously acquire new information for itself by “reading” documents rather than having information added through programming efforts or other manual intervention.

Machine Learning
An LLM is a form of Machine Learning. Machine learning software produces decisions or predictions through algorithms and statistical inferences driven by the processing of large data sets. These inferences are typically refined over time through continual data analysis rather than prior programming or explicit instructions. Machine learning comes in a variety of forms other than LLMs and can be “supervised” or “unsupervised.” Supervised systems use labeled data sets identifying some information relationships, while unsupervised systems use unlabeled data where the system detects relationships on its own

One of the exciting features of machine learning is its capacity to detect patterns that go unrecognized by human analysts. For example, a collaboration between M.I.T. and Massachusetts General Hospital has produced a machine learning tool named Sybil⁸ that, using a single low-radiation chest scan, can predict lung cancer risk for the following six years without input from a radiologist. Dr. Lecia Sequist, medical oncologist, noted, “In our study, Sybil was able to detect patterns of risk…that were not visible to the human eye.”⁹ Sybil has, on occasion, detected early lung cancer signs that radiologists did not recognize until lung nodules were visible on scans years later.¹⁰ This could potentially direct personalized screening programs to make earlier cancer diagnoses and improve outcomes.

Artificial Neural Networks

Sybil calculates its lung cancer prediction using an Artificial Neural Network¹¹ (ANN), a subset of machine learning. Though there are several types of ANNs, a common feature is layers of artificial neurons (also called nodes). Like biological neurons, artificial neurons communicate with one another and operate within multiple layers (see Figure 2). However, an artificial neuron is an individual software module that, as part of its collaboration on a shared ANN task, receives and processes an input. The processing is a mathematical calculation that will determine whether the neuron will activate, i.e. pass data to one or more other neurons.¹² A numeric constant, known as a bias, may be added to the values being calculated to affect the neuron’s propensity for activation. When activation occurs, the result is weighted before it is passed to the next neuron. The weighting value affects the importance of the activated neuron’s output for the next neuron receiving it, and weighting itself is based on previous training of the ANN. This data point, in some cases, may also be modified by the results produced by other nodes.

The representation of an ANN in Figure 2 has individual circles signifying individual artificial neurons, and these are stacked in columns to represent separate layers. The lines emanating each individual neuron portray potential data exchanges between the neuron and other neurons belonging to the next layer within the ANN. These exchanges occur when the starting neuron is activated and passes an input to the next neural layer. The nature of a neuron’s activation determines to which neurons the activated neuron’s input is communicated. For a neuron receiving input from multiple neuron activations, the weights applied to the connections influence the importance of each input.

Most ANNs process data from the input layer to the output layer in a movement called feedforward or forward propagation. The reverse movement is known as backpropagation. Backpropagation, where the system works from predetermined outputs toward earlier layers, may be used to train a system to produce the desired outcomes by fine-tuning weights within neurons.¹³

The multiple layer architecture within an ANN is advantageous for performing progressive tasks. For example, in the case of an X-ray or medical scan, the first layer in an ANN may establish the edges or basic form of a physical feature (e.g. a lung nodule). Subsequent ANN layers, using those initial determinations, may identify more complicated structures¹⁴ as part of its overall pattern recognition. When the ANN has at least two intermediate layers of nodes between the input and output layers, the result expressed by the output layer is described as an instance of “Deep Learning.” As demonstrated by systems such as Sybil, Deep Learning may perform feats beyond the capacity of human clinicians.

Generative AI
Generative AI leverages ANNs to produce original creations such as the speech responses of chatbots. The creations produced by generative AI resemble the artifacts on which they are based, whether textual, visual, or acoustic. The characteristics themselves are procured through Deep Learning on large data sets and may, in some cases, be processed by a Generative Adversarial Network (GAN)¹⁵ to produce a new object of the same type. A GAN possesses a generator neural network for creating novel content and a discriminator neural network for the content’s evaluation. The discriminator is adversarial because it compares the new content to examples confirmed to be authentic. If the discriminator can distinguish the new content from the confirmed content, the generator must improve the content through subsequent iterations until the discriminator cannot reliably differentiate between the new content and the confirmed content.

Like LLMs, generative AI in health care is often associated with chatbots that can answer patient questions, thus improving matters of health care access as well as customer service. With respect to administrative work, this type of AI can automate front office customer service (e.g. appointment scheduling, patient intake, patient discharge instructions) and back-office administrative tasks (e.g. billing, prior authorization requests). Surprisingly, the types of operations allowing generative AI to imitate images and conversational language can also be used in the pharmaceutical industry. Drug manufacturers are using generative AI and Machine Learning not only to design new drugs but also predict which patients will benefit (i.e. be a drug responder) or not benefit from these medications.¹⁶

There is little dispute that AI technologies can improve health care, though improvements do not guarantee lower costs. In fact, the worst-case cost scenario is AI adding to existing expenditures as a result of medical providers purchasing new AI-enabled systems and passing those costs to consumers. To avoid this prospect, it is necessary to consider three ways AI may reduce medical costs: productivity gains, quality improvements, and autonomous care. Together these areas share considerable savings opportunities but have meaningful differences regarding the barriers existing between savings and consumers’ wallets.

Productivity Gains
Administrative activity is perhaps the most rudimentary of AI’s health care applications given the technology’s analytical sophistication. Nevertheless, administrative labor is estimated to account for 15-30 percent of total U.S. health care costs.¹⁷ New AI systems adapted for administrative tasks are promising to perform front-office and back-office work at greater speed than would be the case for human staff and should reduce the staffing needed for such work. Prior authorization requests, for example, consume considerable labor on the part of physician offices as well as health plans. Such a request seeks advanced approval of a medical treatment (or medication) for a patient so it may be paid for under the terms of the health plan. Existing prior authorization processes are time-consuming and require considerable data entry on the part of the provider. An analysis by McKinsey & Company estimates that an AI-enabled prior authorization process could decrease manual effort by 50-75 percent.¹⁸

Similar opportunities for productivity gains abound in health care administration. Chatbots, in the form of virtual assistants, can handle a portion of patient questions and reduce the labor currently expended on inbound calls and emails. The same underlying technologies could also listen to patient sessions and populate electronic health records for doctors, subsequently generating summaries of medical events and health status,¹⁹ freeing time for physicians. AI can also benefit the back-office. For example, a machine learning system working with an LLM can automate claims processing and denial management,²⁰ which currently consume dozens of hours per practice every month.

In the pharmaceutical industry, AI is being used to design new drugs in less time than traditional methods, lowering development costs and accelerating time to market. For example, a collaboration between researchers at Stanford University and McMasters University employed generative AI to combat deadly strains of Acinetobacter baumannii that are resistant to existing antibiotics.²¹ Their generative AI system developed the chemical recipes for multiple new drugs that killed the pathogen.²² The system itself produced 25,000 different drug candidates in less than nine hours, giving researchers the luxury of manufacturing and testing the options believed most promising.²³ Given that the costs of drug development have been rising, generative AI’s ability to compress the design portion of drug development is timely and welcome. Aside from this, AI could be further beneficial in the process of finding drug targets—that is, the identification of a molecule associated with a disease that, in turn, allows a drug to alter the function of the disease.

The challenge with productivity improvements in health care is that they can be difficult to translate into lower costs for consumers. Among the reasons contributing to this situation is the nature of medical payment in the United States. With only 7.9 percent of the population uninsured as of 2022,²⁴ most Americans are covered by public or private health plans that determine payments on a mass scale. For a program such as traditional Medicare, there is the Physician Fee Schedule covering the reimbursements (subject to further adjustment for issues such as geography) for over 7,000 medical services. Private health plans, in contrast, have negotiated rates for their in-network clinicians as well as strict rules for the payment of out-of-network care. Such payment arrangements are not easily adaptive to localized efficiencies among individual health care providers. Further, health insurance itself conceals the full price of medical treatments from consumers and, thus, removes the financial incentive to compare prices and select the best value. This absence of consumer price shopping, in turn, diminishes price reduction as a motivation for physicians to attract and retain patients.

Adding to this challenging environment, pharmaceutical companies using AI have their own obstacles between productivity-improved costs and consumer drug price reductions. Drug prices are determined by a wide variety of factors, not the least of which is competition. In fact, a study by the Assistant Secretary for Planning and Evaluation (ASPE) examining prescription drug prices from 2017 through 2022 found “drugs with fewer or only one manufacturer, on average, have higher prices than drugs with multiple manufacturers, holding all else constant.”²⁵ With respect to the drugs with the highest costs (i.e. the top 10 percent in price per prescription), the study “found that the vast majority of the highest priced drugs had only one manufacturer.”²⁶ Single manufacturer drugs are often brand name drugs whose patent are active and protects them from other pharmaceutical companies making the same medication. While AI might reduce research and development costs for drug manufacturers, it does not determine the level of competition for their individual drugs.

There is the related question of whether AI’s ability to reduce drug development time and expense will increase the number of competing drugs within individual therapeutic classes. While increased competition is a definite possibility, there are drug industry trends that moderate hopes, such as the preference to avoid competition through developing first-in-class drugs (as well as drugs that treat rare diseases where there is little to no competition). A first-in-class drug refers to a drug whose approval by the Food and Drug Administration (FDA) effectively creates a new therapeutic category. The FDA’s Center for Drug Evaluation and Research noted that 54 percent of new drug approvals in 2022 were for first-in-class drugs.²⁷ As analysts at venture capital firm Andreessen Horowitz have noted, “Typically, the first drug to get across the finish line, wins: first-in-class drugs become the standard of care, dominate market share, and tend to maintain that advantage even after new entrants arrive.”²⁸ The financial rewards²⁹ of being first-in-class may explain a lack of competitors (“follow-ons”) within new therapeutic drug classes. For example, a study of new FDA-approved drugs for use within established therapeutic drug classes from 1986 to 2018 found 64 percent of first-in-class drugs did not have follow-on entries from competitors. Given the rate of failure in new drug development,³⁰ it is unsurprising that pharmaceutical companies exhibit an inclination to pursue low-competition opportunities where the financial rewards of success are assumed to be best.

These factors—along with lingering issues related to health care price transparency and the opaque pricing negotiations (i.e. rebates, discounts, spread pricing, artificial list prices)³¹ of pharmacy benefit managers—have substantially impaired price competition within the American health care market.³² Consequently, it is questionable whether AI productivity gains at medical facilities and drug manufacturers will exert downward pressure on consumer-facing medical costs. With respect to administrative savings, they are more probable to be passed on to consumers from payers (i.e. government and private insurers) when they use AI to detect fraud, waste, and abuse within medical bills.³³ Such an application of AI might save billions through more thorough claim policing, especially when one considers that a recent case of Medicare fraud involving catheters alone accounted for over $2 billion in fraudulent spending.³⁴ Back-office savings from fraud, waste, and abuse analysis will not lower provider charges on the consumer level but can have a positive effect on health plan premiums.

In summary, there are significant obstacles between AI administrative savings and consumer savings. However, the general pessimism on these savings reaching consumers does not extinguish AI’s possibilities for reducing health care spending. Rather, it focuses attention beyond productivity gains within health care administration.

Quality Improvements
As was the case for AI’s administrative productivity, its quality improvements in health care have attracted high hopes for lower costs. The improvements themselves can be astonishing, such as where AI demonstrates higher accuracy than human specialists in cancer prediction.³⁵ For example, the “Google DeepMind algorithm” outperformed radiologists in breast cancer detection with respect to fewer missed cancers and fewer false positives.³⁶ Economically, fewer false positives mean fewer expenditures for unnecessary treatment and avoiding unnecessary patient anxiety. Fewer missed cancers (false negatives), on the other hand, may save money by treating a breast cancer immediately rather than at a later stage when treatment costs may be more expensive.³⁷ A similar quality improvement example regards skin cancer detection. According to researchers presenting at the European Academy of Dermatology and Venereology Congress 2023, a particular AI-based software had a 100 percent accuracy rate for detecting melanoma and a 99.5 percent accuracy for all skin cancers.³⁸ With respect to pre-cancerous lesions, its accuracy was 92.5 percent.³⁹ Perhaps equally impressive as its accuracy rate was the rapid improvement of the software’s performance. Testing of an earlier version of the software in 2021 observed a melanoma detection rate of 85.9 percent, an 83.8 percent detection rate for all skin cancers, and a 54.1 percent detection rate for precancerous lesions.⁴⁰

AI’s clinical applications are often in diagnostic work such as radiology, pathology, and dermatology.⁴¹ In fact, more than 80 percent of the 692 AI medical algorithms cleared by the FDA (through July 2023) pertained to medical imaging.⁴² Despite this, the deep learning and flexibility of AI’s neural networks have allowed the technology to improve patient outcomes across many other medical settings. In surgery, AI is being examined for use in guiding surgical procedures and providing real-time decision support during surgery itself.⁴³ Non-surgical procedures can equally benefit from AI inasmuch as the technology can absorb enormous amounts of data (including a patient’s genetic information) and analyze numerous intervention possibilities to personalize a treatment likely to produce the best outcome.⁴⁴ With respect to post-treatment, AI can perform patient monitoring activities far beyond human ability. Specifically, AI can review more variables (pulse, blood pressure, respiration rate, temperature, etc.) at greater frequency than a human and detect patient deterioration earlier than nurses can. Moreover, an AI system monitoring patients can also predict health issues (based on historical data trends) before they happen and alert clinicians to a patient’s risk.⁴⁵

While more instances of AI-based improvements can be supplied, the effect these improvements will have on medical costs is unclear. Positively, AI’s success in early disease detection may save individual consumers money on treatment, though early disease detection does not always produce better outcomes⁴⁶ or lower costs. Additionally, what saves money for the individual does not always save money for the population. This is because, at the population level, the costs of broad AI screening must be considered against the number of early disease detections achieved and the savings (or costs due to overdiagnoses) generated from those detections. This kind of scrutiny is not unique to AI. A ready analogy can be found in the age recommendation for colonoscopies. Colorectal cancer is both deadly and relatively common.⁴⁷ The screening procedure for colorectal cancer, a colonoscopy, is an expensive procedure with an average cost over $2,000.⁴⁸ Whether colonoscopies for the general population should be recommended at a younger age (e.g. 40 or 45 instead of 50) has been periodically debated, with researchers weighing the treatment savings and lives prolonged against the incremental costs—including from adverse outcomes—of more screenings. Ultimately, researchers recognize there is an age threshold before which the benefits of early detection fail to justify the additional screening expenses.⁴⁹

While proponents of AI in health care are hopeful that improved patient outcomes will lower health care costs, the extent of savings will be inconsistent from one AI application to the next and together they may not be transformative for the market.

Autonomous Care (i.e. Self-Service)
Autonomous care—the delivery of a medical service via a self-service system a consumer uses without clinician assistance—is not unique to AI. For years, pharmacies have offered simplistic blood pressure machines for unassisted consumer use, with some systems having additional functionality such as body mass index calculation or vision evaluation. AI’s differentiation is the sophistication of its software and the additional services it enables. Compared to productivity gains and quality improvements, gains from autonomous AI medical services may have the best long-term prospects for consumer savings. Factors contributing to this perspective include the nation’s high physician expense and the lack of health care system competition. A landmark 2018 study concluded that the prices of labor and goods was the leading reason for high U.S. medical costs as compared to other high-income countries.⁵⁰ Medscape’s 2023 Physician Compensation Report found the average salary was $265,000 for a primary care doctor and $382,000 for specialists.⁵¹ Previous studies have found American physicians earn much more than doctors in European nations,⁵² including one that found the average U.S. doctor salary over twice the average in Germany.⁵³ This cost burden consumers bear for higher-paid physicians is inflated further by consolidation among health systems in the U.S. Between 1998 and 2021, 1,887 hospital mergers were announced.⁵⁴ A 2023 study from Harvard and the National Bureau of Economic Research found physician services provided by large consolidated health systems priced 12-26 percent more than at independent physician practices, while hospital services averaged 31 percent more in consolidated systems than independent hospitals.⁵⁵ Unfortunately, hospital consolidation has left many Americans lacking alternatives. A recent American Medical Association analysis found “low competition in virtually all hospital markets.”⁵⁶

AI has introduced a potentially disruptive element into this landscape. The combination of neural networks, natural language processing, machine learning, and generative AI has moved autonomous health care delivery further from science fiction and closer to reality. While AI is only at the dawn of self-service applications, the savings implications are substantial. Every time a medical service can be performed independently of a clinician, a sizable expense within health care delivery can be eliminated. Because medical devices aren’t licensed by individual states, there is also the possibility of AI-enabled health care devices competing for patients across the nation, giving alternatives to regions dominated by a consolidated hospital system and lowering prices through market competition.

There should also be savings from efficient scaling within self-service applications. Unlike humans, software can increase its delivery of a service with minor marginal costs. Marginal cost—the additional expense for software to supply a new service for a patient—is often quite low because the same programming is reused and computer infrastructure is relatively inexpensive to adapt for more simultaneous user sessions. Scaling human-generated service is quite expensive in comparison because the doubling of capacity often requires a doubling or more of employees.

Self-service medicine has additional implications for infrastructure-related costs. Physical medical facilities are extremely expensive to build and maintain, and these costs contribute to the nation’s high medical expenses. While self-service AI tools will not be able to reproduce some of the functions delivered through a hospital or doctor’s office, it is realistic that a subset of their activities can be performed through internet-based AI devices. In the case of a cell phone medical app, peripheral devices could be attached to capture data (e.g. pulse, oxygen level, and blood pressure) the microphone and camera cannot acquire. Dislocating care from a medical facility cannot only relieve the expense facilities add to medical care but can introduce another source of savings. Under current rules, Medicare often pays more for the same medical service if that service is delivered in a hospital (as opposed to a physician’s office).⁵⁷ This site-specific price adjustment may double the service cost. The practice has incentivized hospitals to acquire private medical practices and rechristen them hospital outpatient facilities even when they are not located on a hospital campus.⁵⁸ Self-service medical devices can offer an alternative to such inflated charges.

Some initial ventures have already been launched for self-service health care facilitated by AI, including instances where a clinician may be available but optional. Primary care company Forward has released automated CarePods using AI for screenings and diagnosis.⁵⁹ Marketed as the “first AI doctor’s office,”⁶⁰ these single-person compartments leverage self-service AI tools for most of their services, reserving doctor assistance for about 5 percent of its care. While this venture has a very limited range of medical care, Forward has announced that it will be expanding CarePod sensors and services in the future.⁶¹

Complementing this effort on AI primary care, virtual medical assistants diagnose illness and provide feedback based on their analyses. One estimate of the medical virtual assistant market expects revenues to reach $2.2 billion by 2030.⁶² Symptom checker and health management app Ada Health already has 13 million users and has performed 30 million symptom assessments.⁶³ Buoy Health, a medical AI platform developed by a Harvard Medical School team, markets over three dozen⁶⁴ care solutions addressing comprehensive physical, mental, and specialized health needs. The system elicits a patient’s symptoms and responds with “personalized information on possible causes and treatments for your illness or health problems, and proactive next steps to seek care.”⁶⁵

An AI medical assistant typically directs patients to clinicians for matters of diagnosis confirmation and treatment. The growing capabilities of AI, however, will force the issue of autonomous medical care, as well as its proper limits, into public debate. This debate, and the resulting regulations, will have profound implications for whether self-service AI lives up to its savings potential or becomes one more cost in an unsustainably expensive U.S. medical system.

Closely related to this issue will be the matter of legal liability, since this too influences health care costs through malpractice insurance premiums and legal judgments. As is the case for human clinicians, there will be occasions when the medical decisions and advice of autonomous AI will result in patient harm since much of medical advice and decisions is based on assessments of probabilistic diagnoses and outcomes. In the absence of clinician assistance or referral, the company providing the medical care will face liability if error, negligence, or improper omission is established. Such a determination might require greater transparency on the part of developers when there is a need to examine algorithms and underlying datasets.

Where autonomous AI complicates the traditional malpractice investigation is the issue of standard of care. “The standard of care,” as noted by health care risk management expert Donna Vanderpool, “is a legal term, not a medical term. Basically, it refers to the degree of care a prudent and reasonable person would exercise under the circumstances.”⁶⁶ Adherence to a standard of care may afford clinicians with a degree of malpractice protection⁶⁷ while negligence exposes the medical provider to liability actions.⁶⁸ Autonomous AI presents difficulties in not being a person and in having the capacity to exceed human decision making by virtue of its breadth of data assimilation and statistical calculations. In a scenario where an AI system (acting within the parameters of law and regulation) makes a medical recommendation that results in unintended harm, policymakers need to clarify the liability rules if the recommendation did not reflect a clinician’s standard of care but the system itself could demonstrate superior information and reasoning.

While there are dimensions of existing malpractice law that can be readily applied to autonomous AI, further evolution of this framework will be needed. Existing clinical expectations around diagnosis, treatment, and image analysis have been developed for a human model of medical practice. To maximize the health care benefits of AI while maintaining standards for patient safety, lawmakers and regulators will need to update malpractice rules to treat fairly the realities of nonhuman medical care. This would include, but not be limited to, the type of patient risk disclosures needed for autonomous AI since it is very important to avoid “strict liability” rules where a system or product is found liable for personal injuries, regardless of steps taken to avoid harm or intention.

AI has the potential for numerous quality and cost improvements in health care, but the technology’s potential within autonomous self-service settings may have the best long-term prospects for meaningful savings rather than producing incremental expenses. This hypothesis, of course, rests on many assumptions, the most important of which is the continued proliferation of AI tools for autonomous medical self-service. Beyond this are more practical matters regarding which conditions autonomous AI medicine requires to flourish. Two of the most important of these concern regulation and intellectual property.

Regulation of Health Care AI
The regulation of AI is an evolving landscape both inside and outside⁶⁹ of health care. Currently, there is not a comprehensive regulatory framework for medical AI, though there are state laws pertaining to the subject as well as ongoing work by the FDA.⁷⁰ One of the challenges facing regulators in their oversight is a deficit of internal programming expertise as well as the mathematical complexity behind the different varieties of AI technology. This is further complicated by transparency issues on the part of AI developers. Micky Tripathi, national coordinator for health information technology at the U.S. Department of Health and Human Services, has commented on government approval of AI technologies stating, “Right now, there’s a resistance to some of these tools because of the black box nature of them.”⁷¹ In fact, some companies are reluctant to share the specific data sets from which an AI algorithm learned because of the potential for competitors to reverse engineer AI algorithms by means of the data to be reverse engineered by competitors.⁷²

A similar situation had emerged for regulators with the rise of quantitative (or quant) funds.⁷³ Quant funds combine intricate multivariate math models with algorithmic securities trading. As a consequence, U.S. Securities and Exchange Commission regulators were initially ill-equipped to deal with these funds given that legal expertise was the governing skill set among them, not higher-level math or software programming. Further, the quants themselves were reluctant to expose their algorithmic secrets.

One of the dangers attending knowledge asymmetries between technology developers and regulatory agencies is that resulting industry rules produced by the agencies can be excessively cautious, impeding innovation and potentially robbing consumers of benefits that might have otherwise made it to market. Therefore, sound regulatory work on AI medical technology requires staffing adjustments within the FDA so the agency has solid internal AI expertise. Given the advances within AI, it is conceivable that AI itself could assist the FDA’s own AI approval processes through administrative automation as well as improvements in review and testing.

If AI expertise is inadequately represented within the FDA, there are several unattractive scenarios for regulation. First, final rules on AI issued by the agency may be made based on the most emotionally compelling and sensationalistic claims submitted during regulation comment periods, regardless of their inherent merit. Alternatively, the FDA may cede some portion of regulatory activity to another body. If this leads to the creation of an additional federal agency for AI compliance, there is the prospect of duplicative oversight costs (for both regulator and regulatee) resulting from overlapping agency efforts.⁷⁴ If, alternatively, a deficit of internal expertise leads the FDA to employ a third party in the vetting of AI technology then different challenges present themselves. For example, many institutions qualified to evaluate medical AI are themselves AI developers, thus creating a potential conflict of interest.⁷⁵ In order to avoid a situation where an AI developer determine “how difficult and costly it will be for competitors to measure up,”⁷⁶ the FDA should avoid consolidating the power of a small group of health systems and academic medical centers by giving them primary responsibility for health care AI validation.⁷⁷ A 2024 article published on STAT, “AI assurance labs intended to test health care technology have an equity problem,” has argued for a more distributed model of governance equipping health care providers with the resources and support to perform localized validations of AI functions.⁷⁸

Regardless of the entity validating AI, there are aspects of the technology that present unique challenges. Instead of evaluating a fixed foundation of source code and data, AI is evolutionary. Machine learning technology, for example, may add data to its system over time and iteratively refine its predictions. Generative AI is even more complex because it can create novel insights over time and produce outputs that have original content outside the software’s initial data and possibly beyond its foreseen outputs. This capacity, while delivering high value, also reduces the system’s predictability to regulators. Unlike the case for many traditional software systems, all probable outputs from a generative AI system may not be able to be anticipated from a Quality Assurance test scheme. However, as a software system, AI does have the possibility of reporting data back to a regulatory agency in automated fashion to confirm its proper ongoing operation. Such monitoring would be analogous to reporting obligations the FDA requires, post-approval, for certain elevated risk medications.⁷⁹

Another regulatory challenge of AI is the possibility of the system developers not fully understanding why an AI system is as effective as it is. ANN may have an enormous number of nodes as well as training data that produced its outcome. This “black box” problem is not unique to AI technology, though. It is widely accepted that there are FDA-approved medications where the drugs’ operation is not understood.⁸⁰ For the FDA, the central concerns are efficacy and safety, and these principles can apply equally to the approval of AI systems. Approval, though, has its own set of complications.

As of mid-2024, FDA regulators do not have a single clearance pathway for the approval of medical devices using AI,⁸¹ but rather several tracks (premarket clearance (510(k)), De Novo classification, and premarket approval). The FDA may also “may also review and clear modifications to medical devices, including software as a medical device, depending on the significance or risk posed to patients of that modification.”⁸² There are three classifications of risk, with the level of risk affecting the review process.⁸³

Against this background, the White House issued an executive order in late 2023 concerning AI to promote, among other objectives, “rigorous regulation”⁸⁴ of the technology. The order has many laudable aims on subjects such as privacy and consumer safety but often lacks detail regarding the kinds of measures needed to achieve them. Instead, the order articulates multiple goals without their prioritization or thought regarding possible conflicts among these goals. Additionally, the order seeks a determination of the “appropriate human oversight of the application of AI-generated output” without clarifying the principles on which that determination should be made. Absent from the order—and the HHS grantmaking it encourages for AI development—is a directive to identify ways AI can reduce health care costs.

There is a need for future regulation on medical AI’s supervision where the combination of AI system accuracy and safety concerns (such as hallucinations⁸⁵) necessitate clinician oversight. In order for AI to flourish, however, it is crucial that clinical assistance not be made a regulatory requirement in instances where an AI system can empirically demonstrate three criteria:

• Accuracy levels (or patient outcomes depending on the nature of the application) equal to or exceeding the average rate for clinicians performing the same function
• No amplification of health risks as compared to when the same function is performed by a clinician
• Output communications that are comprehensible and actionable for the patient without the additional explanation from a physician

Given the continual evolution of AI software systems, it is also advisable that regulators provide an economical pathway for innovators to re-apply for FDA approval on their devices where the functionality remains the same, but system autonomy increases over time. On this front, the FDA could leverage work already performed by the U.S. Department of Transportation for self-driving vehicles with differing levels of system autonomy (e.g. driver-assistance vs self-driving).⁸⁶

Through attention to real-world technology instances, these guidelines reflect an innovation-sensitive approach to regulating self-service medical AI, departing from the Europe Union’s Artificial Intelligence Act⁸⁷ because these guidelines are more flexible and better able to accommodate breakthrough advances while recognizing important differences in risk dictated by medical context (as opposed to political sensibilities). The direction of the Artificial Intelligence Act, in contrast, follows more of a politicized top-down regulatory preference that, even at this nascent period of the industry, creates “substantial obligations” for European AI developers that may put them at a competitive disadvantage internationally.⁸⁸

These regulatory recommendations intend to protect self-service AI development from rules impairing consumer cost-savings without materially benefitting those consumers with enhanced safety or quality. Nevertheless, inasmuch as savings inevitably involve revenue declines within the health care ecosystem, these principles will attract strong criticism from industry groups whose finances are most threatened by self-service AI. Policymakers should evaluate such lobbying efforts in light of the financial interests of the lobbying firms as well as the nation’s health care spending trajectory and the obvious limits on its sustainability.

AI and Intellectual Property
Intellectual property (IP) is also crucially important for AI’s success, and rules around IP can promote or constrain the rate of invention within the field. Positively, IP protections preserve the incentives for research and development investments. Negatively, these same protections, if misapplied, can inhibit innovation as well as encourage the use of AI patents as legal weapons to bleed funds away from AI start-ups rather than protect the rights of new products.⁸⁹

During the dot-com period, start-up technology companies faced a surge of infringement lawsuit threats from so-called patent trolls. Making matters worse is that 60 percent of the companies targeted by trolls were small to mid-sized.⁹⁰ For small companies, there was the likelihood of limited resources for legal defense and, therefore, an incentive to settle even if the suit had low merit. While patent litigation was not unique to the dot-com industry, complaints arose regarding the quality of many patents that had been issued for software and the same issues now shadow the AI health care market.

Because of the Patent and Trademark Office’s failures to properly review existing software use and convention, some patents were granted for concepts that were vague or inappropriately broad.⁹¹ The U.S. Code stipulates that “Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor.”⁹² The “process” dimension of this statement is what typically applies to software patents (including the AI variety) and may include methods of doing business as facilitated by software. Processes, unfortunately, may not be unique inventions so much as intuitive extensions of a programming language’s capability or the mathematical models that inform programming’s implementation. When this context is not properly taken into account, bad patents are granted, and unproductive infringement lawsuits abound.

A famous example is Amazon’s “single-click purchase” patent to expedite ecommerce transaction. Critics objected to the patent on the basis that underlying processes entailed in a single-click purchase were obvious extensions of existing software functionality and ecommerce conventions. The fundamental legal criteria for a patent, critics complained, are novelty (i.e. originality) and non-obviousness. These criteria are meant to protect the public not only from patents claiming invention of pre-existing items, but from logical extensions of those items. Imagine, for the purpose of illustration, if someone invented a new vacuum technology that was conceived to clean wall-to-wall carpet. The same invention could not be patented by a different party for the purpose vacuuming area rugs. Using the wall-to-wall carpet vacuum for area carpets would be an obvious application of the existing technology.

A study of patent validity disputes handled by the U.S. district courts and the Patent Trial and Appeal Board found that, when challenged,⁹³ patents were found invalid in approximately 40 to 45 percent of the cases under examination⁹⁴ (the district court statistics were based on all patent complaint rulings between 2008 and 2009, which accounted for nearly 1,000 cases).⁹⁵ The invalidation frequency, however, is not encouraging for AI health care start-ups whose capacity for financing patent disputes is constrained by limited funding and sales that have not reached an inflection point.

All the same problems that apply to software patents equally apply to AI patents as a subspecies of software. There is a risk for AI patents being issued on previously developed mathematical and logic systems that are now claimed to be novel by virtue of the specific datasets applied to them and the outputs they produce. Equally unhelpful are patents where the claims of novelty are allowed to be unduly generous. Critics have pointed to Google’s Dropout patent⁹⁶ and Microsoft’s Active Machine Learning⁹⁷ as examples of overly broad patents that might negatively affect progress in the AI market.

Another IP issue that can affect the progress of AI pertains to inventions created with the assistance of a third-party AI system. Unlike many software systems that perform predetermined operations (such as a calculator or a statistical analysis system), AI systems have the possibility to contribute material innovations unanticipated by the end user. Specifically, what happens if an inventor uses a third-party AI system and generates something patentable but the patent’s novel elements were generated by the AI system without the direction or intent of the inventor? Does the inventor have sole ownership of the invention, is ownership shared between the inventor and the AI system, or does the AI system alone own the invention? The way the government approaches these questions will have profound implications for the AI market, though the same could be said for the government’s treatment of AI patents in general. Ideally the government will:

• Increase internal AI expertise within the Patent and Trademark Office;
• Allocate appropriate resources to accommodate extensive prior invention searches needed for AI patent claims;
• Promote a mild skepticism around process patents without disallowing them altogether, favoring trade secrets as the preferred method of protection when algorithm claims are heavily indebted to existing theory and practice in the AI field;
• Protect AI inventions that are materially novel and not logical extrapolations from foundational math and theory in the public domain; and
• Incentivize self-service medical AI with high-cost savings potential with a faster patent review and approval process

AI is a multi-dimensional reality already transforming American health care and, thankfully, it has the potential to effect significant medical savings at a time when they are desperately needed. Additionally, the benefits of lower-cost AI health care would extend beyond our borders to nations where clinicians are scarce and poverty abundant. Realizing these savings at a meaningful scale, however, will require regulatory flexibility and IP conditions that promote free market innovation alongside the obligatory work to determine which AI approaches are truly effective. The recommendations contained in this paper fundamentally seek to:

• promote regulation from personnel who have an adequate technical grasp of AI’s potential and risks,
• avoid models of supplemental AI vetting that consolidate industry influence among a few major AI developers and produce unnecessary conflicts of interest,
• shield autonomous AI from requirements that will drive up health care costs without improving patient safety,
• advocate regulatory adjustments that respond to issues that emerge from real-world applications of AI in medicine rather than a massive defensive expansion of the existing regulatory framework, and
• support IP protections for true innovations and reduce AI companies’ exposure to patent trolls

The above recommendations, if implemented, would create a supportive environment conducive for the development of autonomous AI health care solutions. This environment would be further enhanced by complementary efforts establishing a model liability framework for states⁹⁸ in their efforts to address instances when patient harm results from medical AI use that is “proper and on-label.”⁹⁹ The liability model would ideally incorporate input from regulators, AI developers, clinicians, and malpractice legal experts. Such an endeavor would have multiple benefits, the first and foremost being improved consumer safety inasmuch as a liability framework communicates to the AI industry expectations around the technologies’ function, as well as accountability for failures to satisfy these expectations. Additionally, such a model would help to reduce uncertainty around AI liability and, in turn, further encourage investment in the field as well as market adoption.

The potential for AI to contain medical costs meaningfully is realistic but will remain hypothetical without a supportive environment. Should the government and private sector approach AI savings as an afterthought, rather than an explicit goal, new regulation and IP practices will reflect this neglect and unintentionally perpetuate the high health care costs that have burdened the nation for decades. Given the unsustainability of these costs, this would be a tragedy—and worse, an avoidable one.