Lowest price from $10 per page Start saving now →
Annotated Bibliography

Annotated Bibliography on Ethical AI in Healthcare

Annotated Bibliography Assignment: Ethical Use of Artificial Intelligence in Healthcare

Institutional model: New York City College of Technology, City University of New York (CUNY)

Course: PHIL 2203ID, Health Care Ethics D050

Assessment: Case Project, Annotated Bibliography

Model study period: Fall 2026 adaptation of a published Fall 2023 course assignment

Assessment format: Individual research-preparation submission supporting a group case report

Suggested length: Approximately 1,600–2,050 words, excluding the reference details

Referencing: Chicago Manual of Style, Author-Date

Institutional Context and Assignment Purpose

New York City College of Technology’s PHIL 2203ID Health Care Ethics D050 course used an annotated bibliography as a preparatory assessment within a group case project. The published assignment required at least five peer-reviewed articles, alphabetised entries and explicit analysis of each source’s thesis or conclusions, arguments or evidence, and relevance to the ethical problem. The instructor identified Chicago Author-Date as the preferred citation system. The present Fall 2026 model retains that distinctive structure and adds a defined case scenario, a search record and a short comparative synthesis. It is an original teaching model, not an official current CUNY course handout.

The purpose of the assessment is to build a reliable evidence base before students write a case report. Each entry must show what the author claims, what evidence or reasoning supports that claim, and how the source helps resolve the ethical problem. A paragraph that only repeats an abstract does not meet the task.

Case Scenario: Algorithmic Allocation of Care Management

A hypothetical hospital network is considering a commercial algorithm that predicts each patient’s future healthcare costs. Managers propose using the score to identify patients who should receive intensive care-management support, including additional nurse contact, medication coordination and follow-up appointments. The vendor argues that the system can identify people likely to need expensive services and help the network use limited resources efficiently.

The hospital’s ethics committee has raised concerns about whether predicted spending is a fair measure of clinical need. A published study by Obermeyer and colleagues found that a widely used population-health algorithm underestimated the health needs of Black patients because it used healthcare costs as a proxy for illness. The case therefore asks students to examine the ethical consequences of proxy measures, unequal access to care, biased data, explainability and accountability.

Central research question: Should a hospital use an algorithm that predicts future healthcare costs to identify patients for intensive care-management support, and what ethical safeguards should govern its use?

Learning Outcomes

Successful work will demonstrate that the student can:

  • Locate and select peer-reviewed literature relevant to a defined healthcare ethics case.
  • Distinguish empirical research from normative or philosophical argument.
  • State an article’s thesis or conclusions accurately and in the student’s own words.
  • Explain how an author supports a claim through evidence, results or ethical reasoning.
  • Evaluate the source’s limitations, relevance and contribution to the case.
  • Compare sources that agree, disagree or address different parts of the ethical problem.
  • Use Chicago Author-Date referencing consistently.

Assignment Instructions

Prepare an annotated bibliography containing at least five peer-reviewed academic articles relevant to the case. Entries must be arranged alphabetically by the first author’s surname. Each entry must contain a complete bibliographic reference followed by the student’s own annotation. Each entry should be at least half a page of single-spaced text and will usually require approximately 200–250 words. The complete submission will normally contain 1,600–2,050 words, including the introduction, search record and synthesis.

Part A: Research Focus and Search Record

  1. Use the central research question or refine it into a specific question about algorithmic bias, explainability, fairness, patient autonomy or accountability.
  2. Search academic databases such as PubMed, CINAHL, Scopus, Web of Science or the university library catalogue. Google Scholar may help identify material, but verify each item through the journal or a recognised scholarly database.
  3. Use combinations of terms such as “artificial intelligence” AND healthcare AND ethics; “clinical decision support” AND algorithmic bias; healthcare AND explainability; algorithmic fairness AND patient care; and resource allocation AND health equity.
  4. Prioritise peer-reviewed work published from 2018 to 2026. Include foundational research when it directly informs the case, but explain why it remains relevant.
  5. Record each database, search date, principal search terms, filters, approximate number of results screened and reasons for selecting the final sources.

Part B: Select and Classify the Sources

  • Use a minimum of five peer-reviewed articles from relevant academic disciplines, including healthcare ethics, philosophy, medicine, health policy, medical informatics or public health.
  • Include a meaningful range of source types. At least one source must report empirical research or a data-based case study, and at least one must present a normative or philosophical argument.
  • Review articles and scoping reviews may be included, but identify them accurately. Do not describe a review as a primary empirical study.
  • Classify each source as empirical, normative, systematic review, scoping review or another clearly identified type. Explain the classification briefly where the type may be unclear.
  • Do not count blogs, general news stories, vendor marketing material or AI-generated summaries toward the five-source minimum.

Part C: Write Each Annotation

Each annotation must explicitly label the following three components. The labels must appear in the submitted work.

  1. Thesis or Conclusions: For a normative or philosophical source, state the author’s central thesis as a concise, specific claim. For an empirical source, state the study’s principal conclusion in the present tense.
  2. Arguments or Evidence/Results: Explain the reasoning used to support a normative thesis or the method, data and findings that support an empirical conclusion. State empirical observations in the past tense. Include relevant details such as the study design, population, measures or analytical approach when the source provides them.
  3. Relevance to the Ethical Problem: Identify the particular part of the case that the source helps address. Explain whether it supports, qualifies or challenges a proposed response. State how the source may be used in the later case report.

Include a critical comment on the source’s strengths and limitations within the relevant labelled component. Assess the source’s authority, method, scope, assumptions, evidence quality and applicability. Avoid generic praise such as “the article is credible” unless the annotation explains why.

Part D: Comparative Synthesis

After the five or more entries, write a 250–300-word synthesis identifying the main areas of agreement and disagreement. Explain whether the literature supports deployment, conditional deployment or non-deployment of the proposed algorithm. Identify evidence gaps that the later case report must address. Do not write five disconnected summaries and call them a synthesis.

Part E: Working Position

End with a short provisional position of approximately 100–150 words. State the ethical position that appears best supported by the sources and identify one issue that requires further research. The position may change as the group develops its case report.

Search and Source-Selection Standards

Criterion Required standard Reason
Peer review Five or more peer-reviewed academic articles Ensures the bibliography rests on scholarly work rather than general commentary or marketing.
Topic fit Each source must address at least one part of the research question Prevents sources from being included merely because they mention artificial intelligence.
Source mix At least one empirical or data-based source and one normative or philosophical source Connects evidence about real-world outcomes with arguments about what healthcare organisations ought to do.
Publication date Prioritise 2018–2026 publications Reflects recent development in clinical AI while retaining relevant foundational findings.
Critical appraisal Identify method, evidence, assumptions, limitations and applicability Demonstrates evaluation rather than abstract-level description.
Traceability Provide DOI or stable publisher/database links where available Allows the reader to locate and verify each source.

Use a source only after checking its full text or a reliable full-text record. An abstract can help determine relevance, but it is not enough to evaluate study limitations or the full argument. Verify author names, publication year, journal title, volume, issue, page range or article number, and DOI against the publisher or database record.

Required Document Structure

  1. Title page with student name, course, section, instructor and submission date.
  2. Introduction of approximately 150–200 words, stating the case topic, research question and scope.
  3. Search record of approximately 100–150 words, identifying databases, search terms, date limits and selection approach.
  4. Annotated bibliography with at least five alphabetised entries, each at least half a page and normally 200–250 words.
  5. Comparative synthesis of approximately 250–300 words.
  6. Provisional ethical position of approximately 100–150 words.
  7. Reference details in Chicago Author-Date format.

Presentation: Use a readable 12-point font, one-inch margins and single spacing within each annotation. Leave a clear blank line between entries. Use hanging indents for bibliographic references. The published City Tech model specifies that each of its five entries should be at least half a page and single-spaced. The introduction, search record and synthesis are additions to this adapted model.

Submission Requirements

  • Submit one Word document or accessible PDF through the course learning-management system.
  • Use the course’s assigned case and research question. Do not substitute a different case without instructor approval.
  • Include at least five peer-reviewed sources and all three required labels in every annotation.
  • Arrange entries alphabetically by the surname of the first author.
  • Use Chicago Author-Date consistently. The instructor’s published course page identifies this as the preferred system.
  • Use the student’s own language. Do not copy abstracts, database summaries, other students’ work or AI-generated annotations.
  • Check every DOI and bibliographic detail before submission.

Academic integrity and AI: The published Fall 2023 course instructions state that copied material from AI sources is treated as plagiarism. For this model, students should not submit AI-generated annotations as their own work. If the current instructor permits limited AI support for search planning or language editing, students must follow the course disclosure rules and independently verify every source and claim. The current course policy takes precedence.

Assessment Rubric

The published City Tech page provides detailed annotation components but does not publish a point-weighted rubric for this assignment. The following 100-point rubric is an original model for a centre or instructor to adapt; it is not an official CUNY marking scheme.

Criterion Weight Excellent performance Insufficient performance
Source quality and relevance 20 Selects at least five peer-reviewed sources with a clear, defensible relationship to the case. Includes an appropriate range of empirical and normative material. Uses too few scholarly sources, includes weakly related material or relies on non-academic sources.
Thesis or conclusions 15 States each source’s thesis or main conclusion accurately, specifically and in the correct form for its source type. Misstates the author’s position, gives only a topic label or confuses a study question with its conclusion.
Arguments or evidence/results 20 Explains how reasoning or results support the claim. Identifies relevant methods and findings and distinguishes empirical from normative reasoning. Repeats the abstract, omits evidence or fails to connect evidence with the stated conclusion.
Relevance to the ethical case 20 Explains precisely how each source informs the hospital’s decision, identifies the ethical issue addressed and states how the source may be used in the case report. Offers generic statements of relevance without connecting the source to the case or research question.
Critical evaluation and synthesis 15 Identifies significant limitations and compares points of agreement, disagreement and evidence gaps across sources. Provides disconnected summaries, overlooks limitations or makes unsupported comparisons.
Documentation, organisation and academic integrity 10 Uses accurate Chicago Author-Date citations, alphabetical order, clear labels and original prose. Contains incomplete references, inconsistent style, missing labels or unattributed copied material.
Total 100 Model assessment scale

Sample Annotated Bibliography Entries

The following entries illustrate how a student could classify and evaluate the five sources. They are model examples for study and assessment design. Students must read the full articles and write their own annotations rather than submit these examples as original work.

Ghassemi, Oakden-Rayner and Beam (2021)

Bibliographic reference: Ghassemi, M., Oakden-Rayner, L. and Beam, A.L. (2021) ‘The false hope of current approaches to explainable artificial intelligence in health care’, The Lancet Digital Health, 3(11), pp. e745–e750. https://doi.org/10.1016/S2589-7500(21)00208-9.

Source type: Critical review and normative analysis.

Thesis or Conclusions: Ghassemi, Oakden-Rayner and Beam argue that current explainable-AI techniques can create misplaced confidence in clinical decision systems because an explanation may not faithfully represent how a model reaches its output. Their central concern is that apparent transparency should not be treated as proof that a system is reliable or clinically safe.

Arguments or Evidence/Results: The authors examine the limits of post-hoc explanation methods and explain why simplified explanations may mislead clinicians about model behaviour. Their article is a critical review rather than a trial of a single hospital algorithm, so it does not establish the frequency of harm in a defined patient population. Its strength lies in identifying a technical and ethical problem that can be overlooked when a vendor presents a readable explanation as sufficient accountability.

Relevance to the Ethical Problem: The hospital should not accept an explanation screen as the only safeguard for its patient-priority score. The source supports a requirement for independent validation, clinician oversight and a route for challenging decisions. It also helps distinguish technical explainability from the wider question of whether the model uses an ethically acceptable measure of patient need.

Obermeyer, Powers, Vogeli and Mullainathan (2019)

Bibliographic reference: Obermeyer, Z., Powers, B., Vogeli, C. and Mullainathan, S. (2019) ‘Dissecting racial bias in an algorithm used to manage the health of populations’, Science, 366(6464), pp. 447–453. https://doi.org/10.1126/science.aax2342.

Source type: Empirical research study.

Thesis or Conclusions: The study concludes that a widely used population-health algorithm produced racial bias because it predicted healthcare costs rather than illness or clinical need. The authors show that a measure can appear operationally useful while systematically understating the needs of a group that historically receives less healthcare spending.

Arguments or Evidence/Results: The researchers analysed algorithmic risk scores and health information for patients enrolled in a large health system. At the same risk score, Black patients were considerably sicker than White patients. The authors estimated that correcting the proxy would increase the proportion of Black patients receiving additional help from 17.7% to 46.5%. The study connects the observed disparity to unequal spending, showing why cost is an unsafe substitute for health need.

Relevance to the Ethical Problem: This study directly informs the proposed hospital case because the hypothetical vendor also uses predicted costs to allocate support. It gives the student a concrete empirical basis for questioning the model’s target variable. The study does not prove that every cost-based algorithm will produce the same disparity, but it shows why the hospital must test outcomes across patient groups and validate the clinical meaning of its prediction target before deployment.

Freyer, Groß and Lipprandt (2024)

Bibliographic reference: Freyer, N., Groß, D. and Lipprandt, M. (2024) ‘The ethical requirement of explainability for AI-DSS in healthcare: a systematic review of reasons’, BMC Medical Ethics, 25(1), article 104. https://doi.org/10.1186/s12910-024-01103-2.

Source type: Systematic review of ethical arguments.

Thesis or Conclusions: The review concludes that there is no settled agreement on whether healthcare AI decision-support systems should always be explainable or which level of explanation is ethically necessary. The authors show that the debate includes competing positions rather than a single accepted rule.

Arguments or Evidence/Results: The authors searched PubMed, BASE and Scopus for peer-reviewed publications from 2016 to 2024. From 1,662 records, they included 44 documents after eligibility screening. Seventeen records supported explainable methods or models, while 27 argued against a general requirement. Fourteen supported context-dependent levels of explainability, whereas 30 favoured standards that do not vary by context. These findings make the article useful for showing that explainability requirements remain contested.

Relevance to the Ethical Problem: The source helps the student avoid a simplistic conclusion that every algorithm must disclose every technical detail in the same way. The hospital should instead identify which people need to understand which parts of the decision, including clinicians, patients, auditors and regulators. The review is directly relevant to accountability, but it does not itself demonstrate that one particular explanation method improves patient outcomes.

Ratti, Morrison and Jakab (2025)

Bibliographic reference: Ratti, E., Morrison, M. and Jakab, I. (2025) ‘Ethical and social considerations of applying artificial intelligence in healthcare: a two-pronged scoping review’, BMC Medical Ethics, 26, article 68. https://doi.org/10.1186/s12910-025-01198-1.

Source type: Two-pronged scoping review.

Thesis or Conclusions: Ratti, Morrison and Jakab argue that general ethical principles alone are unlikely to address the specific problems created when AI tools move from abstract discussion into real healthcare settings. Their conclusion favours analysis that takes account of the application, affected groups and conditions of deployment.

Arguments or Evidence/Results: The authors used two review strands. The first examined recent academic literature from 2021 to 2023. The second examined earlier systematic and scoping reviews from 2014 to 2024 to identify themes that recur over time. The design allows the review to consider new developments alongside continuing ethical concerns. The authors found substantial variation across healthcare fields, audiences and ethical issues, which makes broad principles difficult to apply without context.

Relevance to the Ethical Problem: The article supports a case analysis that examines the hospital’s actual allocation process rather than merely listing principles such as fairness and transparency. It directs attention to who receives extra services, who is excluded, how clinicians use the score and whether patients can question the result. Its broad scope is useful for framing the problem, although the student should pair it with empirical evidence about the specific algorithm under review.

Vandersluis and Savulescu (2024)

Bibliographic reference: Vandersluis, R. and Savulescu, J. (2024) ‘The selective deployment of AI in healthcare: an ethical algorithm for algorithms’, Bioethics, 38(5), pp. 391–400. https://doi.org/10.1111/bioe.13281.

Source type: Normative ethical analysis.

Thesis or Conclusions: Vandersluis and Savulescu argue that selective deployment of a clinical algorithm can sometimes be ethically defensible when its performance is reliable for some groups but materially worse for underrepresented groups. Their argument challenges the assumption that a system must always be either deployed for everyone or withheld from everyone.

Arguments or Evidence/Results: The authors examine two clinical case studies involving breast cancer and melanoma. They weigh the harms of using a system that performs poorly for some groups against the possible harms of delaying a tool that may benefit other patients. Their proposed approach includes transparency with patients, additional services for excluded groups and urgent collection of more representative training data. The argument is normative: it evaluates what decision-makers ought to do rather than reporting a controlled trial of an allocation system.

Relevance to the Ethical Problem: The article supplies a serious counterposition to a blanket ban on the hospital’s algorithm. It requires the student to assess whether selective or conditional deployment could be justified, what support must remain available to excluded patients and how quickly the hospital must correct its evidence gaps. Its argument should be tested against the facts of the case rather than accepted as a universal permission to deploy biased systems.

Case Analysis Writing Sample

Algorithmic Cost and Clinical Need

A hospital should not use predicted healthcare costs as the sole basis for deciding which patients receive intensive care-management support. Cost records reflect previous access to services as well as illness, so a model trained on expenditure may reproduce unequal treatment rather than identify need. Obermeyer and colleagues’ study of racial bias in a population-health algorithm found that Black patients were substantially sicker than White patients at the same risk score because the algorithm used cost as a proxy for health. The researchers estimated that correcting the proxy would raise the share of Black patients receiving additional help from 17.7% to 46.5%. Those results do not establish that every algorithm will produce the same pattern, but they demonstrate why the hospital must examine the target variable before relying on its score. An independent audit should compare model performance and referral outcomes across relevant patient groups. Until that audit establishes that the model identifies clinical need fairly, the score should not determine access to care by itself.

Fairness, Benefit and Conditional Use

The decision also involves a conflict between avoiding discriminatory outcomes and making beneficial tools available when they can improve care. Vandersluis and Savulescu’s ethical analysis of selective AI deployment argues that limited deployment may sometimes be preferable to universal use when a model performs poorly for underrepresented groups. Their position is contestable because excluding a group from an algorithm can preserve unequal access if the hospital does not provide equivalent support. Ratti, Morrison and Jakab’s 2025 scoping review further indicates that general principles need to be connected to the specific circumstances of healthcare deployment. A defensible policy would therefore require the hospital to retain non-algorithmic routes to care, explain any restriction to affected patients and commit resources to improving the model’s performance. Conditional use could be considered only after independent review, monitoring and a clear process for human override.

Further Research and Critical Considerations

Explainability and Accountability

Explainability should be assessed according to the decisions that people need to make, not treated as a single technical feature. Freyer, Groß and Lipprandt’s systematic review of explainability in healthcare decision support found disagreement about whether every system requires the same form of explanation. Ratti and colleagues’ scoping review of ethical and social considerations in healthcare AI argues that broad principles need to be tested against the circumstances of real clinical use. Ghassemi, Oakden-Rayner and Beam warn that simplified explanations can create confidence without showing how a model actually behaves. The hospital should document the system’s intended use, validation results and known limitations, while patients should be told when an algorithm materially influences their care. A named clinical team must remain responsible for reviewing disputed decisions and responding to unequal outcomes. These safeguards make accountability an operational responsibility rather than a statement in a policy document.

Is Predictive Accuracy Enough?

Predictive accuracy alone does not establish that a healthcare algorithm is ethically acceptable. A high overall accuracy score can conceal poor performance for smaller patient groups or for people whose clinical information is incomplete. Students should ask whether the target variable measures the outcome the organisation actually values, whether the training data represent the people affected and whether clinicians can challenge the result. The analysis should consider at least these safeguards:

  1. Use clinical indicators of need rather than relying on prior healthcare spending as the only target.
  2. Report performance and referral rates across relevant demographic and clinical groups.
  3. Require independent validation, human review, patient-facing explanations and a route to appeal or correction.

These questions turn a general discussion of bias into an assessment of the hospital’s actual decision process. They also help students identify evidence that is missing from a source rather than treating every article as equally applicable.

Balancing Competing Ethical Claims

The strongest case report will acknowledge that algorithmic tools may improve consistency, support early intervention and help clinicians manage limited resources. A fair evaluation must also ask who benefits, who bears the risk and whether the same benefit can be provided through less harmful methods. Vandersluis and Savulescu offer a challenging argument for selective deployment, while Obermeyer and colleagues provide empirical evidence of how a seemingly neutral cost proxy can produce unequal access. The student should compare the kind of claim each source makes before weighing them together. A philosophical argument can clarify the moral reasons for a policy, but it cannot establish the local model’s performance. An empirical study can document an observed disparity, but it does not by itself decide which allocation rule is morally justified.

Conclusion

An annotated bibliography for a healthcare ethics case should make the research process visible and show how each source contributes to the decision under review. The five-source minimum provides a starting evidence base, but source quality depends on relevance, method and critical evaluation rather than quantity alone. The City Tech model is distinctive because it requires students to separate empirical conclusions from normative theses and explain how each source informs the case. Applied to AI-based care allocation, the assignment asks students to examine whether predicted spending measures patient need and whether the proposed safeguards protect people who may be poorly represented in the data. The strongest submission will connect evidence about algorithmic performance with ethical reasoning about fairness, accountability and access to care. Its provisional position should follow from the sources and remain open to revision when further evidence becomes available.

References and Learning Materials

  1. Freyer, N., Groß, D. and Lipprandt, M. (2024) ‘The ethical requirement of explainability for AI-DSS in healthcare: a systematic review of reasons’, BMC Medical Ethics, 25(1), article 104. https://doi.org/10.1186/s12910-024-01103-2.
  2. Ghassemi, M., Oakden-Rayner, L. and Beam, A.L. (2021) ‘The false hope of current approaches to explainable artificial intelligence in health care’, The Lancet Digital Health, 3(11), pp. e745–e750. https://doi.org/10.1016/S2589-7500(21)00208-9.
  3. Obermeyer, Z., Powers, B., Vogeli, C. and Mullainathan, S. (2019) ‘Dissecting racial bias in an algorithm used to manage the health of populations’, Science, 366(6464), pp. 447–453. https://doi.org/10.1126/science.aax2342.
  4. Ratti, E., Morrison, M. and Jakab, I. (2025) ‘Ethical and social considerations of applying artificial intelligence in healthcare: a two-pronged scoping review’, BMC Medical Ethics, 26, article 68. https://doi.org/10.1186/s12910-025-01198-1.
  5. Vandersluis, R. and Savulescu, J. (2024) ‘The selective deployment of AI in healthcare: an ethical algorithm for algorithms’, Bioethics, 38(5), pp. 391–400. https://doi.org/10.1111/bioe.13281.

Annotated Bibliography: Ethical Use of Artificial Intelligence in Healthcare

Referencing follows Chicago Author-Date, as the brief requires, in place of the Harvard or APA styles named in the template. Verify each DOI, page range and figure against the full text before submission.

Introduction

Hospitals now buy software that scores patients before a clinician has met them. A hypothetical network wants to use one such score, a prediction of future healthcare cost, to decide who receives intensive care management: extra nurse contact, medication coordination and follow-up visits. The vendor presents cost as a sensible stand-in for need. The ethics committee doubts it. Obermeyer et al. (2019) found that a widely used population-health algorithm underestimated the needs of Black patients because it ranked people by spending, and spending reflects access to care as much as illness. The research question follows from that finding. Should the network use a cost-prediction algorithm to select patients for care management, and which safeguards should govern its use? The bibliography draws on five peer-reviewed sources: one empirical study, two evidence reviews and two normative analyses. Together they address the proxy problem, explainability, accountability and conditional deployment. Legal compliance and vendor pricing fall outside the scope.

Search Record

Databases searched: PubMed, Scopus and the library catalogue [add search dates]. Principal terms combined “artificial intelligence” with healthcare and ethics, “clinical decision support” with algorithmic bias, explainability with healthcare, and resource allocation with health equity. Filters limited results to peer-reviewed English-language articles from 2018 to 2026, with one exception, a 2019 study retained because it anticipates the exact design of the case. Approximately [add number] records were screened by title and abstract, then [add number] by full text. Selection favoured sources that address a distinct part of the problem: the target variable, explanation, deployment strategy or governance. Each source was checked against the journal record for author names, volume, pages and DOI.

Annotated Bibliography

Freyer, Groß and Lipprandt (2024)

Freyer, Nils, Hendrik Groß, and Myriam Lipprandt. 2024. “The Ethical Requirement of Explainability for AI-DSS in Healthcare: A Systematic Review of Reasons.” BMC Medical Ethics 25 (1): article 104. https://doi.org/10.1186/s12910-024-01103-2.

Source type: Systematic review of ethical arguments.

Thesis or Conclusions: The review concludes that no settled position exists on whether healthcare decision-support systems must always be explainable. Authors disagree about the level of explanation that ethics requires. The result is a contested debate rather than a single rule, which limits any claim that explainability alone resolves the hospital’s problem.

Arguments or Evidence/Results: The authors searched PubMed, BASE and Scopus for peer-reviewed publications from 2016 to 2024. They screened 1,662 records and included 44 documents after eligibility checks. Reviewing arguments rather than clinical outcomes gives the article breadth. It also means the findings describe what scholars claim, not what explanation methods achieve for patients. Counting positions can also overstate agreement when authors cite one another.

Relevance to the Ethical Problem: The source supports an accountability plan built around audiences. Clinicians, patients, auditors and regulators need different information about the care-management score. The case report can use the review to argue against a blanket disclosure rule and for tiered explanation. It offers no evidence that one explanation method improves outcomes.

Ghassemi, Oakden-Rayner and Beam (2021)

Ghassemi, Marzyeh, Luke Oakden-Rayner, and Andrew L. Beam. 2021. “The False Hope of Current Approaches to Explainable Artificial Intelligence in Health Care.” The Lancet Digital Health 3 (11): e745–50. https://doi.org/10.1016/S2589-7500(21)00208-9.

Source type: Critical review and normative analysis.

Thesis or Conclusions: The authors argue that current explainable-AI techniques can create misplaced confidence because an explanation may not faithfully describe how a model reached its output. A readable explanation does not prove that a system is reliable or safe. Validation must carry that burden instead.

Arguments or Evidence/Results: The article examines the limits of post-hoc explanation methods and shows how a simplified account can mislead clinicians about model behaviour. It tests no single hospital algorithm, so it cannot estimate how often such misreading causes harm. Its strength lies in naming a failure that vendors rarely disclose. A reader should treat it as a warning grounded in technical argument, not as measured evidence.

Relevance to the Ethical Problem: A vendor dashboard that explains each patient’s score cannot count as the hospital’s main safeguard. The source supports independent validation, clinician oversight and a route to contest decisions. It also separates technical transparency from the harder question of whether cost is an acceptable measure of need.

Obermeyer, Powers, Vogeli and Mullainathan (2019)

Obermeyer, Ziad, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. 2019. “Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations.” Science 366 (6464): 447–53. https://doi.org/10.1126/science.aax2342.

Source type: Empirical research study.

Thesis or Conclusions: The study concludes that a widely used population-health algorithm produced racial bias because it predicted healthcare costs rather than illness. A measure can look operationally useful while understating the needs of a group that historically receives less spending.

Arguments or Evidence/Results: The researchers analysed algorithmic risk scores alongside health data for patients in a large health system. At the same risk score, Black patients were considerably sicker than White patients. The authors estimated that correcting the proxy would raise the share of Black patients receiving additional help from 17.7% to 46.5%. The design is strong because it compares the score against direct measures of health. Its limit is scope: one algorithm in one system tells the reader little about every cost-based tool.

Relevance to the Ethical Problem: The hospital’s vendor uses the same target variable, so this study speaks to the core of the case. It gives the report an empirical basis for questioning predicted cost as a proxy for need. It also justifies a pre-deployment audit that compares referral rates and illness burden across patient groups.

Ratti, Morrison and Jakab (2025)

Ratti, Emanuele, Mark Morrison, and Ivett Jakab. 2025. “Ethical and Social Considerations of Applying Artificial Intelligence in Healthcare: A Two-Pronged Scoping Review.” BMC Medical Ethics 26: article 68. https://doi.org/10.1186/s12910-025-01198-1.

Source type: Two-pronged scoping review.

Thesis or Conclusions: The authors argue that general ethical principles cannot, alone, settle the problems that arise when AI enters real healthcare settings. Analysis should account for the application, the affected groups and the conditions of deployment.

Arguments or Evidence/Results: One strand examined recent academic literature from 2021 to 2023. A second strand examined earlier systematic and scoping reviews from 2014 to 2024 to identify themes that recur. That structure lets the review compare new concerns with long-standing ones. The authors found wide variation across clinical fields, audiences and ethical issues. A scoping review maps topics without grading the quality of the studies it includes, so its conclusions carry less evidential weight than a systematic appraisal.

Relevance to the Ethical Problem: The review supports an analysis of the hospital’s actual process. Who receives extra services, who is excluded, how clinicians use the score and whether patients can question it all matter more than a list of principles. Its breadth helps frame the report. The student should pair it with specific evidence on cost-based allocation.

Vandersluis and Savulescu (2024)

Vandersluis, Robert, and Julian Savulescu. 2024. “The Selective Deployment of AI in Healthcare: An Ethical Algorithm for Algorithms.” Bioethics 38 (5): 391–400. https://doi.org/10.1111/bioe.13281.

Source type: Normative ethical analysis.

Thesis or Conclusions: Vandersluis and Savulescu argue that selective deployment of a clinical algorithm can be ethically defensible when it performs reliably for some groups and materially worse for underrepresented ones. Their position rejects the assumption that a system must either serve everyone or no one.

Arguments or Evidence/Results: The authors work through two clinical cases, breast cancer and melanoma. They weigh the harm of using a tool that performs poorly for some patients against the harm of delaying a tool that helps others. Their approach adds transparency with patients, additional services for excluded groups and urgent collection of more representative data. The argument is normative, so it shows what decision-makers ought to do without testing it on an allocation system. Its cases involve diagnosis, which differs from distributing scarce support.

Relevance to the Ethical Problem: The article offers the strongest counterposition to a flat ban on the hospital’s algorithm. It forces the report to ask whether conditional use could be justified, what care must remain for excluded patients and how fast the evidence gap must close. The argument needs testing against the case facts, because it does not license deploying a biased system.

Comparative Synthesis

The sources agree that a model’s apparent sophistication offers no assurance of fairness. Obermeyer et al. (2019) supply the clearest evidence: the target variable, not the code, produced the disparity. Ghassemi, Oakden-Rayner and Beam (2021) extend the point to explanation, arguing that a persuasive display can hide the same flaw. Ratti, Morrison and Jakab (2025) reach a compatible conclusion from a different direction, since general principles fail without attention to the setting where a tool operates.

Disagreement appears over remedy. Freyer, Groß and Lipprandt (2024) show that scholars divide on how much explanation ethics demands, which weakens any argument that transparency is the primary safeguard. Vandersluis and Savulescu (2024) disagree with a strict non-deployment stance, proposing that selective use may serve patients better than withholding a tool. Their position conflicts with the audit logic implied by Obermeyer et al., though both can coexist if selective use depends on verified performance.

Taken together, the literature supports conditional deployment at most. The cost target should not drive referrals without evidence that it tracks clinical need. Independent validation, group-level reporting, human review and an appeal route form the minimum conditions. Evidence gaps remain. No source tests a cost-based allocation tool inside a hospital like the one described, and none measures how clinicians act on scores. The empirical work covers one health system, and the normative work relies on diagnostic examples. The case report must address both gaps directly.

Working Position

The sources best support non-reliance on predicted cost as the sole selection criterion, with conditional use after independent review. The hospital should replace or supplement the cost target with clinical indicators of need, report referral rates across patient groups and keep a human route to care for anyone the score overlooks. Explanation tools may assist clinicians but cannot substitute for validation. Selective deployment remains arguable, yet it requires equivalent support for patients outside the tool’s reliable range. One issue needs further research: whether clinicians who see the score override it in ways that reduce or reproduce disparity. The group should revise this position if local audit data contradict the published findings.

References

Freyer, Nils, Hendrik Groß, and Myriam Lipprandt. 2024. “The Ethical Requirement of Explainability for AI-DSS in Healthcare: A Systematic Review of Reasons.” BMC Medical Ethics 25 (1): article 104. https://doi.org/10.1186/s12910-024-01103-2.

Ghassemi, Marzyeh, Luke Oakden-Rayner, and Andrew L. Beam. 2021. “The False Hope of Current Approaches to Explainable Artificial Intelligence in Health Care.” The Lancet Digital Health 3 (11): e745–50. https://doi.org/10.1016/S2589-7500(21)00208-9.

Obermeyer, Ziad, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. 2019. “Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations.” Science 366 (6464): 447–53. https://doi.org/10.1126/science.aax2342.

Ratti, Emanuele, Mark Morrison, and Ivett Jakab. 2025. “Ethical and Social Considerations of Applying Artificial Intelligence in Healthcare: A Two-Pronged Scoping Review.” BMC Medical Ethics 26: article 68. https://doi.org/10.1186/s12910-025-01198-1.

Vandersluis, Robert, and Julian Savulescu. 2024. “The Selective Deployment of AI in Healthcare: An Ethical Algorithm for Algorithms.” Bioethics 38 (5): 391–400. https://doi.org/10.1111/bioe.13281.

Need a custom paper on this topic? Our experts can help.

Order now

Ready to get your paper started?

Share your instructions, pick a deadline and we will match you with a qualified writer.