The fiscal frontier

Introduction

Human history is marked by repeated waves of technological innovation and progress, with each wave reshaping civilization in its wake. From the mastery of electricity and the invention of the wheel to the printing press and the steam engine, each era-defining technological advancement has expanded the horizons of human capability. Not only does each of these waves add to society’s toolkit but they also profoundly influence culture, society, and individual lives. Now, in the 21st century, we sit at the crest of another wave, marked by advancements in artificial intelligence (AI).

In recent years, AI has transformed from a niche tool to a mainstream application. This transformation has been hailed by techno-optimists who envision a future in which AI not only enhances productivity and economic efficiency but also solves complex societal problems. With each algorithmic breakthrough and successful application, the promise of AI reinforces the hope that we might overcome some of the most daunting challenges facing humanity. This sentiment is widely shared by many who are witnessing the formidable strides made by AI. Our study is motivated by the fact that, amidst the latest wave of techno-optimism fueled by AI, virtually no attention is being paid to how AI might shape the fiscal outlook.

The impact of AI on federal spending and revenues is highly uncertain, given the nascent evolution of AI and the unpredictable impact of AI on economic activity. From the outset, we acknowledge this uncertainty and aim to model a series of representative shocks that provide an illustration of AI’s potential to impact federal old-age entitlement spending. While the nature and magnitude of these shocks vary, several implicit assumptions frame the collection of shocks. To start, we implicitly assume that AI’s economic impact will be at least moderate—if not more substantial—and rising over time as the technology is adopted more widely and its capabilities continue to develop. In addition, we assume—in line with virtually every other major technological shock—that the net impact of widespread AI adoption will be productivity enhancing. However, we also note that the evidence to date suggests that AI may impact the fiscal outlook through substantially different channels than in prior technological revolutions. Specifically, while AI may ultimately have a profound impact on productivity, AI has already shown the potential for dramatically changing health care delivery, effectiveness, and cost—which could translate into changes in mortality, morbidity, price of care, and care utilization. Given that such changes could have profound impacts on Social Security and public health program outlays, policymakers would benefit from proactively integrating AI’s capabilities into fiscal planning and projections.

In this paper, we model the potential impact of AI on Social Security outlays, Medicare outlays, and the subsequent change in net interest payments. These components of the federal budget comprise the vast bulk of entitlement outlays, with an increasing share over time.¹ However, there is potential for AI to impact various other health-related fiscal elements including payroll taxes to fund major entitlements, defense spending, other health programs such as Medicaid, premium support, and the Children’s Health Insurance Program. Analyzing the impact of AI on these fiscal elements presents an opportunity for future research.

Recent developments and medical applications

This research is especially timely as the field of AI has evolved rapidly over the last decade. Particularly significant advances in AI technology include the 2017 development of the Transformer model.² The latter breakthrough provided the foundation for modern Large AI Models (LAMs) such as OpenAI’s Generative Pre-Trained Transformer 4 (GPT-4).

LAMs (or, colloquially, “large language models” (LLMs)) based on Transformer have continuously progressed in size, complexity, and capability. In 2023, OpenAI released its GPT-4 model which is rumored to have 1.76 trillion parameters,³ making it hundreds of times larger than the 340 million parameters of Google’s Bidirectional Encoder Representations from Transformers (BERT) LAM model released in 2018.⁴ While model size is imperfectly correlated to capability, larger AI models tend to be more capable than smaller ones. AI models’ number of “modalities” has also expanded; while earlier LAMs were only trained on language data (LLMs) or image data (large vision models; LVMs), LAMs can now be trained on two or more types of data (large multi-modal models; LMMs).⁵

On the less technical front, AI products like ChatGPT have seized the public’s imagination and taken AI mainstream as a topic of both excitement and concern. This widespread discussion has intensified existing academic interest in AI’s applications in a variety of fields, including law,⁶ finance,⁷ economics,⁸ and most significantly for our purposes, medicine.

AI products have already impacted health care. For example, the Alphafold 2 program, based on a Transformer model, has revolutionized protein structure prediction (“protein folding”) since its release in 2020. Before Alphafold, decades of experimentation had left researchers with a complete structural understanding of only about 17% of the protein residues in the human body. In contrast, Alphafold was able to quickly develop a confident structural prediction for 58% of proteins.⁹ Alphafold offers researchers unprecedented insight into the building blocks of human life, which could accelerate the speed of medical research and drug discovery.¹⁰

LAMs are also increasingly being leveraged towards improved medical diagnoses, an area in which they show incredible promise. Google’s Articulate Medical Intelligence Explorer (AMIE), an LLM-based system, performed better than human clinicians when evaluating over 300 challenging diagnostic cases drawn from the New England Journal of Medicine. This LLM listed the correct diagnosis among its top-10 predictions 59.1% of the time, significantly outscoring human clinicians’ 33.6% top-10 accuracy.¹¹ AMIE also excels at interacting with patients; patient actors scored text-based consultations with AMIE as being significantly better than those provided by primary care physicians across the vast majority of evaluation axes, including empathy and sensitivity.¹²

Beyond protein folding and diagnosis, current LAMs can perform better than previously state-of-the-art methods in categories of health care tasks such as medical imaging, medical informatics, medical education, and public health.¹³ For their part, Generative Adversarial Networks (GANs) are used for synthetic medical image generation, augmenting otherwise limited sets of training data and thereby improving the performance of other neural networks.¹⁴ Furthermore, GAN-synthesized data can help anonymize health care data, ameliorating AI-related privacy concerns.¹⁵ Although the future capabilities of LAMs are difficult to forecast, the growth in capabilities of frontier AI models in the past two years makes clear the direction of travel. As the capabilities of AI models expand so do their implications for society through a variety of channels including health care and the federal budget.

Potential impact of AI on health care and longevity

One of AI’s largest potential impacts will be in accelerating the efficacy of preventive medicine. The use of AI in preventive care and early detection of diseases could lead to a reduction in morbidity rates, contributing to a healthier population that requires less medical intervention over time. AI algorithms have shown remarkable success in diagnosing diseases from images (such as radiology scans) and predicting patient outcomes based on historical health data. AI’s ability to improve diagnostic accuracy can not only improve patient outcomes but also reduce wasteful spending on inappropriate treatments. These tools can assist clinicians in detecting conditions earlier and with greater precision, potentially enabling earlier interventions that extend longevity.

Additionally, AI shows significant promise in optimizing treatment plans. By rapidly analyzing massive amounts of data from a wide range of sources, AI can help identify the most effective and cost-efficient individualized treatment plans for patients. This includes determining which medications are likely to be most effective based on a patient’s unique profile, thus avoiding costly and ineffective treatments.

Similarly, AI applications in monitoring patient health and predicting flare-ups of chronic conditions can lead to better management of chronic diseases and reduce the need for expensive hospitalizations and treatments. Wearable devices and mobile health apps, powered by AI, enable real-time monitoring and can alert patients and health care providers to potential health issues before they require more serious intervention. The aim is to integrate data from wearable devices, patient records, and call transcripts into unified systems that act as “co-pilots” for health care providers, keeping them informed about their patients’ conditions in real-time. By reducing the need for in-person health care, this can alleviate capacity constraints across the entire health care system.

These advances in AI have the potential to dramatically alter the scope of federal spending on old-age entitlement programs, which can subsequently alter the fiscal trajectory. From a more optimistic perspective, existing AI systems may lower expenditures on all health spending, including Medicare, with cost reductions occurring through several channels—with personalized medicine being a prominent example. AI enables the analysis of vast amounts of data, including genetic information, lifestyle factors, and environmental exposures, to tailor treatments to individual patients. This personalized approach may significantly improve outcomes by targeting therapies that are most likely to be effective for a particular patient, reducing the trial-and-error approach that characterizes much of current medical practice. AI may further reduce health care costs by avoiding unnecessary treatments and hospital admissions, thus lowering the financial burden on the public health care system. AI can also help identify and prevent fraudulent Medicare claims, saving costs for the program.

Beyond direct patient care, AI may enhance health care quality by improving hospital and clinic operations. From optimizing appointment scheduling to managing patient flow and predicting peak times for different services, AI can help reduce wait times and improve the patient experience. Similarly, AI could potentially automate administrative tasks such as data entry, appointment scheduling, and even preliminary data analysis for diagnostic purposes. By reducing the burden of repetitive tasks on health care professionals, AI allows them to focus more on patient care, thereby increasing the efficiency of health care delivery and reducing labor costs.

The rapid investments in AI-based health care by the world’s leading technology companies signify the direction of the industry, and the ensuring competition that could significantly enhance medical care and operational efficiency. For example, Google, a leader in leveraging AI for health advancements, is at the forefront with projects like Med-Gemini. This health-specific LLM focuses on providing accurate responses to medical inquiries and facilitating the summarization of information during pivotal moments such as patient handoffs and staff shift changes. Microsoft’s recent string of strategic acquisitions further exemplifies the tech industry’s drive towards integrating AI into health care. In 2021, Microsoft acquired Nuance, software designed to assist health care professionals with administrative tasks such as generating clinical notes and managing electronic health records. Similarly, Amazon’s collaboration with Anthropic aims to introduce a version of Claude AI—Anthropic’s LLM-powered chatbot—to augment health care services.

The race to harness AI for health care advances isn’t just happening among American tech giants—it’s also a point of competition and collaboration among nations. Chinese technology companies are also venturing into this arena. A 2022 McKinsey report highlighted AI’s potential to revolutionize health care in China, projecting that AI’s integration into diagnostic predictions and clinical decision support could generate approximately $5 billion in economic value.¹⁶ This underscores a global recognition of AI’s transformative potential in health care.

Importantly, these positive developments are not guaranteed to translate into fiscal improvements. Under a scenario where AI leads to reductions in mortality rates but not cost savings in per beneficiary health costs, the fiscal outlook for entitlement spending could deteriorate as an expanded old-age population implies higher fiscal outlays. Moreover, improved efficacy of health care delivery could potentially increase health care utilization by driving up demand for such services—although such a scenario would likely be accompanied by price reductions due to improved health care productivity. In short, the AI revolution in health care could counterintuitively increase both per capita spending on entitlements and the population of beneficiaries receiving these services.

This paper is based on exploratory analysis which attempts to place structure on forecasts of AI’s potential impact on medicine and fiscal budgets. Our analyses should be regarded as an initial attempt to scope the potential magnitude of an AI shock on the long-term budget outlook for entitlement spending—which we consider to be a 20-year time frame. Our initial estimates suggest that the nature of the shock is critical, as the impact of the shock on annual budget deficits could range from an increase of roughly 1.6% of GDP to a decrease of around 0.8% of GDP by 2044, with the latter reducing annual budget deficits in 2044 by roughly one ninth. In the next section we review the literature around the impact of AI on various aspects of health care and longevity. Section 3 presents a theoretical model. In Section 4 we lay out our methodology, and our results are presented in Section 5. Section 6 briefly concludes.

Download the final report here