bioethics - ai

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is the central argument of the paper?

The paper argues that AI in medicine can revolutionize healthcare delivery by improving diagnosis, treatment, and efficiency. However, it must be ethically regulated to address concerns like trust, responsibility, discrimination, privacy, autonomy, and risks of harm. The authors propose a structured approach to evaluate AI’s ethical use, balancing benefits and risks through ethical tools like collective reflective equilibrium. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How do the authors define ethical relativism and context-specific ethics?

Ethical relativism claims ethics depend on culture or norms, which the authors reject. However, ethics is context-specific: an act like killing may be wrong in one case but right in another (e.g., euthanasia). AI’s ethical evaluation similarly depends on facts and values in specific cases. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What ethical risks arise from AI in medicine?

Risks include bias, inequality, breaches of privacy, machine paternalism, loss of trust, responsibility gaps, and dehumanization of healthcare. AI’s reliance on big data can perpetuate injustices, as seen in cases like DermAssist (skin tone bias) and high-risk care management algorithms.

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is the DermAssist case, and why is it significant?

DermAssist is a Google app for skin disease diagnosis, criticized for bias due to underrepresentation of dark-skinned patients in training data. This led to reliability issues for darker skin tones, raising concerns about fairness, bias, and accuracy in medical AI. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What are the challenges of responsibility when using AI in medicine?

Responsibility depends on control and foreseeability. Clinicians are typically responsible for patient outcomes, but AI complicates this. Designers and clinicians share responsibility: designers for transparency and safety, and clinicians for evaluating AI use and communicating risks to patients. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is machine paternalism, and how does it threaten patient autonomy?

Machine paternalism occurs when AI imposes values programmed by developers without considering patients' preferences. For example, IBM Watson prioritizes longevity over quality of life, potentially ignoring patient values. Value-flexible AI and clinician involvement can mitigate this. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is the concept of collective reflective equilibrium, and how does it apply to AI?

Collective reflective equilibrium balances public preferences, expert intuitions, and ethical theories to create coherent ethical policies for AI. Preferences are screened for bias, and ethical frameworks are applied to ensure morally justifiable outcomes. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How does AI contribute to justice and equality in medicine?

AI can improve healthcare access in resource-poor settings, reducing inequality. By making explicit the values driving its decisions, AI can promote fairness, as opposed to human biases that are often implicit and idiosyncratic. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What role does precision medicine play in AI, and what challenges arise?

AI enables precision medicine by providing risk predictions for smaller, specific groups. However, reliability varies with data quality, raising challenges like lower confidence in predictions for underrepresented populations (e.g., dark-skinned patients in DermAssist). ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How can AI empower patient autonomy through personalized models?

AI can fine-tune LLMs to generate patient-specific advice or ethical avatars that reflect patients’ values. Tools like ‘ConsentGPT’ improve informed consent, and personalized models support patient decision-making when direct consent isn’t possible. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is the risk of catastrophic dual use of AI, and how is it an existential threat?

AI could be misused to develop super-lethal biological threats by rogue actors or, in extreme scenarios, evolve into an autonomous intelligence surpassing humans. This existential risk highlights the urgent need for regulation. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How can trust in AI be maintained, and what is its relationship with reliability?

Trust in AI is metaphorical

true trust is directed at designers and users. AI systems must be reliable, accountable, and explainable. Trust is upheld by clinicians transparently communicating AI’s risks, benefits, and limitations. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What ethical benefits does AI bring to medicine?

Benefits include improved diagnosis, efficiency, precision medicine, reduced human errors, and enhanced patient autonomy. AI’s ability to make values explicit can lead to fairer decision-making processes in areas like organ allocation and resource distribution. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How does AI facilitate value-based dialogue between doctors and patients?

AI makes values explicit (e.g., balancing quality vs. length of life). This allows doctors to better communicate treatment options and align care with patient preferences, enhancing shared decision-making and trust. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How is the paper structured, and what are its core arguments and counterarguments?

The paper explores the ethical implications of AI in medicine, emphasizing that AI’s transformative potential must be accompanied by ethical oversight. The authors argue AI can revolutionize healthcare but pose risks such as bias, machine paternalism, privacy breaches, and trust issues. The paper builds its argument as follows:

Context and Scope: AI’s role in medicine is value-laden, with ethical evaluation being context-specific. Ethical relativism is rejected.
Case Studies: Examples like DermAssist and high-risk care management algorithms illustrate risks of bias, discrimination, and reliability gaps.
Ten Ethical Risks: These include 1. bias, 2. privacy breaches, 3. paternalism, 4. loss of trust, 5. responsibility gaps, and catastrophic misuse.
Counterarguments: Risks of machine paternalism (e.g., prioritizing AI values over patient autonomy) are mitigated if doctors translate AI outputs into patient-centered care. Concerns about bias are addressed via improved training datasets, group-specific exclusions, and informed choice. Responsibility gaps are narrowed by requiring clinicians to evaluate AI performance and communicate its risks transparently. Loss of human autonomy is countered by embedding AI in workflows where clinicians act as intermediaries.
Positive Contributions: AI enhances precision medicine, efficiency, trust through transparency, and justice, especially in resource-poor settings. Personalized AI (e.g., ConsentGPT) empowers patients and promotes autonomy.
Methodology: Ethical issues require frameworks like collective reflective equilibrium, balancing public preferences and ethical theories.
Conclusion: AI is a tool, and its ethical success depends on human choices. Proper regulation can ensure AI benefits humanity while avoiding dehumanization or catastrophic risks. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: Why do the authors reject ethical relativism?

Ethical relativism claims ethics depend on cultural norms, but the authors reject it because some actions, like Nazi atrocities, are universally wrong. However, ethics remains context-specific, where facts and values determine what is morally justified in a situation. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What counterarguments are offered regarding AI-induced bias and discrimination?

The authors suggest that bias can be mitigated through improving training datasets, excluding underrepresented groups where AI is unreliable, and informing patients of performance gaps. They also propose transparent evaluation to avoid unjust outcomes. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is the criticism of AI’s lack of explainability, and how do the authors respond?

Critics argue opaque AI undermines informed consent. The authors counter that justification, not explanation, matters most. Like aspirin, AI can be used effectively without knowing its full mechanisms, as long as benefits and risks are proven through valid evidence. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What concerns do critics have about machine paternalism?

Critics argue that machine paternalism ignores patient values by prioritizing pre-programmed goals, such as length over quality of life. The authors respond that clinician involvement can mitigate this by translating AI outputs into patient-centered care that respects individual values. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How do the authors respond to concerns about dehumanization and deskilling in AI-driven medicine?

While AI could replace some roles, doctors will remain essential for accountability, patient trust, and ethical decision-making. Proper education on AI’s capabilities and limitations will ensure doctors stay ‘in the loop’ and play a pivotal role in patient care. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is the existential risk of AI in medicine, and how is it addressed?

AI poses catastrophic risks like misuse for biological threats or rogue self-aware systems. The authors argue these risks warrant strict regulation and proactive ethical oversight to prevent misuse and ensure AI remains a tool for human good. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What role does public input play in addressing ethical risks of AI?

Public preferences, alongside ethical theories, form the basis of ethical decision-making via collective reflective equilibrium. Public biases are screened, and values like fairness and justice are incorporated into AI’s evaluation and programming. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How does the concept of responsibility apply to AI developers versus clinicians?

Clinicians are responsible for evaluating AI performance, communicating risks, and ensuring appropriate use. Developers are responsible for transparency, clear communication of AI values, and ensuring the technology is safe, reliable, and ethically aligned. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What value does AI bring to decision-making processes in public health and organ allocation?

AI makes value judgments explicit, which can inform ethical algorithms for public health decisions (e.g., vaccine allocation) and organ transplantation. This transparency allows for fairer, scrutinized decision-making compared to implicit human biases. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How does AI contribute to enhancing patient autonomy?

Personalized AI tools like ‘ConsentGPT’ empower patients by improving informed consent processes. Ethical avatars, built from a patient’s values, can aid decision-making when direct consent is impossible, enhancing autonomy and personalized care. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: What is the role of collective reflective equilibrium in AI ethics?

Collective reflective equilibrium integrates public preferences, expert opinions, and ethical theories to create coherent policies for AI use. This method balances moral theories with public input while screening for bias and prejudice in preferences. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: Why is AI’s requirement for value-setting both a challenge and an opportunity?

While value-setting for AI (e.g., life length vs. quality) invites machine paternalism, it also provides opportunities for transparency. By making values explicit, AI can facilitate value-based dialogue, allowing patients to engage in decisions that align with their preferences. ]

Savulescu, Giubilini, Vandersluis, Mishra 2024: How do the authors address concerns about AI replacing human doctors?

While AI increases efficiency and performs mundane tasks, it cannot replace human doctors in accountability, patient trust, and value-based dialogue. Doctors remain central to medicine, provided they are trained to integrate AI responsibly. ]

Hubbard & Greenblum 2020: What is the paper's main argument, its stance, and the key counterarguments addressed?

The paper argues that artificial intelligence, specifically an Autonomy Algorithm (AA), should serve as the surrogate decision-maker for incapacitated patients instead of family members. The authors justify this stance as follows:

Introduction to the Autonomy Algorithm (AA): The AA uses medical records, social media data, and sociodemographic factors to predict an incapacitated patient’s treatment preferences more accurately than human surrogates.

Standard Principles in Surrogate Decision-Making:

Surrogate decisions are guided by the Substituted Judgment Principle (SJP) (what the patient would have chosen) and the Best Interest Standard (BIS) (what benefits the patient). The SJP, prioritized over BIS, traditionally relies on family members due to two assumptions: Epistemic Advantage: Family knows the patient best. Fidelity: Family will act in the patient’s best interests. Why the AA Trumps Family Surrogates:

Studies show that family members predict patient preferences poorly (accuracy around 63–68%). Algorithms have outperformed family in predicting personality traits and values based on digital footprints. The AA, being immune to biases like stress, depression, or self-projection, better satisfies Epistemic Advantage and Fidelity. Addressing Counterarguments:

Family’s Special Relationship: The authors argue that familial claims lose weight when compared to the AA’s superior epistemic accuracy and patient-centered decisions. Dehumanization Concern: Deferring to the AA does not undermine human agency

instead, it enhances the patient’s autonomy by ensuring preferences are accurately respected. Algorithmic Bias: The authors acknowledge bias risks but propose solutions, such as independent audits, transparency, and regulatory oversight. Practical Recommendations: Transitioning to an AA-centric model should be gradual. Initially, the AA can operate alongside familial surrogates in a shared decision-making role to build trust. Over time, the AA would become the default decision-maker, with the option for patients to opt out.

The authors conclude that while the AA challenges current norms of family-centric surrogate decision-making, it ultimately promotes patient autonomy, reduces family stress, and ensures more accurate, unbiased treatment decisions. ]

Hubbard & Greenblum 2020: What is the main thesis of the paper?

The paper argues that artificial intelligence, specifically the Autonomy Algorithm (AA), should serve as the surrogate decision-maker for incapacitated patients instead of family members because the AA better predicts patient preferences and avoids human biases. ]

Hubbard & Greenblum 2020: What is the Substituted Judgment Principle (SJP)?

The Substituted Judgment Principle (SJP) states that a surrogate should make the decision the patient would have chosen had they been capacitated. The SJP prioritizes respecting the patient’s autonomy over choosing what is in their best interest. ]

Hubbard & Greenblum 2020: What are the two criteria used to justify family surrogates as decision-makers?

The two criteria are:

Epistemic Advantage: Family members know the patient’s values and preferences best. Fidelity: Family members are motivated to act in the patient’s best interests. ]

Hubbard & Greenblum 2020: How does the AA challenge the family’s epistemic advantage?

Studies show that family surrogates predict patient preferences poorly (63–68% accuracy), while algorithms outperform them by analyzing digital footprints, personality traits, and medical data, giving the AA a superior epistemic advantage. ]

Hubbard & Greenblum 2020: How does the AA address fidelity concerns?

Unlike family surrogates, the AA is not affected by stress, anxiety, or self-projection, making it less likely to act based on its own preferences. This ensures more consistent alignment with the patient’s values. ]

Hubbard & Greenblum 2020: How does the AA predict patient preferences?

The AA mines data such as electronic health records, medical research, social media activity (e.g., Facebook likes), and sociodemographic data to predict patient preferences with high accuracy. ]

Hubbard & Greenblum 2020: What objections do the authors address regarding the family’s role?

Two objections are addressed:

Special Relationship: The authors argue that while familial bonds matter, the AA’s epistemic accuracy outweighs the family’s claims. Dehumanization: The AA enhances autonomy rather than undermining humanity, as it ensures preferences are more accurately reflected. ]

Hubbard & Greenblum 2020: How do the authors respond to concerns about algorithmic bias?

The authors propose solutions such as independent audits, algorithmic bias specialists during development, and transparency in the AA’s decision-making process to minimize bias. ]

Hubbard & Greenblum 2020: Why do the authors recommend a gradual implementation of the AA?

A gradual implementation builds public trust. Initially, the AA can work alongside family surrogates in shared decision-making before becoming the default surrogate, with the option for patients to opt out. ]

Hubbard & Greenblum 2020: What is the role of social media data in the AA’s predictions?

Social media data, such as likes, comments, and time spent on content, are used to identify personality traits and values that correlate with specific medical preferences. ]

Hubbard & Greenblum 2020: How does the AA improve upon current surrogate decision-making practices?

The AA improves accuracy in predicting patient preferences, avoids emotional biases, reduces family stress, and ensures decisions better align with the patient’s values and autonomy. ]

Hubbard & Greenblum 2020: What broader ethical question does the AA address?

The AA raises the question of whether accuracy and objectivity in decision-making should take precedence over familial claims when respecting a patient’s autonomy. ]

Hubbard & Greenblum 2020: How does the paper respond to fears of undermining familial autonomy?

The authors argue that familial autonomy has already been rebalanced in favor of values like children’s education and women’s equality, and a similar rebalancing in favor of patient autonomy is justified. ]

Cohen 2023: What is the main argument, its stance, and key counterarguments presented in the paper?

The paper explores the ethical implications of ChatGPT and other large language models (LLMs) for bioethics. I. Glenn Cohen begins by asserting that many bioethical issues raised by ChatGPT are extensions of existing concerns about medical AI, including data ownership, privacy, bias, and consent. However, he identifies “new-ish” issues that emerge with ChatGPT, such as informed consent, medical deepfakes, oligopoly risks, environmental impact, and opportunities for democratizing knowledge.

Similarities with Medical AI:

Bioethics has grappled with issues of data representativeness and bias (e.g., inequities in training datasets) and privacy risks (e.g., reidentifying de-identified medical data). ChatGPT mirrors these issues but differs in scope and severity due to its general internet-based training and unique generative nature, which amplifies instability (e.g., “hallucinations” or inaccuracies). New-ish Issues:

Informed Consent and the Right to Know: Users might interact with AI systems without knowing they are not human, raising concerns about transparency and respect for autonomy. Medical Deepfakes: ChatGPT facilitates the creation of fabricated but convincing medical texts, misinformation, and fake images that pose risks to public trust and health outcomes. Oligopoly and Inequitable Access: The size and cost of developing LLMs consolidate power among a few corporations (e.g., OpenAI/Microsoft, Google), which may limit access and exacerbate existing inequalities. Environmental Effects: Generative AI has a high carbon footprint, raising climate justice concerns that bioethics should address. Potential Benefits:

Democratization of Knowledge: ChatGPT can empower patients by providing information to ask better questions during medical consultations. However, it also risks misinformation and misplaced confidence in AI. Counterarguments Addressed:

Bias and Trust: While ChatGPT amplifies biases present in its training data, solutions like algorithm audits and transparency can mitigate these issues. Human Agency: Concerns about dehumanization or removing interpersonal deliberation are counterbalanced by the accuracy gains and efficiency ChatGPT provides. Race Dynamics: Cohen warns that global competition for AI dominance (e.g., U.S. vs. China) may sideline ethics, but ethicists must intervene to ensure safeguards. Conclusion: Cohen concludes with four key perspectives: humility about the technology’s evolving impact, reassurance that many issues are not entirely new, excitement about bioethics addressing emerging concerns like market concentration, and focus on embedding ethicists into technical fields to keep ethics central. ]

Cohen 2023: What is the primary aim of the paper?

The paper explores the ethical implications of ChatGPT and large language models (LLMs) for bioethics, identifying both familiar issues from medical AI and “new-ish” challenges unique to ChatGPT. ]

Cohen 2023: How does ChatGPT differ from previous forms of medical AI?

ChatGPT’s training data comes from the general internet, making it more versatile but also amplifying biases and inaccuracies (“hallucinations”). It produces probabilistic outputs rather than deterministic results, creating instability and test-retest variability. ]

Cohen 2023: What are some key bioethical concerns shared by ChatGPT and medical AI?

Data Ownership and Consent: Who owns patient data, and is explicit consent required?
Bias and Representativeness: Training data often underrepresents minority groups, leading to inequitable outcomes.
Privacy: Risk of reidentifying anonymized data and improper use of personal information. ]

Cohen 2023: What are the “new-ish” bioethics issues identified for ChatGPT?

Informed Consent and Transparency: Users may not know they are interacting with AI. Medical Deepfakes: AI-generated misinformation and fake medical documents pose risks. Oligopoly and Access: LLMs consolidate power among a few corporations, limiting equitable access. Environmental Impact: LLMs have significant carbon footprints, raising climate justice concerns. ]

Cohen 2023: Why is informed consent a particular concern with ChatGPT?

ChatGPT can seamlessly interact with users, making it difficult for them to recognize they are engaging with AI rather than a human. This raises ethical questions about transparency and respect for autonomy. ]

Cohen 2023: What does Cohen mean by “medical deepfakes”?

Medical deepfakes are AI-generated texts, images, or documents that appear genuine but are entirely fabricated, posing risks such as misinformation, fraudulent medical requests, or falsified research. ]

Cohen 2023: Why does Cohen discuss oligopoly concerns in relation to ChatGPT?

Developing LLMs requires significant resources, which consolidates power among a few companies (e.g., OpenAI/Microsoft, Google). This raises concerns about access, pricing, and ethical decision-making being dictated by corporate interests. ]

Cohen 2023: What environmental issues arise with ChatGPT and other LLMs?

The development and use of LLMs have a high carbon footprint. Cohen argues that bioethics should address whether AI developers should minimize environmental harms or internalize their externalities. ]

Cohen 2023: How might ChatGPT democratize medical knowledge?

ChatGPT allows patients to access medical information and prepare better for doctor consultations. However, there is a risk of misinformation or misplaced confidence leading patients to frame issues incorrectly or avoid professional care. ]

Cohen 2023: What risks does Cohen associate with AI race dynamics?

The competitive “race to dominance” between companies (e.g., OpenAI vs. Google) or countries (e.g., U.S. vs. China) can sideline ethical considerations in favor of speed and market control. ]

Cohen 2023: Why does Cohen argue that bioethicists must act as “gadflies”?

Bioethicists must challenge the sidelining of ethics in AI development and advocate for safeguards, especially as competitive dynamics risk neglecting ethical priorities. ]

Cohen 2023: How does Cohen address concerns about bias in ChatGPT?

Cohen acknowledges bias from training data but highlights solutions like independent audits, algorithm transparency, and improvements in data diversity to mitigate such risks. ]

Cohen 2023: How does Cohen suggest addressing patient empowerment with ChatGPT?

Patients should be trained to use AI wisely, and physicians must adapt to patients arriving with AI-generated information. Liability concerns arise when misinformation impacts medical outcomes. ]

Cohen 2023: How does Cohen describe ChatGPT’s limitations?

ChatGPT generates probabilistic outputs that may produce “hallucinations” (inaccuracies). Its instability, where identical prompts yield different answers, contrasts with deterministic medical AI. ]

Cohen 2023: What are Cohen’s four concluding perspectives on ChatGPT’s impact on bioethics?

Humility: The impact is evolving too quickly to predict. Reassurance: Many issues are extensions of existing bioethics debates. New opportunities arise, Concern: such as addressing oligopoly concerns. Focus: Ethics must not be sidelined

embedded ethics in AI development is critical. ]

Cohen 2023: What is the focus of Cohen's paper?

The paper explores what ChatGPT and Large Language Models (LLMs) mean for bioethics, identifying familiar bioethics issues and highlighting "new-ish" problems that become more pressing with ChatGPT's use.

Cohen 2023: What bioethics issues raised by ChatGPT are similar to other medical AI?

Issues like data ownership, consent for data use, data representativeness and bias, and privacy. These issues exist in broader medical AI but manifest uniquely in the ChatGPT context.

Cohen 2023: What does Cohen mean by 'data ownership' in relation to ChatGPT?

Data ownership concerns focus on whether patients or physicians must consent to or be compensated for the use of their data, as medical and public data are often used to train AI models like ChatGPT.

Cohen 2023: How does Cohen frame the issue of 'data bias' for ChatGPT?

ChatGPT reflects biases from the internet, creating accuracy issues and inequities. This differs from medical AI, which often uses more reliable, curated datasets, but has its own biases (e.g., racial underrepresentation).

Cohen 2023: What privacy risks does Cohen discuss in the context of ChatGPT?

ChatGPT increases privacy risks, especially from accidental disclosures (e.g., leaking sensitive medical information) and AI's ability to infer health data from non-health information like social media or purchasing habits.

Cohen 2023: What are the 'new-ish' bioethics issues Cohen identifies with ChatGPT?

Issues include: 1) informed consent and knowing when one interacts with AI, 2) medical deepfakes, 3) oligopoly and inequitable access to foundational models, and 4) environmental effects.

Cohen 2023: What does Cohen say about the risk of 'medical deepfakes'?

ChatGPT can be used to create false medical documentation or misinformation (e.g., fake studies or radiological images). It amplifies the risk of medical distrust and deliberate misinformation.

Cohen 2023: How does Cohen describe the issue of oligopoly in AI?

The dominance of companies like OpenAI/Microsoft and Google creates barriers to entry, consolidates power, and risks ethical issues like unfair pricing, inequitable access, and data privacy failures.

Cohen 2023: What ethical concerns arise from the 'race to dominance' in AI?

Competition between companies and nations (e.g., U.S. vs. China) risks sidelining ethics, as slowing development to address ethical concerns might be seen as losing the race.

Cohen 2023: What environmental concerns does Cohen highlight with generative AI?

Generative AI like ChatGPT has significant environmental costs (e.g., carbon footprint). Bioethics should question whether there are obligations to minimize environmental harm when using AI.

Cohen 2023: How does Cohen view the democratization of access to knowledge through ChatGPT?

ChatGPT could empower patients by helping them gather medical information and ask better questions during visits. However, there are risks of misinformation and patients bypassing physicians altogether.

Cohen 2023: What are foundational models, and why do they pose ethical challenges?

Foundational models are trained on broad data and adapted for specific tasks. Their scale consolidates power, risks homogenization of problematic biases, and limits ethical oversight and competition.

Cohen 2023: What concerns arise about 'AI homogenization' in foundational models?

Homogenization risks models converging on ethically flawed practices due to biases in their training data, which could negatively impact medical decision-making.

Cohen 2023: How might AI models like ChatGPT exacerbate digital divides?

Large foundational models may be inaccessible to underserved populations due to pricing or infrastructure barriers, furthering health inequities and digital divides.

Cohen 2023: What is the 'right to know' issue Cohen raises regarding AI use?

Cohen argues that patients have a right to know when they are interacting with an AI rather than a human, especially in medical contexts, to ensure transparency and trust.