Belmont Report as a Framework for Patient Safety and Artificial Intelligence (AI)

Beneficence, Respect for Persons and Justice in Patient Safety and AI

Patient Safety and Artificial Intelligence have great potential but with any great strides, great potential comes with great responsibility.

The current artificial intelligence landscape as a kind of “wild wild west”. In terms of generative AI, the biggest players in the U.S. are well known: Google, OpenAI, and Anthropic PBC. In terms of AI more broadly, the biggest players also include Facebook, Amazon, Microsoft, IBM, and many others. Additionally, companies such as Baidu, Alibaba, and Tencent (BAT), and DeepSeek are some of China’s most prominent companies. Each player in this space is seeking not only to capture monthly subscriptions from individual and corporate users, but are also seeking to establish themselves as the leader in AI.

Being a leader in this space requires vast sums of capital, land, and access to a supply chain that can source, manufacture, and deliver tens of thousands of graphics cards, servers, and more. The electrical and cooling demands of these AI systems is on a scale that is equally vast. With regard to electricity alone, GoldmanSachs estimates these systems will require 24GW of power by 2030, whereas others have predicted that, with an increase in AI-specific chips, the power requirements could exceed 300GW of power.^.

To put that into perspective, California has a total current power capacity of 86GW. With regard to worries about climate change, Eric Schmidt (one of the founders of Google), has said "We’re not going to hit the climate goals anyway because we’re not organized to do it The takeaway of this comment was that we should go all in on Artificial Intelligence

Patient safety is defined as the absence of preventable harm to a patient and minimization of the risk of harm associated with the health care process Every part of the care-giving process involves a certain degree of inherent risk. Since resolution WHA55.18 on “Quality of Care: Patient Safety” at the 55th World Health Assembly was proposed in 2002, there has been increasing attention paid to patient safety concerns and adverse events in health care settings

Despite the safety initiatives and investments made by federal and local governments, private agencies, and concerned institutions, studies continue to report unfavorable patient safety outcomes.

The Proposal is for the Patient Safety Community to use the Belmont Report in accordance with the principals of cybersecurity to create guidelines for the medical community and the AI community to work together to come with guiding principals of patient safety testing to prevent harm and assist the AI community in coming up a updated guidelines for tolerance testing to prevent harm and prevent medical errors for a healthy community.

The following is some evidence as a call to action from several sources before it is too late.

According to Patton (2025), Establishing an equivalent legal framework for AI is going to take tremendous buy-in from a variety of private and public actors in the United States. The model afforded by the Belmont Report is well suited to generate such buy-in. While this may seem like a daunting task given various polarizing issues at play in society today, the context that produced the Belmont Report was quite fractious itself. It is the position of this paper that a similarly styled approach to AI regulation can succeed in proactively limiting the harms of AI’s use (and abuse).

What is the Belmont Report

The Belmont Report, published in 1979 by the National Commission for the Protection of Human Subjects of Biomedical and Behavioral Research in the United States, outlines foundational ethical principles for conducting research involving human subjects. While it originated within the context of medical and behavioral research, its principles have become increasingly pertinent to the field of artificial intelligence (AI), especially concerning data derived from individuals.

The Belmont Report comes out of 3 important events, The Tuskegee Syphilis study, the Nuremberg Report after World War II and the Henry Beecher study of experiments on vulnerable populations (children and prisoners) and is used mainly for clinical trials research with a bunch of exemptions that can be utilized,.

The report establishes three fundamental ethical principles:

1. Respect for Persons: This principle acknowledges the inherent dignity and autonomy of individuals, requiring researchers to respect participants' decisions and protect those with diminished autonomy. In AI ethics, this translates to respecting user autonomy in interactions with AI systems, ensuring transparency, and safeguarding privacy and personal data.

2. Beneficence: Beneficence involves an obligation to prevent harm and promote the well-being of participants by maximizing potential benefits and minimizing possible risks. In the AI context, this principle guides the development of AI systems that are safe, secure, and designed to benefit users while actively preventing harm.

3. Justice: This principle pertains to the fair distribution of the benefits and burdens of research. It seeks to prevent the exploitation of vulnerable groups and ensure equitable access to the advantages derived from research. For AI, justice entails providing equitable access to AI technologies and ensuring that AI systems do not exacerbate existing societal inequalities or introduce new forms of bias and discrimination.

Principals of Cybersecurity

1. Confidentiality: The principle of confidentiality ensures that sensitive information is kept private and protected from unauthorized access. This includes personal information, financial data, and other confidential information that should only be accessible to authorized individuals.

2. Integrity: The principle of integrity ensures that information is accurate and trustworthy. This involves protecting information from unauthorized modification, ensuring that information is complete and accurate, and maintaining data consistency.

3. Availability: The principle of availability ensures that information and systems are available and accessible to authorized users when needed. This involves protecting against denial of service attacks, ensuring that systems are functioning properly, and implementing disaster recovery and business continuity plans.

4. Non-repudiation: The principle of non-repudiation ensures that an individual or entity cannot deny that they have sent or received a message. This involves using digital signatures and other authentication mechanisms to verify the authenticity

   of messages and transactions.

Intersection of Patient Safety and Artificial Intelligence

Artificial intelligence (AI) provides opportunities to identify the health risks of patients and thus influence patient safety outcomes.

According to Chowdry and Asan from the Stevens Insitute of Technology, who did a literature review of 53 cases over 10 years from 2009 - 2019 and dentified 53 eligible studies, which were summarized concerning their patient safety subcategories, the most frequently used AI, and reported performance metrics. Recognized safety subcategories were clinical alarms (n=9; mainly based on decision tree models), clinical reports (n=21; based on support vector machine models), and drug safety (n=23; mainly based on decision tree models).

Analysis of these 53 studies also identified two essential findings: (1) the lack of a standardized benchmark and (2) heterogeneity in AI reporting. there are 3 factors.

1. The integration of artificial intelligence (AI) into the health care system is not only changing dynamics such as the role of health care providers but is also creating new potential to improve patient safety outcomes [6] and the quality of care The term AI can be broadly defined as a computer program that is capable of making intelligent decisions

2. The operational definition of AI is the ability of a computer or health care device to analyze extensive health care data, reveal hidden knowledge, identify risks, and enhance communication In this regard, AI encompasses machine learning and natural language processing. Machine learning enables computers to utilize labeled (supervised learning) or unlabeled (unsupervised learning) data to identify latent information or make predictions about the data without explicit programming. Among different types of AI, machine learning and natural language processing specifically have societal impacts in the health care domain and are also frequently used in the health care field.

3. The third category within machine learning is known as reinforcement learning, in which an algorithm attempts to accomplish a task while learning from its successes and failures [9]. Machine learning also encompasses artificial neural networks or deep learning [13]. Natural language processing focuses on building a computer’s ability to understand human language and consecutively transform text to machine-readable structured data, which can then be analyzed by machine-learning techniques [14]. In the literature, the boundary defining natural language processing and machine learning is not clearly defined. However, as illustrated in studies in the field of health care have been using natural language processing in conjunction with machine-learning algorithms

AI has potential to assist clinicians in making better diagnoses [16-18], and has contributed to the fields of drug development [19-21], personalized medicine, and patient care monitoring [14,22-24]. AI has also been embedded in electronic health record (EHR) systems to identify, assess, and mitigate threats to patient safety [25]. However, with the deployment of AI in health care, several risks and challenges can emerge at an individual level (eg, awareness, education, trust), macrolevel (eg, regulation and policies, risk of injuries due to AI errors), and technical level (eg, usability, performance, data privacy and security).

The measure of AI accuracy does not necessarily indicate clinical efficiency [26]. Another common measure, the area under the receiver operating characteristic curve (AUROC), is also not necessarily the best metric for clinical applicability [27]. Such AI metrics might not be easily understood by clinicians or might not be clinically meaningful [28]. Moreover, AI models have been evaluated using a variety of parameters and report different measure(s) such as the F1 score, accuracy, and false-positive rate, which are indicative of different aspects of AI’s analytical performance. Understanding the functioning of complex AI requires technical knowledge that is not common among clinicians. Moreover, clinicians do not necessarily have the training to identify underlying glitches of the AI, such as data bias, overfitting, or other software errors that might result in misleading outcomes. Such flaws in AI can result in incorrect medication dosage and poor treatment

AHRQ Interview with Patrick Tighe in 2024

Ensuring patient safety in modern healthcare is a complex task, with numerous interrelated factors contributing to numerous potential harms. These factors, including disorganized data, overburdened clinicians, and complex clinical cases, create challenges that require sophisticated solutions. The integration of artificial intelligence (AI) into health information technology (IT) systems offers the promise that some of the challenges can be reduced or overcome. AI can analyze a vast amount of data from various sources, optimize workflows, and offer evidence-based recommendations to clinicians. While certain specialties have already found success by implementing AI, and the research continues to progress, widespread AI adoption in daily clinical practice is still on the horizon. Not only is there a lack of peer-reviewed prospective evidence of its effects on patient care and the clinician experience, integrating AI poses a number of ethical and technical challenges. Nonetheless, the potential of AI to enhance patient safety and improve healthcare is great.

The potential for enhancing patient safety across various specialties and settings is substantial. A scoping review in 2021 explored the impact of AI on eight main patient safety domains, suggesting that AI's influence would be most pronounced in domains where existing strategies have proven insufficient and where integration and analysis of new, unstructured data is crucial for accurate predictions. Such domains include adverse drug events, clinical decompensation, and diagnostic errors.

1. One of the most common potential applications of AI for patient safety is risk prediction—for example, predicting the likelihood that a patient will decompensate, have an adverse reaction to a medication.

2. AI-powered data models excel in this task, as they can process real-time data from various sources within the electronic health records (EHRs) and biometrics and dynamically adjust their predictions based on new data about patients.

3. A novel data source being explored is video taken in clinical environments.⁷ In this application, cameras or movement sensors gather data on what is happening in the healthcare setting and can alert staff when a patient falls or a critical checklist step is skipped, for example.

4. A final example of an application that may have a large impact on many clinicians’ day-to-day life is AI auto-charting. Instead of needing to complete the EHR as they perform a procedure, with the computer between them and the patient, the AI system can be listening along and completing the chart for them. , its adaptation to healthcare holds promise, especially with the continued improvement of large language models and ongoing exploration of how to implement these technologies in live clinical practice.

Potential Harms to Patient Safety

While the potential for improving patient safety is high, AI also comes with risks that must be carefully considered before and during implementation.

First and foremost, as with all models, it is important to ensure that the AI model goals are in alignment with specific patient safety goals, such as identifying patient decompensation. If the AI model is not precisely aligned with this patient safety goal, it may either miss critical signs of decompensation or generate false alarm

Springer Call Action

According to Patton “While AI is not a researcher and the public are not its subjects, viewing the ethical principles that we need to agree on as prima facie is useful in the context of AI as well. Given that there will be so many varied applications of AI to our lives, it will likely turn out that the weighting we have of certain values in one area will not map neatly into another. This will be due to the specific details of each application of AI and the ways in which such applications will impact our lives. This is completely consistent with a principles-based approach. The move from prima facie to all things considered requires the incorporation of all circumstantially relevant details. How such values will be weighted in, for example, the implementation of AI into medical diagnosing is likely to be very different than how those same values will be weighted in cases where AI is utilized to determine employability. In fact, it may turn out that for different applications of AI, there are just different values that we agree to. The kind of cases that will arise due to AI being utilized in medicine, for example, may simply fall under the current principles already at play in contemporary medical ethics. With regard to AI being utilized in decisions about employment, however, principles of medical ethics may not be applicable. AI is malleable with regard to its uses since the data that it is trained on can come from nearly any domain. Hence, having different groupings of relevant ethical principles given different domains of applications is perhaps inevitable. This may make something like a Belmont Report for AI a more complicated and multifaceted project. All the more reason to get started now.

In this paper, I have argued that the National Research Act and Belmont Report can serve as a model form which AI regulation can follow. What we need for effective AI regulation is both ethical consensus among varying parties and external enforcement. With regard to the former, there is no shortage of candidates on offer in the literature. With regard to the latter, however, either we trust the companies to regulate themselves or we must rely on legislation. I take that most are hesitant to trust these companies to self-police effectively, and so we must focus on legislation. Ethics and legislation thus comprise the two sides of the AI regulation coin. To that end, I propose that we should establish a National AI Act with a corresponding document expounding the relevant ethical principles at play.

What we need for effective AI regulation is both ethical consensus among varying parties and external enforcement. With regard to the former, there is no shortage of candidates on offer in the literature. With regard to the latter, however, either we trust the companies to regulate themselves or we must rely on legislation. I take that most are hesitant to trust these companies to self-police effectively, and so we must focus on legislation. Ethics and legislation thus comprise the two sides of the AI regulation coin. To that end, I propose that we should establish a National AI Act with a corresponding document expounding the relevant ethical principles at play.”

NIST Call to Action

“We looked at existing principles of human subjects research and explored how they could apply to AI,” said Kristen Greene, a NIST social scientist and one of the paper’s authors. “There’s no need to reinvent the wheel. We can apply an established paradigm to make sure we are being transparent with research participants, as their data may be used to train AI.”

The Belmont Report arose from an effort to respond to unethical research studies, such as the Tuskegee syphilis study, involving human subjects. In 1974, the U.S. created the National Commission for the Protection of Human Subjects of Biomedical and Behavioral Research, and it identified the basic ethical principles for protecting people in research studies. A U.S. federal regulation later codified these principles in 1991’s Common Rule, which requires that researchers get informed consent from research participants. Adopted by many federal departments and agencies, the Common Rule was revised in 2017 to take into account changes and developments in research.

There is a limitation to the Belmont Report and Common Rule, though: The regulations that require application of the Belmont Report’s principles apply only to government research. Industry, however, is not bound by them.

The NIST authors are suggesting that the concepts be applied more broadly to all research that includes human subjects. Databases used to train AI can hold information scraped from the web, but the people who are the source of this data may not have consented to its use — a violation of the “respect for persons” principle.

“For the private sector, it is a choice whether or not to adopt ethical review principles,” Greene said.

While the Belmont Report was largely concerned with inappropriate inclusion of certain individuals, the NIST authors mention that a major concern with AI research is inappropriate exclusion, which can create bias in a dataset against certain demographics. Past research has shown that face recognition algorithms trained primarily on one demographic will be less capable of distinguishing individuals in other demographics.

Applying the report’s three principles to AI research could be fairly straightforward, the authors suggest. Respect for persons would require subjects to provide informed consent for what happens to them and their data, while beneficence would imply that studies be designed to minimize risk to participants. Justice would require that subjects be selected fairly, with a mind to avoiding inappropriate exclusion.

Greene said the paper is best seen as a starting point for a discussion about AI and our data, one that will help companies and the people who use their products alike.

“We’re not advocating more government regulation. We’re advocating thoughtfulness,” she said. “We should do this because it’s the right thing to do.”

Conclusion

We really need to move fast on Patient Safety and Health Care Systems Issues. As in “To Err is Human”. Patient Safety has been a guiding principle for 25 years. The patient safety community is a team of professionals that can address this head on and come together as a community to get ahead of AI before it gets ahead of us I join the chorus in sounding the alarm before it is too late. AI is coming fast and there will be casualties. It is up to the patient safety community to come together to make sure that harm is prevented before it happens. We are all in this together.

Sources

Role of Artificial Intelligence in Patient Safety Outcomes: Systematic Literature Review - PMC

Rutgers AI lab: Belmont Report

Paper: K.K. Greene, M.F. Theofanos, C. Watson, A. Andrews and E. Barron. Avoiding Past Mistakes in Unethical Human Subjects Research: Moving from AI Principles to Practice. Computer. February 2024. DOI: 10.1109/MC.2023.3327653

Patton, K. We need a Belmont report for AI. AI & Soc (2025). https://doi.org/10.1007/s00146-025-02461-0

Tighe P, Mossburg S, Gale B. Artificial Intelligence and Patient Safety: Promise and Challenges. PSNet [internet]. Rockville (MD): Agency for Healthcare Research and Quality, US Department of Health and Human Services. 2024.

What Are The 4 Principles Of Cybersecurity? - Decoding Cybersecurity