Before using generative AI to develop Key Feature Problems (KFPs), educators should first define clear, measurable learning outcomes and derive the corresponding key features. Key features represent the critical decisions or actions that determine effective clinical management ( Farmer & Page, 2005; Nayer et al., 2018). Establishing these foundations ensures that AI-generated content is grounded in explicit educational intent and that each scenario targets competencies essential to clinical reasoning. Developing learning outcomes and key features in advance prevents the creation of unfocused or misaligned cases and supports validity by ensuring that each question assesses a decision point directly related to the intended outcome.

Once the initial key features are identified, AI can assist in refining and expanding them. By analyzing large datasets or educational case repositories, AI can identify additional high-yield decision points that may not be apparent through manual analysis. Drawing upon diverse clinical information allows educators to uncover patterns and associations that enhance authenticity and completeness. This process strengthens alignment between curricular objectives and the reasoning steps that differentiate expert from novice performance ( Farmer & Page, 2005; Nayer et al., 2018).

Example 1:

Learning Objective:

Demonstrate the ability to diagnose and manage acute asthma in adult patients.

Identified Key Features: 1.

Assess severity of the asthma exacerbation.

AI-Generated Case (Short Clinical Vignette):

A 30-year-old patient presents to the emergency department with shortness of breath and audible wheezing for the past two hours. The patient has a known history of asthma and seasonal allergies.

Key Feature Questions: 1.

(Write-in) What two clinical assessments are most important for determining the severity of this exacerbation?

This sequence (learning outcome → key features → case → questions) illustrates the structured logic of KFP design and specifies the item format and number of responses required, consistent with established methodology ( Farmer & Page, 2005; Nayer et al., 2018).

Example 2:

Educators who identify preliminary key features for managing acute chest pain, such as: 1.

Obtaining an appropriate history and identifying red-flag symptoms,

Large language models may reveal additional decision points, including: •

Differentiating cardiac from non-cardiac causes (for example, pulmonary embolism or aortic dissection).

These refinements help ensure that the resulting KFPs capture a broader spectrum of clinical complexity and reflect authentic decision-making challenges encountered in practice ( Farmer & Page, 2005; Nayer et al., 2018).

Note: KFPs may be presented as write-in or short-menu (SM) items. In SM formats, response options and the number of required selections must always be explicitly stated. At this stage, AI assists in improving the quality and breadth of key features, but the educator retains responsibility for selecting which AI-suggested features to include when constructing the final clinical vignette and corresponding questions.

Creating realistic and contextually grounded clinical scenarios is essential to the educational value of KFPs. Once key features have been identified and refined, generative AI can be used to construct authentic vignettes that situate these decisions within believable clinical contexts ( Berbenyuk et al., 2024; Qiu & Liu, 2025). By incorporating relevant demographic, environmental, and psychosocial details, AI helps simulate the complexity of real-world medical encounters ( Potter & Jefferies, 2024; Sardesai et al., 2024).

AI tools can also vary contextual parameters, such as disease stage, comorbidities, or resource limitations, to produce multiple versions of the same case. This contextual diversity strengthens students’ ability to transfer reasoning across scenarios and enhances case authenticity without adding to faculty workload ( Berbenyuk et al., 2024; Indran et al., 2024).

Example:

AI-Enhanced Realistic KFP Scenario (Short Vignette)

Mr. Ali K., a 58-year-old taxi driver with long-standing hypertension and type 2 diabetes, arrives at a community clinic complaining of mild chest discomfort radiating to his jaw. He reports the pain began after climbing stairs 30 minutes ago and has gradually subsided. He takes metformin and amlodipine irregularly. Vital signs: BP 160/95 mmHg, HR 88 bpm, SpO ₂ 97%, BMI 31 kg/m ². The nearest hospital is 25 km away.

Key Feature Questions: 1.

(Write-in) What initial clinical assessments are essential before deciding whether this patient can safely remain in the clinic? List up to two.

By prompting AI to integrate demographic, psychosocial, and logistic details, educators can generate scenarios that are not only clinically coherent but also contextually realistic ( Potter & Jefferies, 2024; Qiu & Liu, 2025). Such authenticity strengthens cognitive fidelity, meaning that decisions made in the scenario closely mirror real clinical reasoning, thereby enhancing learners’ engagement and readiness for practice ( Preiksaitis & Rose, 2023; Sardesai et al., 2024).

Note: While AI can enhance realism, each generated scenario must undergo expert review to verify clinical accuracy and appropriateness for the target learner level ( Farmer & Page, 2005; Nayer et al., 2018).

Generative AI can be strategically used to create multiple, pedagogically distinct versions of clinical scenarios centered on the same medical condition. This approach promotes both educational richness and psychometric robustness by exposing learners to varied but conceptually equivalent challenges ( Berbenyuk et al., 2024; Indran et al., 2024).

By varying contextual elements such as patient demographics, comorbidities, access to resources, and disease stage, AI helps educators design cases that assess the transfer of learning rather than rote recall ( Hrynchak et al., 2014). For instance, a single learning outcome on “acute coronary syndrome management” can be represented through different case variants: a young woman with atypical chest pain, an elderly diabetic with silent ischemia, or a middle-aged smoker with classic symptoms. Each variant targets the same underlying key features but tests adaptive reasoning in distinct contexts ( Farmer & Page, 2005; Nayer et al., 2018).

AI can also support psychometric balance by generating parallel cases matched on cognitive level and difficulty, aiding blueprinting and longitudinal assessment across cohorts ( Indran et al., 2024). Through controlled prompting, educators can maintain item equivalence while ensuring content freshness and reduced cueing effects. This capacity is especially useful for formative assessments, progress tests, and multi-institutional benchmarking.

Example:

Learning Objective: Manage patients presenting with myocardial infarction.

Common Key Features: 1.

Identify ischemic symptoms and risk factors.

By generating structured variants like these, AI helps educators evaluate consistency in reasoning across different contexts while maintaining construct validity. Moreover, such diversity supports inclusivity, ensuring exposure to a range of patient profiles and system-level challenges ( Mishra et al., 2024; Teferi et al., 2023).

When educators vary contextual parameters such as disease stage, comorbidities, or resource limitations, AI can produce multiple case versions to strengthen transfer of reasoning ( Table 1).

Note: Each AI-generated variant should be reviewed for alignment with curricular outcomes and calibrated for difficulty using item analysis or expert consensus ( Farmer & Page, 2005; Nayer et al., 2018).

Effective KFPs go beyond factual recall and assess a learner’s ability to analyze, synthesize, and evaluate complex clinical information at the upper levels of Bloom’s taxonomy ( Zaidi et al., 2018). Generative AI can assist educators in scaffolding these higher-order cognitive processes by helping design questions that explicitly demand interpretation, prioritization, and reasoning rather than mere recognition ( Berbenyuk et al., 2024; Indran et al., 2024).

By adjusting prompts and parameters, educators can use AI to generate versions of KFP that target specific cognitive levels, for example, distinguishing between tasks that ask students to identify key findings (lower order) and those that require them to justify management decisions or evaluate competing interventions (higher order) ( Farmer & Page, 2005; Nayer et al., 2018). This calibrated complexity enhances both formative and summative assessment design within competency-based curricula ( Jantausch et al., 2023).

AI can also suggest reasoning scaffolds such as stepwise justification prompts, conditional branching, or “what-if” variations that help learners articulate the logic behind their choices. When combined with faculty validation, these features turn KFPs into active reasoning exercises that closely resemble real-world diagnostic and management decision-making ( Araújo et al., 2024).

Example:

Learning Objective: Apply critical reasoning to prioritize diagnostic steps in a patient with acute shortness of breath.

Key Features: 1.

Interpret initial presentation and vital signs.

AI-Generated Higher-Order Question Sequence 1.

(Write-in) Based on this patient’s presentation, what is your leading differential diagnosis? List up to two.

This structure progresses from analysis to evaluation, showing how AI can scaffold increasing levels of cognitive complexity within a single clinical context.

By refining prompts to elicit reasoning, educators ensure that AI-generated KFP assess how students think, not just what they know ( Araújo et al., 2024; Jantausch et al., 2023).

Note: While AI can help generate cognitively rich content, final validation by subject-matter experts is essential to confirm that each question targets the intended cognitive level and aligns with learning outcomes ( Farmer & Page, 2005; Nayer et al., 2018).

Generative AI can accelerate the development of draft Key Feature Problems (KFPs), but educator oversight remains essential to ensure that items are constructively aligned with curricular outcomes, competency frameworks, and learner progression ( Farmer & Page, 2005; Harden et al., 1999). AI-generated content should also be contextualized to the institution’s clinical setting, patient population, and healthcare realities, enhancing authenticity and local relevance ( Berbenyuk et al., 2024; McLaughlin et al., 2019).

Item complexity must match the learner’s cognitive and experiential readiness. Early-phase students benefit from single-decision questions emphasizing recognition, while advanced learners should tackle multi-step cases demanding integration and prioritization ( Farmer & Page, 2005; Nayer et al., 2018). AI can scaffold difficulty by varying diagnostic ambiguity, patient stability, or data availability, supporting progressive learning across preclinical and clinical phases ( Berbenyuk et al., 2024; Indran et al., 2024; Tolsgaard et al., 2023).

An illustrative example of progressive complexity across learner levels focused on managing diabetic ketoacidosis (DKA) is provided in Table 2. This example demonstrates how item difficulty can be structured according to learner competence, from early recognition to advanced management and prioritization.

AI can further diversify assessment by reformatting a single clinical concept into multiple item types, such as short-answer, extended-matching, or multiple-response questions, while maintaining the same cognitive intent ( Indran et al., 2024; Javaeed, 2018). This enhances reliability and fairness by sampling reasoning across modalities and supports triangulation in programmatic assessment frameworks ( Connor et al., 2020; Fatima et al., 2024; Tolsgaard et al., 2023).

Example:

A 28-year-old man presents with sudden onset of severe shortness of breath after a long-haul flight. He is tachycardic and mildly hypoxic.

Original KFP:

What is the most likely diagnosis, and what is the next immediate investigation? ( Write-in ).

An illustrative transformation of a single vignette into multiple formats is provided in Table 3.

By aligning complexity, format, and learner stage, AI enables coherent and longitudinal assessment design that reinforces stage-appropriate competencies while maintaining curricular coherence ( Berbenyuk et al., 2024; Harden et al., 1999; Indran et al., 2024).

Note: While AI can automate scaffolding and reformatting, faculty judgment remains indispensable to verify that each item accurately represents the intended cognitive process and meets clinical and psychometric standards ( Farmer & Page, 2005; Nayer et al., 2018).

The rapid generation of KFPs by generative AI demands a structured and defensible validation process to ensure that the resulting items meet accepted standards of quality, fairness, and educational relevance. To address this need, we adapted an evidence-based framework for validating AI-generated assessment content, drawing upon widely recognized validity models from Messick (1995), Kane (2013), Downing (2002), and Cook et al. (2015).

This adapted process integrates principles of content validity, cognitive process verification, response process accuracy, internal structure coherence, and consequential validity, contextualized for AI-assisted item generation ( Farmer & Page, 2005; Nayer et al., 2018; Tolsgaard et al., 2023). It provides educators with a transparent and replicable structure for reviewing and approving AI-generated questions prior to implementation ( Table 4).

This structured approach does not replace psychometric analysis but provides a pragmatic validity chain that educators can apply before large-scale deployment. Each step contributes evidence toward construct validity, ensuring that AI-generated KFPs assess genuine clinical reasoning rather than superficial pattern recognition ( Farmer & Page, 2005; Nayer et al., 2018; Wade et al., 2012).

The overall process is visualized in Figure 1, which outlines the adapted five-step validation workflow for AI-generated assessment items.

Example Application:

Suppose AI generates a KFP on managing community-acquired pneumonia. •

Stage 1: SMEs confirm the key features (diagnosis, antibiotic choice, admission criteria) match curricular outcomes.

Note: The five-step validation process is an adaptation of established assessment validity frameworks ( Cook et al., 2015; Downing, 2002; Kane, 2013; Messick, 1995), contextualized for the use of generative AI in question development. It aims to provide a practical quality-assurance model for educators rather than propose a novel psychometric paradigm.

Effective feedback in KFPs must be decision-specific, concise, and actionable, focusing on each key feature rather than the case as a whole ( Farmer & Page, 2005; Hrynchak et al., 2014; Nayer et al., 2018). Well-designed feedback helps learners understand why a particular decision is correct and why alternatives are less appropriate. Generative AI can assist in drafting such targeted feedback rapidly, but its output must always undergo SME review to verify clinical accuracy, tone, and contextual sensitivity ( Farmer & Page, 2005; Nayer et al., 2018; Zhang et al., 2025).

Generative AI can be prompted to produce feedback at different levels of granularity, as summarized in Figure 2, which illustrates how prompts can generate decision-specific feedback messages tailored to each key feature.

Per-key feature rationales explaining both correct and incorrect choices, particularly valuable for short-menu (SM) items where learners must select a specified number of responses ( Farmer & Page, 2005; Nayer et al., 2018). •

Tiered feedback messages for correct, partially correct, and incorrect responses that identify common reasoning errors and suggest appropriate next steps in decision-making ( Burner et al., 2025; Lee & Moore, 2024).

Timing also matters. For formative KFP, immediate, key feature level feedback enhances learning efficiency and self-regulation ( Burner et al., 2025; Lee & Moore, 2024). For summative KFP, delayed or aggregate feedback preserves item security while still supporting post-exam reflection ( Farmer & Page, 2005; Nayer et al., 2018).

Despite its efficiency, AI-generated feedback may lack nuance and contextual sensitivity in complex or atypical cases, which highlights the need for human oversight, particularly in edge scenarios ( Burner et al., 2025). SMEs should verify that AI feedback accurately targets the intended reasoning process and does not introduce misleading or unsafe guidance.

Illustrative Example (Write-in + Short-Menu with Feedback)

Scenario (abridged): A 28-year-old presents with fever, headache, and neck stiffness.

KF-Q1 (write-in): What is the most likely diagnosis? •

Correct feedback: “Bacterial meningitis is most consistent with fever and neck stiffness; treat urgently with empiric antibiotics.”

KF-Q2 (SM; select 2): Which initial diagnostic investigations are required? •

Lumbar puncture

Continuous improvement of AI-generated KFPs depend on systematic analysis of response data and psychometric evidence. Educators should employ both quantitative and qualitative data to identify items that require revision, strengthening validity, reliability, and alignment with learning outcomes ( Farmer & Page, 2005; Kim et al., 2022; Nayer et al., 2018; Tolsgaard et al., 2023).

Data sources include item statistics from pilot tests (difficulty, discrimination, non-functioning options) and learner feedback on clarity and realism. When analyzed together, these indicators reveal whether each KFP effectively assesses the intended decision point ( Almansour & Alfhaid, 2024; Tolsgaard et al., 2023). For example, very low discrimination may indicate that the question does not differentiate between competent and struggling learners, while an unexpectedly high success rate may suggest over-cueing or insufficient cognitive demand ( Kim et al., 2022).

AI can support this process by generating revised item versions based on educator feedback or psychometric findings. Prompted appropriately, the model can reword stems for clarity, modify distractors for plausibility, or adjust contextual parameters to correct misalignment ( Berbenyuk et al., 2024; Indran et al., 2024). These revisions must then be revalidated by SMEs before reuse.

Illustrative Example (KFP Improvement via Data Review)

Original AI-Generated KFP (Pre-Revision)

Scenario: A 35-year-old patient presents with pleuritic chest pain and mild dyspnea.

Question (Write-in): What is the most likely diagnosis?

Issue: Student response data showed poor discrimination (r = 0.05); many misidentified pneumonia or pneumothorax.

Data Insight: Qualitative feedback revealed insufficient contextual clues to differentiate pulmonary embolism from other causes of chest pain.

Revised KFP (Post-Review)

Scenario: A 35-year-old female on oral contraceptives presents with sudden pleuritic chest pain and mild dyspnea after a 10-hour flight.

Question (Write-in): What is the most likely diagnosis?

Rationale: Added risk factor and temporal trigger clarified the intended decision focus (PE) without making the question easier. SME review confirmed improved alignment and realism.

This example demonstrates how data-driven iteration enhances clarity, construct validity, and clinical authenticity ( Farmer & Page, 2005; Nayer et al., 2018). The implementation steps for this iterative process are illustrated in Figure 3, which presents the data-driven KFP improvement workflow.

Implementation Steps for Data-Driven KFP Improvement 1.

Collect data from pilot or formative use (difficulty index, discrimination, and student feedback).

Note: This process focuses solely on psychometric and content improvement. Considerations of inclusivity and bias mitigation are addressed separately (see Tip 10).

Equity, diversity, and inclusion (EDI) are essential principles in assessment design. In the context of Key Feature Problems (KFP), EDI ensures that all learners engage with clinically authentic yet culturally fair scenarios that reflect the diversity of real-world patient populations ( Kim et al., 2024; Tolsgaard et al., 2023). When generative AI is used to create KFP, additional vigilance is required to prevent the unintentional introduction or amplification of bias in case content, patient descriptors, or reasoning expectations ( Kim et al., 2024; Rodman et al., 2024).

Identify and Mitigate Potential Bias in AI Outputs

AI models can inadvertently reproduce societal or dataset biases, leading to stereotypical patient profiles, imbalanced demographic representation, or culturally narrow assumptions ( Kim et al., 2024).

To prevent this, educators should: •

Audit AI-generated cases for demographic balance across age, gender, ethnicity, and socioeconomic background.

Promote Inclusive Case Representation

EDI-aligned KFP should expose learners to the breadth of human variation and social determinants that influence diagnosis and management. AI can assist by generating case variants that represent different demographic or psychosocial contexts while maintaining equivalent cognitive challenge ( Berbenyuk et al., 2024; Kim et al., 2024).

For example, a case on myocardial infarction can be rendered across: •

A younger female with atypical presentation,

Such diversity fosters equitable preparedness and reduces bias in clinical decision-making ( Kim et al., 2024; Rodman et al., 2024; Tolsgaard et al., 2023).

To operationalize inclusivity in AI-assisted KFP design, educators should follow a structured EDI review sequence illustrated in Figure 4, which outlines the bias-mitigation checkpoints during AI generation and validation.

Integrate EDI Checks Into the KFP Workflow

To operationalize inclusivity in AI-assisted KFP design: 1.

Set EDI parameters before prompting AI, specifying desired demographic distribution and case diversity.

Original AI Output:

A 45-year-old South Asian man with poorly controlled diabetes presents with chest pain after eating a heavy meal.

Issue: The AI model consistently associated “South Asian” with “diabetes,” reinforcing a stereotype without instructional purpose.

Revised Prompt and Case:

Generate a case of a 45-year-old adult presenting with chest pain unrelated to ethnicity. Include relevant lifestyle and risk factors.

Result: The AI produced a balanced scenario highlighting modifiable risks (sedentary lifestyle, hypertension) rather than cultural identity, aligning better with fairness and learning objectives.

Note: EDI alignment is not a single review step but a continuous design principle that parallels psychometric validation. Each AI-generated KFP should undergo both content and equity review before use to ensure fairness, representation, and clinical authenticity ( Kim et al., 2024; Rodman et al., 2024).

Use existing pre-trained models rather than developing new ones. Concentrate faculty effort on prompt design, SME review, and validity checks so AI output meets curricular and clinical standards ( Berbenyuk et al., 2024; Kovari, 2024; Tolsgaard et al., 2023).

Document for auditability. For each item, record the tool/model and version used, prompt template (and key settings), SME comments, and validation outcomes (see Tip 6). This enables reproducibility and external review by faculty and accreditors ( Kovari, 2024; Rodman et al., 2024; Tolsgaard et al., 2023).

Protect boundaries. Never upload identifiable learner or patient data to external tools; clarify authorship when AI contributes text or drafts; require human sign-off on all exam materials ( Kovari, 2024; Tolsgaard et al., 2023).

Build capacity. Provide ongoing faculty development in responsible prompting, data stewardship, and bias awareness so that AI augments educational expertise rather than replacing it ( Berbenyuk et al., 2024; Indran et al., 2024; Kovari, 2024; Tolsgaard et al., 2023).

Be pragmatic about clinical data. Until secure educational data environments mature, prefer synthetic or de-identified sources and simulated EHR interfaces; full interoperability with real systems is generally not feasible yet ( Blau et al., 2024; Razmi, 2024; Tolsgaard et al., 2023).

Quick checklist (for your item bank record): •

Tool/model + version

This keeps AI use ethical, transparent, and sustainable while preserving assessment integrity.

A consolidated overview of all ten tips summarizing their purposes, recommended educator actions, and common pitfalls is presented in Table 5.

•

Start with outcomes and key features; keep AI inside those boundaries.

This paper is intended as a practice-oriented guide rather than an empirical or psychometric validation study. Its focus is on the educational design and responsible use of generative artificial intelligence (AI) to assist in developing Key Feature Problems (KFP) within undergraduate (UME) and postgraduate medical education (PGME) contexts. The recommendations emphasize conceptual alignment, item quality, and governance rather than quantitative analysis of reliability, validity coefficients, or statistical performance metrics.

The scope of guidance also excludes blueprinting logistics, standard setting, and scoring procedures, which vary across institutions and are beyond the current discussion. While examples provided illustrate typical clinical reasoning domains, they are intended to demonstrate design principles rather than to serve as validated assessment items.

Implementation feasibility may differ depending on institutional infrastructure, data governance maturity, and faculty readiness. The principles described should therefore be adapted to local curricular frameworks, regulatory requirements, and available AI tools. Educators should interpret these tips as a foundation for responsible innovation and not as a prescriptive or exhaustive model for KFP development.