Generative AI in medical practice: in-depth exploration of privacy and security challenges

Probably not a substantial claim: may not pass S3 (load-bearing).

Adversarial training uses carefully crafted adversarial examples to attack the training model to check for the integrity and robustness of outputs.

“The method uses carefully crafted adversarial examples to attack the training model to check for the integrity and robustness of outputs [57,79].”

• cites [57] sourceNot checked — no full text of the source
• cites [79] sourceDisputed

  The Contributions section (p. 2) describes adversarial training as transforming real endoscopy images into synthetic-like images for domain adaptation. Section 4 (p. 4) says the discriminator classifies images as real or synthetic and gives feedback to the transformer. That is a different method from using crafted adversarial examples to attack a model and test the robustness of its outputs. (LLM, conf=0.87)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Ghosheh et al suggest that the use of GANs to create synthetic electronic health records creates many privacy challenges, such as reidentification and membership attacks.

“Ghosheh et al [46] suggest that the use of GANs to create synthetic EHRs creates many privacy challenges (eg, reidentification and membership attacks).”

• cites [46] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

In research settings, generative AI techniques can hypothesize promising new directions by creatively combining concepts.

“In research settings, these generative AI techniques can hypothesize promising new directions by creatively combining concepts [12].”

• cites [12] sourcePartly supported

  The analogy-based design synthesis section (p. 13) proposes using generative models to create new designs guided by analogies. The Applications section (p. 9) also describes tools that combine ideas for design ideation. Neither passage says these techniques hypothesize promising new research directions; the paper's application is engineering design. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Hernandez et al discuss privacy concerns related to synthetic tabular data generation in health care.

“Hernandez et al [47] discuss privacy concerns related to synthetic tabular data generation in health care.”

• cites [47] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Various methods and evaluation metrics are used to assess the privacy dimension of synthetic tabular data generation approaches, including identity disclosure, attribute disclosure, distance to the closest record, membership attack, maximum real-to-synthetic similarity, differential privacy cost, and GANs.

“Various methods and evaluation metrics are used to assess the privacy dimension of the synthetic tabular data generation approaches. These methods include identity disclosure, attribute disclosure, distance to the closest record, membership attack, maximum real-to-synthetic similarity, differential privacy cost, and GANs.”

• cites [47] sourcePartly supported

  Section 6.2.3 (page unavailable) lists identity and attribute disclosure, distance to the closest record, membership attack, maximum real-to-synthetic similarity, and differential privacy cost as privacy-evaluation methods. Sections 6.2 and 6.2.3 treat GANs as approaches being evaluated, not as a privacy-evaluation method, so the final item in the claim is unsupported. (LLM, conf=0.89)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative AI promises to accelerate drug discovery by inventing optimized molecular candidates.

“Generative AI also promises to accelerate drug discovery by inventing optimized molecular candidates [11].”

• cites [11] sourceSupported

May combine two claims.

Two solutions have been proposed to address privacy leakage in models: adding a regularization term in a loss function to avoid overfitting and improve robustness, and adding acceptable noise into model parameters to hinder privacy inference attacks.

“To address the problem of privacy leakage in the models, two solutions have been proposed: (1) adding a regularization term in a loss function to avoid overfitting and improve robustness; for example, this method can be applied to defend against membership inference attacks, [37] and (2) adding acceptable noise into the model parameters to hinder privacy inference attacks.”

• cites [37] sourceSupported
• cites [38] sourceCould not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Challenges remain around privacy, information accuracy, and integration into provider workflows for virtual health assistants.

“However, challenges remain around privacy, information accuracy, and integration into provider workflows [25].”

• cites [25] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Examples of GAN-based defense models include generative adversarial privacy, privacy-preserving adversarial networks, compressive adversarial privacy, and reconstructive adversarial network.

“Examples of defense models include generative adversarial privacy [30], privacy-preserving adversarial networks [31], compressive adversarial privacy [32], and reconstructive adversarial network [33].”

• cites [30] sourceSupported
• cites [31] sourceSupported
• cites [32] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative AI has applications in engaging patients through natural conversation powered by large language models.

“Generative AI also has applications in engaging patients through natural conversation powered by LLMs [2].”

• cites [2] sourcePartly supported

  The Introduction describes ChatGPT as a generative language model capable of conversation and notes that conversational agents are available in medicine and to consumers. The Interview focuses on medical education and simulated patients; it does not describe using LLMs to engage real patients through conversation. No page numbers are available in our copy. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Federated learning is a distributed AI paradigm that offers privacy preservation in smart health care systems by allowing models to be trained without accessing the local data of participants.

“FL is a distributed AI paradigm that offers privacy preservation in smart health care systems by allowing models to be trained without accessing the local data of participants.”

• cites [49] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

The target of federated learning in health care AI applications is to preserve the privacy of sensitive patient information communicated between hospitals and end users, particularly through Internet of Medical Things devices.

“The target of FL in health care AI applications is to preserve the privacy of sensitive patient information communicated between hospitals and end users, particularly through Internet of Medical Things (IoMT) devices.”

• cites [49] sourcePartly supported

  The Introduction says FL protects privacy in smart-healthcare networks by training across IoMT networks without accessing local data, mitigating disclosure of sensitive patient information. The copy does not specifically describe FL’s target as protecting information communicated between hospitals and end users; that communication path is more specific than the cited passage supports. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

Deep-faked patient face images could violate patient privacy and lead to the leakage or exploitation of protected health information.

“Deep-faked patient face images could violate patient privacy and lead to the leakage or exploitation of PHI [75].”

• cites [75] sourcePartly supported

  The Introduction (p. 1) says membership inference against GANs used with health records or medical images may reveal a patient’s disease history. The Setup (p. 7) treats face images and patient records as separate datasets. The paper supports a risk of patient-data leakage, but does not address deep-faked patient faces specifically or exploitation of protected health information. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Implementing federated learning in Internet of Medical Things networks presents challenges, such as the need for robust federated learning for diffused health data sets, integration of federated learning with next-generation Internet of Medical Things networks, and use of blockchain for decentralized and secure data storage.

“However, implementing FL in IoMT networks presents challenges, such as the need for robust FL for diffused health data sets, the integration of FL with next-generation IoMT networks, and the use of blockchain for decentralized and secure data storage.”

• cites [49] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

In practice, generative AI systems in health care have been found to experience integrity threats, such as generating disinformation and misinformation, and making biased decisions.

“In practice, generative AI systems in health care have been found to experiencing integrity threats, such as generating disinformation and misinformation, and making biased decisions [81].”

• cites [81] sourcePartly supported

  The Results (Consistency of ChatGPT Answers to Guideline Recommendations; page unavailable) report that ChatGPT’s answers agreed with clinical guidelines in 15 of 25 cases, supporting concern about inaccurate medical information. The Discussion (Principal Findings; page unavailable) warns of a danger of perpetuating biases, but does not report observed biased decisions or establish that ChatGPT generated deliberate disinformation. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

Assembling genomic databases and chemical compound or protein structure databases for drug discovery is extremely challenging and could be error ridden because many genomic and protein databases lack necessary annotations, are inconsistent in formats, and may be poor in data quality.

“For example, assembling genomic databases and chemical compound or protein structure databases for drug discovery is extremely challenging and could be error ridden because many genomic and protein databases lack necessary annotations, are inconsistent in formats, and may be poor in data quality [64].”

• cites [64] sourcePartly supported

  Step 4 (Characterizing the data) describes incomplete or erroneous metadata, and Step 5 (Data quality) describes poorly formatted or low-quality cancer genomic data that can prevent analysis. The paper does not establish the same problems for chemical-compound or protein-structure databases, or address assembling those databases for drug discovery. No page numbers are available in our copy. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Many generative AI applications in health care rely on LLMs and are trained on large amounts of internet data without being properly screened and filtered.

“Moreover, many generative AI applications in health care rely on LLMs and are trained on large amounts of internet data without being properly screened and filtered [70].”

• cites [70] sourcePartly supported

  The AI ethics section says many publicly deployed generative AI applications have used internet data scraped without review or training data curated with insufficient attention to bias and misinformation. But the AI technology section distinguishes non-medical LLMs from medical ones, saying that scraping internet data sight unseen may be illegal or impossible for medical LLMs. The paper therefore does not establish that many healthcare applications are trained that way. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative AI is projected to build and use next-generation synthetic gene networks for various AI applications in health care, including medical diagnostics, drug discovery, and medical research.

“From a broader perspective, generative AI is projected to build and use next-generation synthetic gene networks for various AI applications in health care, including medical diagnostics, drug discovery, and medical research [73].”

• cites [73] sourcePartly supported

  The Abstract and Conclusions project medical applications for next-generation synthetic gene networks, and the Whole-Cell Biosensors and Response Systems section discusses potential medical sensing and imaging uses; page numbers are unavailable in our copy. The paper does not project that generative AI will build or use these networks, nor does it identify drug discovery as an AI application. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Deep generative models are used to create synthetic images such as skin lesions, pathology slides, colon mucosa, and chest x-rays, thereby greatly improving the reproducibility of medical data.

“For instance, deep generative models are used to create synthetic images such as skin lesions, pathology slides, colon mucosa, and chest x-rays, thereby greatly improving the reproducibility of medical data [71].”

• cites [71] sourcePartly supported

  The Fidelity tests section lists synthetic images of skin lesions, pathology slides, colon mucosa and chest X-rays. It says synthetic data could contribute to improving reproducibility, but does not show that generating these images has greatly improved it; the section also cautions that current generative models are not ready for off-the-shelf use. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Robustness, generalizability, and other vulnerability tests can prevent integrity and availability attacks and data breaches.

“Robustness, generalizability, and other vulnerability tests (eg, black-box and white-box tests) can further prevent integrity and availability attacks and data breaches [88].”

• cites [88] sourcePartly supported

  The discussion on p. 3 says data-free approaches reveal model vulnerabilities, and §4.2 (p. 10) evaluates white-box and black-box attacks. The paper also calls for robust models against adversarial attacks (p. 2), but it does not say that vulnerability tests prevent integrity or availability attacks or data breaches. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Trust is a complex concept and is built upon human users' interaction and consequent assessment of the system from cognitive, emotional, and social dimensions.

“Nevertheless, trust is a complex concept and is built upon human users’ interaction and consequent assessment of the system from cognitive, emotional, and social dimensions [91-93].”

• cites [91] sourceNot checked — no full text of the source
• cites [92] sourcePartly supported

  The Introduction and Section 2.2 describe users assessing technology through human-like and system-like trusting beliefs; the discussion of social presence also connects social cues and emotion-based qualities to trust. The paper does not present cognitive, emotional, and social dimensions as a three-part account of how trust is built, as the citing sentence implies. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [93] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The current consensus is that AI trustworthiness itself is a complex measurement with multiple dimensions, such as reliability, resilience, accuracy, and completeness.

“The current consensus is that AI trustworthiness itself is a complex measurement with multiple dimensions, such as reliability, resilience, accuracy, and completeness [63,90].”

• cites [63] Artificial Intelligence Risk Management Framework (AIRMF1.0)National Institute of Standards and Technology20232023-09-20https://doi.org/10.6028/NIST.AI.100-1 sourcePartly supported

  Section 3, “AI Risks and Trustworthiness” (page not supplied), describes trustworthiness through multiple characteristics, including reliability and resilience; the framework also says accuracy contributes to trustworthiness. The cited passages do not identify completeness as a characteristic, and “Availability of reliable metrics” says there is currently a lack of consensus on robust, verifiable ways to measure trustworthiness. Thus the multidimensional point is supported, but the claim’s list and measurement-consensus framing are not fully supported. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [90] MattioliJSohierHDelabordeAAmokrane-FerkaKAwadidAChihaniZKhalfaouiSPedrozaGTowards a holistic approach for AI trustworthiness assessment based upon aids for multi-criteria aggregationProceedings of the Safe AI 2023-The AAAI’s Workshop on Artificial Intelligence Safety2023SafeAI '23February 13-14, 2023Washington, DCNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

AI responsibility is centered on the uniqueness of AI, especially generative AI, in "human centricity, social responsibility, and sustainability".

“Moreover, this measure is centered on the uniqueness of AI, especially generative AI, in “human centricity, social responsibility, and sustainability” [63].”

• cites [63] Artificial Intelligence Risk Management Framework (AIRMF1.0)National Institute of Standards and Technology20232023-09-20https://doi.org/10.6028/NIST.AI.100-1 sourcePartly supported

  The Executive Summary (page unavailable) explicitly says that core concepts in responsible AI emphasize human centricity, social responsibility, and sustainability. It also describes AI risks as distinctive, and Appendix B mentions challenging risks related to generative AI, but neither passage singles out generative AI in connection with those three concepts. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S4 (based).

By systematically analyzing the diverse range of generative AI systems across health care settings using the key factors, the heterogeneous landscape of generative AI in health care can be classified into 5 overarching categories: medical diagnostics, drug discovery, virtual health assistants, medical research, and clinical decision support.

“By systematically analyzing the diverse range of generative AI systems across health care settings using the key factors discussed previously, we can classify the heterogeneous landscape of generative AI in health care into 5 overarching categories: medical diagnostics, drug discovery, virtual health assistants, medical research, and clinical decision support.”

• AssumptionRests on an assumption: The key factors used (setting, users, input data, output data, personalization level, workflow integration, validation needs, impact, risks, and human-AI collaboration) are sufficient to differentiate and classify generative AI applications in health care.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Fui-Hoon et al discuss privacy issues related to generative AI, such as the potential disclosure of sensitive or private information, the widening of the digital divide, and the collection of personal and organizational data.

“They discuss several privacy issues related to generative AI, such as the potential disclosure of sensitive or private information by generative AI systems, the widening of the digital divide, and the collection of personal and organizational data by these systems, which raises concerns about security and confidentiality.”

• cites [48] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Clinical decision support with generative AI could improve outcomes and reduce errors.

“This could improve outcomes and reduce errors.”

May combine two claims.

CorGAN enhances predictive accuracy, providing more effective assistance to medical professionals.

“This approach enhances predictive accuracy, providing more effective assistance to medical professionals [43].”

• cites [43] sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Generative adversarial networks can be used to address increasing privacy concerns and strengthen privacy regulations in various applications, including medical image analysis.

“It highlights how GANs can be used to address increasing privacy concerns and strengthen privacy regulations in various applications, including medical image analysis.”

• cites [29] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Differential privacy is an approach that adds noise to the data to prevent the identification of individuals.

“For instance, differential privacy is an approach that adds noise to the data to prevent the identification of individuals.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

GANs can create new and nonreal data points.

“GANs can create new and nonreal data points.”

May combine two claims.

In privacy and security models using generative adversarial networks, the generator can be modeled as an attacker aiming to fool a defender or as a defender resisting a powerful attacker.

“In the privacy and security models using GANs, the generator can be modeled in two ways: (1) as an attacker aiming to fool a defender (the discriminator) to simulate an attack scenario and (2) as a defender resisting a powerful attacker (the discriminator) to simulate a defense scenario.”

• cites [29] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Protection mechanisms for decentralized learning include embedding a buried point layer in local models to detect abnormal changes and block attackers and integrating GAN with federated learning to produce realistic data without privacy leakage.

“Protection mechanisms include embedding a “buried point layer” in local models to detect abnormal changes and block attackers and integrating GAN with FL to produce realistic data without privacy leakage.”

Classified as probably not a substantial claim. May combine two claims.

The proposed framework clarifies the heterogeneous landscape of generative AI in health care and enables a trend analysis across categories.

“The proposed framework clarifies the heterogeneous landscape of generative AI in health care and enables a trend analysis across categories.”

• AssumptionRests on an assumption: The key factors used (setting, users, input data, output data, personalization level, workflow integration, validation needs, impact, risks, and human-AI collaboration) are sufficient to differentiate and classify generative AI applications in health care.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Other advanced statistical and machine learning techniques attempt to balance data utility and privacy.

“Other advanced statistical and machine learning techniques attempt to balance data utility and privacy.”

May combine two claims.

medGAN produces synthetic medical data suitable for distribution analysis, predictive modeling, and medical expert evaluations, minimizing privacy risks associated with identity and attributes.

“The outcome is synthetic medical data suitable for various uses, such as distribution analysis, predictive modeling, and medical expert evaluations, minimizing the privacy risks associated with both identity and attributes.”

• cites [40] ChoiEBiswalSMalinBDukeJStewartWFSunJGenerating multi-label discrete patient records using generative adversarial networksProceedings of the 2017 Machine Learning for Health Care Conference2017MLHC '17August 18-19, 2017Boston, MA120Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

In medical research, the widespread use of medical data, particularly in image analysis, raises significant concerns about the potential exposure of individual identities.

“In medical research, the widespread use of medical data, particularly in image analysis, raises significant concerns about the potential exposure of individual identities.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In research settings, generative AI can formulate novel hypotheses by making unexpected combinations of concepts, mimicking human creativity and intuition.

“In research settings, generative AI can formulate novel hypotheses by making unexpected combinations of concepts, mimicking human creativity and intuition.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative AI has the potential to make a breakthrough in medical diagnostic applications, given its capability to build models using multimodal medical data.

“This subset of AI has the potential to make a breakthrough in medical diagnostic applications, given its capability to build models using multimodal medical data [5].”

• cites [5] sourceNot checked — no full text of the source

May combine two claims.

Generative AI models can be biased, which could lead to inaccurate diagnoses and treatments.

“Another challenge of using generative AI models is that they can be biased, which could lead to inaccurate diagnoses and treatments [7].”

• cites [7] sourceNot checked — no full text of the source

May combine two claims.

Decentralized learning systems can protect data privacy to some extent, but it is not perfect.

“This setup can protect data privacy to some extent, but it is not perfect.”

• cites [35] sourceNot checked — no full text of the source
• cites [36] McMahanBMooreERamageDHampsonSCommunication-efficient learning of deep networks from decentralized dataProceedings of the 20th International Conference on Artificial Intelligence and Statistics2017AISTATS '17April 20-22, 2017Fort Lauderdale, FLNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

A malicious server can reveal user-level privacy in distributed learning systems by training a multitask GAN with auxiliary identification.

“a malicious server can reveal user-level privacy in distributed learning systems by training a multitask GAN with auxiliary identification.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative AI's unique generative capacity could accelerate scientific advancement.

“This unique generative capacity could accelerate scientific advancement.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Torfi and Fox introduced CorGAN, which focuses on the correlations within medical records.

“In a parallel development, Torfi and Fox [42] introduced Correlation-Capturing Convolutional Generative Adversarial Networks (CorGAN), which focuses on the correlations within medical records.”

• cites [42] Torfi, A, 2020, arXiv Preprint posted online January 25, 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

When integrated into clinical workflows, generative AI may provide physicians with patient-specific treatment suggestions.

“When integrated into clinical workflows, it may also provide physicians with patient-specific treatment suggestions [13].”

• cites [13] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Integrating generative AI into clinical workflows could provide patient-specific suggestions to assist physicians in decision-making.

“Integrating generative AI into clinical workflows could provide patient-specific suggestions to assist physicians in decision-making.”

Probably not a substantial claim: may not pass S4 (based).

The data-intensive nature and complexity of AI systems introduce acute privacy and security vulnerabilities that must be addressed to ensure safe and ethical deployment in clinical settings.

“However, the data-intensive nature and complexity of these systems introduce acute privacy and security vulnerabilities that must be addressed to ensure safe and ethical deployment in clinical settings.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Virtual health assistants increase access and engagement for patients.

“This increases access and engagement for patients.”

May combine two claims.

GAN-based models in decentralized learning systems can mimic data distribution and potentially threaten data privacy.

“The GAN-based models in these systems can mimic data distribution and potentially threaten data privacy.”

• cites [35] sourceNot checked — no full text of the source
• cites [36] McMahanBMooreERamageDHampsonSCommunication-efficient learning of deep networks from decentralized dataProceedings of the 20th International Conference on Artificial Intelligence and Statistics2017AISTATS '17April 20-22, 2017Fort Lauderdale, FLNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Generative AI models can be trained on sensitive, multimodal patient data, which could be exploited by malicious actors.

“Generative AI models can be trained on sensitive, multimodal patient data, which could be exploited by malicious actors.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative AI shows promise for expediting and enhancing drug discovery through inventing optimized molecular structures de novo.

“Generative AI shows promise for expediting and enhancing drug discovery through inventing optimized molecular structures de novo.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Fui-Hoon et al highlight regulatory and policy challenges, such as issues with copyright for AI-generated content, the lack of human control over AI behavior, data fragmentation, and information asymmetries between technology giants and regulatory authorities.

“In addition, they highlight regulatory and policy challenges, such as issues with copyright for AI-generated content, the lack of human control over AI behavior, data fragmentation, and information asymmetries between technology giants and regulatory authorities.”

• cites [48] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The use of electronic health record medical records has significantly advanced medical research while simultaneously amplifying concerns regarding the privacy of this sensitive information.

“The use of EHR medical records has significantly advanced medical research while simultaneously amplifying concerns regarding the privacy of this sensitive information.”

May combine two claims.

Both the EU AI Act and US AI Bill of Rights identify important issues such as potential bias, privacy risks, and the need for human control but tackle them differently-the EU through compliance requirements and the United States through voluntary principles.

“Both identify important issues such as potential bias, privacy risks, and the need for human control but tackle them differently—the EU through compliance requirements and the United States through voluntary principles.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Ethical risks around reliability, accountability, algorithmic bias, and data privacy as well as security risks related to confidentiality, integrity, and availability must be addressed through a human-centric approach.

“Ethical risks around reliability, accountability, algorithmic bias, and data privacy as well as security risks related to confidentiality, integrity, and availability must be addressed through a human-centric approach [55].”

• cites [55] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The US AI Bill of Rights, unveiled in 2023, is a nonbinding set of principles to guide AI development and use focused on algorithmic discrimination awareness, data privacy, notice and explanation of AI, and human alternatives and oversight.

“The US AI Bill of Rights, unveiled in 2023, takes a different approach as a nonbinding set of principles to guide AI development and use focused on concepts such as algorithmic discrimination awareness, data privacy, notice and explanation of AI, and human alternatives and oversight [54].”

• cites [54] Blueprint for an AI bill of rights: making automated systems work for the American peopleThe White House2024-02-19https://www.whitehouse.gov/ostp/ai-bill-of-rights/Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Although generative AI in health care holds great promise, substantial validation is required before real-world deployment.

“Although generative AI in health care holds great promise, substantial validation is required before real-world deployment.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Generative medical research AI may update knowledge and literature databases with "wrong inputs" based on wrong findings in these databases or with synthesized but hallucinated findings.

“For example, generative medical research AI may update knowledge and literature databases with “wrong inputs” based on wrong findings in these databases or with synthesized but hallucinated findings.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Adversarial training is a method to check for the integrity and accountability of AI models.

“Adversarial training is a method to check for the integrity and accountability of AI models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The growth in the use of synthetic data by generative AI creates new concerns about data integrity and AI hallucination.

“The growth in the use of synthetic data by generative AI also creates new concerns about data integrity and AI hallucination.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Generative virtual health assistants may put dangerous advice into knowledge databases based on erroneous data from sources or again put synthesized but hallucinated advice into such databases.

“Similarly, generative virtual health assistants may put dangerous advice into knowledge databases based on erroneous data from sources or again put synthesized but hallucinated advice into such databases.”

May combine two claims.

Generative AI, such as LLMs, often requires large volumes of data to train and therefore faces many existing and new security and privacy threats.

“Nevertheless, because generative AI, such as LLMs, often requires large volumes of data (eg, large volumes of patient data) to train, it faces many existing and new security and privacy threats.”

• AssumptionRests on an assumption: Generative AI such as LLMs often requires large volumes of patient data to train.

May combine two claims.

Malicious use of deepfakes generated by deep generative models could fabricate a patient's medical history to falsely claim insurance or lead to incorrect treatments.

“For instance, the malicious use of deepfakes generated by deep generative models could fabricate a patient’s medical history to falsely claim insurance or lead to incorrect treatments.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Attribute disclosure attacks allow an attacker to infer additional attributes about a patient by learning a subset of other attributes about the same patient.

“(5) attribute disclosure attacks—an attacker can infer additional attributes about a patient by learning a subset of other attributes about the same patient [86].”

• cites [86] Matwin, S, 2015, Advanced Research in Data PrivacyNot checked yet

  Not checked in the free run.

May combine two claims.

The EU law institutes enforceable accountability around risky AI, whereas the US bill espouses aspirational AI ethics principles.

“Although the EU law institutes enforceable accountability around risky AI, the US bill espouses aspirational AI ethics principles.”

May combine two claims.

Deidentified data may become identifiable during the data collection and processing phase, and confidential proprietary medical information may be inferred during the data collection and processing phase, leading to data and privacy breaches.

“Deidentified data may become identifiable during the data collection and processing phase, and confidential proprietary medical information, such as drug development and treatment plans, may be inferred during the data collection and processing phase [76], leading to data and privacy breaches.”

• cites [76] sourceNot checked — no full text of the source

May combine two claims.

Generative AI is built on machine learning and AI techniques and faces similar security and privacy threats as general AI and machine learning systems.

“Generative AI is built upon machine learning and AI techniques and hence faces similar security and privacy threats, as summarized in the studies by Liu et al [56] and Hu et al [57].”

• cites Liu et al [56]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Hu et al [57]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Adversaries can use AI technologies to automatically generate large quantities of fake data to poison data to be fed into LLMs, resulting in deteriorated performance of the models and eventually AI hallucination, misinformation or disinformation, and deepfakes.

“Adversaries can use AI technologies to automatically generate large quantities of fake data to poison data to be fed into LLMs, resulting in deteriorated performance of the models (eg, accuracy and fairness) and eventually AI hallucination, misinformation or disinformation, and deepfakes.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The relationship between AI hallucination by GANs and LLMs and synthetic data use is an unknown territory in research and practice, leading to unknown vulnerabilities such as adversarial attacks.

“The relationship between AI hallucination by GANs and LLMs and synthetic data use is an unknown territory in research and practice, leading to unknown vulnerabilities such as adversarial attacks.”

Probably not a substantial claim: may not pass S2 (checkable).

There is a fine line between harmful AI hallucinations or deepfakes and beneficial synthetic data use by generative AI in health care.

“These applications indicate that there is a fine line between harmful AI hallucinations or deepfakes and beneficial synthetic data use by generative AI in health care.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Adversarial training is used to check for fake or realistic features in synthetic medical images created by GANs to avoid fabrication and misleading in the model training process.

“Adversarial training is used to check for fake or realistic features in synthetic medical images created by GANs to avoid fabrication and misleading in the model training process.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the data processing phase of the AI life cycle, image scaling is the main threat; recommended defense strategies include image reconstruction and data randomization.

“(2) data processing phase—image scaling is the main threat; recommended defense strategies include image reconstruction and data randomization;”

• cites [57] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Adversarial attacks can pose long-term risks, such as thwarting AI innovation in health care because of concerns about misdiagnosis, mistreatment, and patient safety.

“In general, adversarial attacks can pose long-term risks, such as thwarting AI innovation in health care because of concerns about misdiagnosis, mistreatment, and patient safety.”

Probably not a substantial claim: may not pass S3 (load-bearing).

CorGAN synthesizes discrete and continuous health care records for model training.

“CorGAN [42] synthesizes discrete and continuous health care records for model training.”

• cites [42] Torfi, A, 2020, arXiv Preprint posted online January 25, 2020Not checked yet

  Not checked in the free run.

May combine two claims.

When being trained with adversarial examples, a diagnostic AI could make an incorrect diagnosis, a conversational virtual assistant could offer harmful advice to patients, and a clinical decision support AI could make the wrong recommendations.

“When being trained with adversarial examples, a diagnostic AI could make an incorrect diagnosis, a conversational virtual assistant could offer harmful advice to patients, and a clinical decision support AI could make the wrong recommendations, to list a few.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Liu et al surveyed the security and privacy attacks related to machine learning and developed a taxonomy.

“Liu et al [56] surveyed the security and privacy attacks related to machine learning and developed a taxonomy.”

• cites [56] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The growth of synthetic medical data in health AI systems raises concerns about the vulnerabilities of such systems and the challenges of the current regulations and policies.

“In addition, the growth of synthetic medical data in health AI systems raises concerns about the vulnerabilities of such systems and the challenges of the current regulations and policies.”

May combine two claims.

The EU AI Act categorizes systems by risk level and introduces mandatory requirements for high-risk AI related to data and documentation, transparency, human oversight, accuracy, cybersecurity, and so on.

“It categorizes systems by risk level and introduces mandatory requirements for high-risk AI related to data and documentation, transparency, human oversight, accuracy, cybersecurity, and so on.”

• cites [53] EU AI Act: first regulation on artificial intelligenceEuropean Parliament20232024-02-16https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligenceNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Many security and privacy threats associated with generative AI in health care are not unique to generative AI and are also prevalent in broader AI systems and machine learning models in health care and other fields.

“It should be noted that although this study primarily discusses various security and privacy threats associated with generative AI in health care (such as AI hallucination in health care), many of these threats are not unique to generative AI systems and are also prevalent in broader AI systems and machine learning models in health care and other fields.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Currently, there are insufficient use cases, especially for rare use cases, to establish clinical reference standards such as clinical quality measures and evaluation metrics to assess risks and benefits.

“Currently, there are just insufficient use cases, especially for rare use cases, to establish clinical reference standards such as clinical quality measures and evaluation metrics to assess risks and benefits.”

May combine two claims.

Even the benevolent use of synthetic medical data faces privacy and security challenges as well as integrity challenges.

“Nevertheless, even the benevolent use of synthetic medical data faces privacy and security challenges as well as integrity challenges.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The taxonomy classifies attacks into three categories: attacks targeting classifiers; attacks violating integrity, availability, and privacy; and attacks with or without specificity.

“The taxonomy classifies those attacks into three categories: (1) attacks targeting classifiers; (2) attacks violating integrity, availability, and privacy (ie, part of confidentiality); and (3) attacks with or without specificity.”

• cites [56] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Liu et al summarize defense techniques in the training phase and the testing and inferring phase of the life cycle of machine learning, for example, data sanitization techniques against data poisoning attacks in the training phase and privacy-preserving techniques against privacy attacks in the testing or inferring phase.

“They also summarize the defense techniques in the training phase and the testing and inferring phase of the life cycle of machine learning, for example, data sanitization techniques against data poisoning attacks in the training phase and privacy-preserving techniques against privacy attacks in the testing or inferring phase.”

• cites [56] sourceNot checked — no full text of the source

May combine two claims.

Both the EU and US highlight growing policy makers' concern over AI's societal impacts and the emergence of either compulsory or optional frameworks aimed at accountability.

“Despite this schism, both highlight growing policy makers’ concern over AI’s societal impacts and the emergence of either compulsory or optional frameworks aimed at accountability.”

May combine two claims.

LLMs in health care face data quality and bias issues, similar to any machine learning model, leading to erroneous medical conclusions or recommendations.

“LLMs in health care also face data quality and bias issues, similar to any machine learning model, leading to erroneous medical conclusions or recommendations [61].”

• cites [61] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

These studies collectively underscore the need for a balanced approach to leveraging the benefits of AI-driven health care innovations while ensuring robust privacy and security measures.

“These studies collectively underscore the need for a balanced approach to leveraging the benefits of AI-driven health care innovations while ensuring robust privacy and security measures.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The principles mentioned in the US AI Bill of Rights do not specifically address generative AI, but they provide a framework for the ethical and responsible use of all AI technologies, including generative AI.

“Although the principles mentioned in the US AI Bill of Rights do not specifically address generative AI, they provide a framework for the ethical and responsible use of all AI technologies, including generative AI.”

• cites [54] Blueprint for an AI bill of rights: making automated systems work for the American peopleThe White House2024-02-19https://www.whitehouse.gov/ostp/ai-bill-of-rights/Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The EU legislates binding rules that companies must follow, whereas the United States issues guidance that organizations may freely adopt.

“Ultimately, the EU legislates binding rules that companies must follow, whereas the United States issues guidance that organizations may freely adopt.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The EU AI Act, passed in 2023, is the first comprehensive legal framework to specifically regulate AI systems.

“The EU AI Act, passed in 2023, is the first comprehensive legal framework to specifically regulate AI systems [53].”

• cites [53] EU AI Act: first regulation on artificial intelligenceEuropean Parliament20232024-02-16https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligenceNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the data collection phase of the AI life cycle, main security threats include databases, fake data, data breaches, and sensor attacks; defense strategies include data sanitization and data government.

“(1) data collection phase—main security threats include databases, fake data, data breaches, and sensor attacks; defense strategies include data sanitization and data government;”

• cites [57] sourceNot checked — no full text of the source

May combine two claims.

Sensor data can be spoofed, tampered with, and thus poisoned.

“Sensor data can be spoofed [65,66], tampered with, and thus poisoned.”

• cites [65] ParkYSonYShinHKimDThis ain't your dose: sensor spoofing attack on medical infusion pumpProceedings of the 10th USENIX Workshop on Offensive Technologies2016WOOT '16August 8-9, 2016Austin, TXNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [66] sourceNot checked — no full text of the source

May combine two claims.

For generative AI such as ChatGPT, the EU AI Act mandates transparency requirements, such as disclosing AI-generated content, designing models to prevent illegal content generation, and publishing training data summaries.

“Specifically, for generative AI such as ChatGPT, the EU AI Act mandates transparency requirements, such as disclosing AI-generated content, designing models to prevent illegal content generation, and publishing training data summaries.”

• cites [53] EU AI Act: first regulation on artificial intelligenceEuropean Parliament20232024-02-16https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligenceNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Hu et al present an overall framework of attacks and defense strategies based on five phases of the AI life cycle.

“Similarly, Hu et al [57] present an overall framework of attacks and defense strategies based on the following five phases of the AI life cycle:”

• cites [57] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

AI may cause physical harm to individual patients, damage a hospital's reputation owing to AI incidents, and even endanger society if it is weaponized (eg, being used to disrupt the global drug manufacturing and supply chain).

“For example, AI may cause physical harm to individual patients, damage a hospital’s reputation owing to AI incidents, and even endanger society if it is weaponized (eg, being used to disrupt the global drug manufacturing and supply chain).”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Accuracy and reliability help to improve the integrity of an AI system.

“For instance, accuracy and reliability help to improve the integrity of an AI system.”

Classified as probably not a substantial claim. May combine two claims.

AI harm can occur to individuals, organizations, and societies.

“AI harm can occur to individuals, organizations, and societies.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

AI responsibility is another key measure in AI risk assessment.

“AI responsibility is another key measure in AI risk assessment.”

May combine two claims.

Generative AI enhances the ability to profile patients, thereby increasing the risk of privacy violations and attacks, although this capability is not unique to AI.

“Moreover, generative AI enhances the ability to profile patients, thereby increasing the risk of privacy violations and attacks, although this capability is not unique to AI.”

Probably not a substantial claim: may not pass S2 (checkable).

The implementation contexts of generative AI matter.

“The implementation contexts of generative AI also matter.”

• AssumptionRests on an assumption: The implementation contexts of generative AI matter.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

AI responsibility is a multifaceted measure depending on many other metrics and factors such as the ethical framework (eg, biases, fairness, and transparency) and legal perspective (eg, accountability and traceability).

“In other words, AI responsibility is a multifaceted measure depending on many other metrics and factors such as the ethical framework (eg, biases, fairness, and transparency) and legal perspective (eg, accountability and traceability).”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

AI inscrutability adds obfuscations to AI risk assessment to identify threats and vulnerabilities as well as biases owing to the lack of transparency and explainability in AI, especially in generative AI based on deep learning.

“Although AI inscrutability is not directly related to security and privacy, it adds obfuscations to AI risk assessment to identify threats and vulnerabilities as well as biases owing to the lack of transparency and explainability in AI, especially in generative AI based on deep learning.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims. The quote only partly states this claim.

Explainability and interaction experience can be the determinants of the AI trustworthiness of a chatbot application on the patient portal, whereas reliability, accuracy, and completeness are significant factors in the AI trustworthiness of a radiology diagnosis AI for radiologists.

“application on the patient portal, whereas reliability, accuracy, and completeness are significant factors in the AI trustworthiness of a radiology diagnosis AI for radiologists.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Availability attacks such as denial of service can be used to flood conversational health AI applications to assess their resilience and availability before deployment, and integrity attacks with poisoned data can be used to test the stability of model performance and generalizability.

“For example, availability attacks such as denial of service can be used to flood conversational health AI applications to assess their resilience and availability before deployment, and integrity attacks with poisoned data can be used to test the stability of model performance and generalizability [57].”

• cites [57] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Health care data are messy and tend to have organic biases.

“As discussed before, health care data are messy and tend to have organic biases (eg, a hospital specializes in serving a particular patient demographic, attending to gender-specific health requirements or offering dedicated care for rare diseases).”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Algorithmic outputs highlighting potential issues such as biases or errors increase awareness, but this does not intrinsically amount to AI responsibility.

“Similarly, although algorithmic outputs highlighting potential issues such as biases or errors increase awareness, this does not intrinsically amount to AI responsibility.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

AI harm is a risk measure highly related to AI responsibility and trustworthiness.

“Hence, AI harm is a risk measure highly related to AI responsibility and trustworthiness.”

Probably not a substantial claim: may not pass S2 (checkable).

Establishing genuine public trust and accountability in health care ultimately falls on the shoulders of health care administrators, innovators, and engaged communities, rather than solely on the algorithms themselves.

“Although progress indicators may serve as signals for these desired attributes, establishing genuine public trust and accountability in health care ultimately falls on the shoulders of health care administrators, innovators, and engaged communities, rather than solely on the algorithms themselves.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

True accountability involves diligent human investigation of problems that surface, enacting appropriate recourse, and continuous authority oversight.

“True accountability involves diligent human investigation of problems that surface, enacting appropriate recourse, and continuous authority oversight.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Trust emerges from human-centered institutional processes, and responsibility stems from human accountability.

“Although metrics aim to make models appear more trustworthy and responsible in reality, trust emerges from human-centered institutional processes, and responsibility stems from human accountability.”

May combine two claims.

Excessive use of prompts may reveal copyright-protective data, proprietary research findings, and training models or algorithms.

“The excessive use of prompts may reveal copyright-protective data, proprietary research findings (eg, chemical compounds of a new drug), and training models or algorithms.”

Probably not a substantial claim: may not pass S2 (checkable).

Trustworthiness and responsibility are contingent on extensive institutional support structures rather than innate model capabilities.

“Thus, trustworthiness and responsibility are contingent on extensive institutional support structures rather than innate model capabilities.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Strategies such as developing AI risk assessment protocols; formulating specific metrics for generative AI such as inscrutability, trustworthiness, responsibility, and harm; and ongoing model monitoring can help mitigate risks.

“Strategies such as developing AI risk assessment protocols; formulating specific metrics for generative AI such as inscrutability, trustworthiness, responsibility, and harm; and ongoing model monitoring can help mitigate risks.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Various algorithms used in patient matching, diagnosis, and other proprietary applications often lack transparency because of their closed nature or intellectual property constraints.

“Various algorithms used in patient matching, diagnosis, and other proprietary applications often lack transparency because of their closed nature or intellectual property constraints.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

A medical chatbot that was thoroughly tested in adult populations may not be very useful in teenage populations.

“A medical chatbot that was thoroughly tested in adult populations may not be very useful in teenage populations.”

• AssumptionRests on an assumption: A medical chatbot thoroughly tested in adult populations may not be very useful in teenage populations.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

AI risks, including those of generative AI in health care, can emerge in a variety of ways at any phase of an AI project.

“AI risks, including those of generative AI in health care, can emerge in a variety of ways at any phase of an AI project.”

Classified as probably not a substantial claim.

Generative AI models, especially when based on deep learning, can operate as "black boxes," making it even more difficult for practitioners to understand how conclusions or recommendations are derived.

“Generative AI models, especially when based on deep learning, can operate as “black boxes,” making it even more difficult for practitioners to understand how conclusions or recommendations are derived.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Developing trustworthy AI and following responsible AI practices can reduce or avoid AI harm.

“Developing trustworthy AI and following responsible AI practices can reduce or avoid AI harm.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Many other AI metrics or factors, such as transparency, explainability, robustness, fairness, and user interactions or perceptions, can be the antecedents of AI trustworthiness.

“Many other AI metrics or factors, such as transparency, explainability, robustness, fairness, and user interactions or perceptions, can be the antecedents of AI trustworthiness.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Responsibility mainly manifests in organizational commitment to discovering issues, working with experts to properly assess AI harms, and instituting robust redress processes with stakeholder input.

“Metrics may aim for AI to appear more responsible, but responsibility mainly manifests in organizational commitment to discovering issues, working with experts to properly assess AI harms, and instituting robust redress processes with stakeholder input.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Collaborative efforts between AI developers, health care providers, policy makers, and domain experts will be critical to unlocking the benefits of generative AI while also prioritizing ethics, accountability, and safety.

“Collaborative efforts between AI developers, health care providers, policy makers, and domain experts will be critical to unlocking the benefits of generative AI while also prioritizing ethics, accountability, and safety.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The inscrutability of generative AI is a continuation of the broader issue of transparency in health care AI.

“Therefore, although the inscrutability of generative AI presents specific challenges owing to the complexity and novelty of these models, it is a continuation of the broader issue of transparency in health care AI.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Many of the recommendations proposed can be applied to AI in non-health care contexts.

“Hence, many of the recommendations we propose in the subsequent section can be applied to AI in non–health care contexts.”

Probably not a substantial claim: may not pass S4 (based).

Clinical and nonclinical data have significantly different risks.

“The separation of clinical and nonclinical data is necessary, given the significantly different risks in these 2 types of data.”

• AssumptionRests on an assumption: Clinical and nonclinical data have significantly different risks.

Probably not a substantial claim: may not pass S2 (checkable).

It may be challenging to humanize AI systems and transfer attributes such as trustworthiness to the algorithms themselves.

“It may be challenging to humanize AI systems and transfer attributes such as trustworthiness to the algorithms themselves.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Many AI algorithms, even those not based on generative techniques, face similar scrutiny regarding their lack of transparency.

“Therefore, many of them, even those that are not based on generative techniques, face similar scrutiny regarding their lack of transparency.”

Classified as probably not a substantial claim.

AI trustworthiness helps reduce the probability and magnitude of security and privacy risks throughout the life cycle of generative AI in health care.

“Similar to AI inscrutability, although AI trustworthiness is not a direct measure of security and privacy risks, it helps reduce the probability and magnitude of such risks throughout the life cycle of generative AI in health care.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

In some cases, verification and validation are about not only factual accuracy but also communications and perceptions as well as cultures.

“In some cases, verification and validation are about not only factual accuracy but also communications and perceptions as well as cultures.”

Probably not a substantial claim: may not pass S2 (checkable).

Public trust fundamentally arises from the ethical data governance, risk communication, and oversight procedures instantiated by organizations.

“Indicators of model transparency, reliability, or accuracy may engender confidence among stakeholders, but public trust fundamentally arises from the ethical data governance, risk communication, and oversight procedures instantiated by organizations.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Modifying training processes and models as well as training methods may help to control AI risks in the model training and building phase.

“Modifying training processes and models as well as training methods may also help to control AI risks in this phase [57].”

• cites [57] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

Without robust governance and review processes overseeing development, data practices, and risk monitoring, claims of AI trustworthiness lack substantiation.

“Without robust governance and review processes overseeing development, data practices, and risk monitoring, claims of AI trustworthiness lack substantiation.”

Probably not a substantial claim: may not pass S2 (checkable).

Deploying generative AI technologies introduces complex privacy and security challenges that must be proactively addressed to ensure safe and effective use.

“However, deploying these data-intensive technologies also introduces complex privacy and security challenges that must be proactively addressed to ensure the safe and effective use of these systems.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A deep learning model can carry significantly different risks than a probabilistic learning model.

“For example, a deep learning model can carry significantly different risks than a probabilistic learning model.”

• AssumptionRests on an assumption: A deep learning model can carry significantly different risks than a probabilistic learning model.

Probably not a substantial claim: may not pass S3 (load-bearing).

AI trustworthiness can be context dependent.

“AI trustworthiness can also be context dependent.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

A generative drug discovery AI may find a new molecular formula for a biochemical weapon.

“For example, a generative drug discovery AI may find a new molecular formula for a biochemical weapon.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

If an IT system automatically runs behind the scenes to assist the work and decisions of human users, a trusting relationship must be established for users to interact with and rely on the system.

“The fundamental argument is that if an IT system automatically runs behind the scenes to assist the work and decisions of human users, a trusting relationship must be established for users to interact with and rely on the system [91].”

• cites [91] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Gesture and face recognition AI for medical diagnosis may need to be culturally sensitive to be useful.

“Gesture and face recognition AI for medical diagnosis may need to be culturally sensitive to be useful.”

• AssumptionRests on an assumption: Gesture and face recognition AI for medical diagnosis may need to be culturally sensitive to be useful.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Integrating generative AI systems into health care offers immense potential to transform medical diagnostics, research, treatment planning, and patient care.

“Integrating generative AI systems into health care offers immense potential to transform medical diagnostics, research, treatment planning, and patient care.”

Probably not a substantial claim: may not pass S3 (load-bearing).

AI responsibility endorses responsible AI practices that facilitate the discovery of the negative consequences and risks of AI, including the security and privacy risks of generative AI.

“Again, although this measure does not directly evaluate security and privacy risks, it endorses responsible AI practices that facilitate the discovery of the negative consequences and risks of AI, including the security and privacy risks of generative AI.”

Probably not a substantial claim: may not pass S2 (checkable).

The development and deployment of generative AI add complexity to the AI responsibility measure owing to its possible, unintended, but profound negative consequences and risks to human society.

“The development and deployment of generative AI add complexity to this measure owing to its possible, unintended, but profound negative consequences and risks to human society.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The challenge of inscrutability is not unique to generative AI and has been a long-standing issue in the broader field of AI, particularly in health care.

“we acknowledge that the challenge of inscrutability is not unique to generative AI and has been a long-standing issue in the broader field of AI, particularly in health care.”

Probably not a substantial claim: may not pass S2 (checkable).

The problem of inscrutability becomes pronounced in the context of generative AI because of its complex and often opaque decision-making processes, which can amplify the challenges already faced in health care AI.

“Nevertheless, the problem of inscrutability becomes pronounced in the context of generative AI because of its complex and often opaque decision-making processes, which can amplify the challenges already faced in health care AI.”