Artificial intelligence (AI) is a transformative technology that enables systems to generate outputs, such as content, recommendations, or decisions, based on human-defined objectives [1,2]. A key branch, generative AI (GenAI), produces human-like media (text, images, videos, and code) in response to prompts [3]. At the forefront are large language models (LLMs) such as ChatGPT (OpenAI), Microsoft Copilot, and Gemini (Google), which use deep learning and vast datasets to understand context, generate coherent responses, and adapt tone and style [4-8]. These models improve continuously through reinforcement learning with human feedback [9-11].

In higher education, LLMs have rapidly gained attention among academics and students [8,12-14], supporting assessment design, automated marking, and curriculum development [14-17], improving feedback timeliness and reducing grading errors [14,18,19]. They also help students refine language, generate ideas, and start research inquiries [20]. Overall, recent evidence in medical education highlights the rapid diversification of GenAI applications across learning resources, instructional methods, and assessment practices, offering tangible opportunities in areas such as documentation support, simulation-based training, and personalized learning [21-23].

Despite the benefits, the integration of GenAI technologies into higher education can present significant ethical challenges [24]. These include ChatGPT’s use of inaccurate content (including fictitious reference material) [25], concerns regarding human teacher replacement [11], and negative repercussions on students’ critical thinking and problem-solving skills [26]. Most significantly, the use of ChatGPT can pose a threat to academic integrity and ethics [27]. A recent 2024 scoping review has highlighted that traditional assessment methods do not operate effectively in GenAI-facilitated learning environments, prompting the need for innovative and refocused assessment designs that foster career-driven competencies and lifelong learning skills [28]. The ethical framework surrounding students’ use of AI in educational assessments primarily involves transparency, accountability, and reliability [29]. Transparency requires that students acknowledge the use of AI appropriately in their submissions. Students must also be accountable for the content they submit, as “outputs of AI tools can include biased, inaccurate, or incorrect content that users should be aware of” [29]. In some instances, GenAI can fabricate information, which raises the question of the reliability of the outputs [7].

The accessibility of AI tools and the difficulty in detecting AI-generated content [7,16] may tempt students to engage in academic misconduct. This raises concerns about assessment fairness, equity, and the credibility of academic credentials earned through digital platforms. Educators, therefore, face challenges in maintaining the integrity of assessments and ensuring educational quality [16,30]. Recent studies also report learners’ desire for clear institutional guidance on appropriate AI use, aligning with calls for explicit policy and training [31,32].

La Trobe University, Australia, offers a wide range of health programs. The Digital Health and Health Information Management (DIGHIM) programs provide formal education in disciplines that support contemporary health care and medical practice, aiming to develop students’ technical, professional, and interpersonal skills [33,34]. Typical enrolments include undergraduate and master’s-level students from clinical or quasi-clinical backgrounds such as nursing, allied health, health sciences, health administration, biomedical sciences, and related fields, as well as graduates seeking roles at the intersection of health care delivery, health data, and digital systems. The curriculum is aligned with medical education priorities, including clinical documentation, health classification, health data governance, digital health (DH) system design, and the ethical use of health information, all of which directly support contemporary medical practice and health service delivery. The DIGHIM programs emphasize discipline-specific and generic competencies, which are inclusive of problem-solving, critical thinking, and ethical decision-making that are typically assessed through academic essays, reports, multiple-choice quizzes, presentations, case studies, programming exercises, and practical simulations [35,36]. The proliferation of AI-generated content threatens the validity and reliability of assessments in most of these areas, potentially compromising the effectiveness of evaluating students’ comprehension and competence [16]. In this context, the broader medical-education literature now differentiates by task type and competency domain: from automated short-answer scoring and case generation to simulation and reflective work, strengthening the rationale for DIGHIM-specific analysis of assessment susceptibility and redesign [21,37-39].

Previous research on the impact of GenAI tools in higher education has primarily focused on specific types of assessments, such as multiple-choice questions or essay evaluations, mainly within the medical [40-42] and business fields [4]. For example, Chaudhry et al [16] undertook an evaluation of AI-generated assessments in a Bachelor of Business Administration, including case analysis, empirical study report, self-reflection, group work, and calculation-based assessments. These findings do not necessarily translate well to the more specialized fields of DH and health information management (HIM), each of which requires a distinct combination of technical, analytical, and decision-making skills. DIGHIM programs, for instance, demand expertise in areas such as health classification, epidemiology and biostatistics, DH solution design, health data governance, and interoperability [33,34]. These specialist areas of study necessitate a rigorous and diverse set of assessment methods, incorporating scenario-based problem-solving, data interpretation, and system implementation tasks that test both theoretical knowledge and practical application. Given the interdisciplinary nature of DH, which aims to bridge clinical practice, data science, and IT, there is a need for a separate evaluation of how GenAI interacts with these unique assessments [43]. This study responds to that gap by examining assessment type susceptibility in DIGHIM and aligning implications with contemporary redesign strategies and governance recommendations emerging from the 2024-2025 literature [21,31,32,44].

This study examined how GenAI, specifically ChatGPT, performed across different DIGHIM assessment types and how its outputs compared with student work, with the aim of identifying assessment formats most susceptible to AI-generated content. It further sought to inform assessment redesign by identifying approaches that promoted authenticity, higher-order reasoning, and responsible AI use, thereby strengthening academic integrity through design rather than detection.

A quasi-experimental design was piloted to evaluate ChatGPT’s performance on a range of assessment types, comparing its results with submitted assessments of past students. This exploratory approach was intended to provide preliminary insights and inform the design of larger-scale studies. While similar methods have been applied in ChatGPT response evaluation within the education sector, this study pilots the approach in the context of DIGHIM, where it has not yet been examined [11,16,41].

Assessments from 5 DIGHIM subjects were purposively selected because of specific assessment characteristics, to maximize diversity and breadth in the type and nature of assessments chosen (Table 1). Past students’ deidentified assessments, originally submitted for academic credit, were reevaluated for research purposes as a benchmark against which the ChatGPT-generated responses were compared. Markers, blinded to the origin of each assessment, were responsible for evaluating both the ChatGPT-generated assessments and the past students’ assessments. Markers were academic staff, both full-time and casual, with postgraduate qualifications and advanced professional and teaching experience in DH or HIM relevant to the specific assessments they were assigned, and with demonstrated expertise in evaluating student work using standardized marking rubrics.

Marks and qualitative feedback provided by academic markers were compiled for each assessment type. Descriptive statistics (means, percentages, and score ranges) were calculated across all student grade bands (high, medium, and low) to support the blinded marking process and ensure natural variation within the control group. While ChatGPT outputs were compared against all remarked student assessments, the primary analytical focus was on the highest-graded submissions, as these offered a stable and meaningful benchmark for assessing whether AI-generated work could approximate high-quality human performance. Performance was examined both within individual assessment types and across assessment categories.

Comparative trend analysis was undertaken to evaluate progression across ChatGPT versions (V1-V4), focusing on improvements associated with structured prompting, schema inclusion, and rubric integration. In addition, rubric-level synthesis was conducted to identify recurring strengths and weaknesses.

Feedback comments for both student and AI-generated submissions were reviewed and compared, with particular focus on top-performing student work. This comparison helped identify where the evaluative feedback on AI outputs differed most noticeably from that of human-authored work.

Results were further synthesized into cross-assessment comparisons, with performance grouped into broader task categories (objective, reflective, descriptive analytical, scenario-based analytical, communication/referencing, and programming) to identify where AI demonstrated relative strengths versus persistent limitations.

The study used GPT-4 (free version/limited) and o1 Preview (pro/paid version) for the experimentation. The researchers opted out of ChatGPT’s AI training model using an available feature, ensuring that the submitted assessment guidelines and instructions were not used to further train the AI. All necessary privacy settings were enabled to maintain data confidentiality throughout the research process.

The research study was approved by La Trobe University’s Human Research Ethics Committee (application number HEC24286). No student assessment submissions, student-generated content, or identifiable educational data were entered into any GenAI tools as part of this study. Only publicly available assessment instructions and marking rubrics were provided to ChatGPT to generate AI responses. All student work used for comparison was deidentified, stored securely, and assessed offline by academic markers.

This research aimed to pilot the evaluation of ChatGPT’s performance across diverse DH and HIM assessment tasks, with a focus on promoting academic integrity. This analysis not only revealed the specific contexts in which ChatGPT excelled or fell short but also suggested broader implications for designing assessments that foster critical thinking, ethical practice, and real-world applicability. By situating ChatGPT’s performance within the study’s original goals, we were able to gain clearer insights into the potential and the limitations of GenAI in academic and professional environments. However, because this pilot examined only the final outputs generated by ChatGPT, it does not evaluate AI’s role as a formative learning partner, such as when students iteratively critique, refine, or codevelop AI-generated drafts.

The DH case study and solution design assessments revealed clear limitations in ChatGPT’s capacity to address complex, context-specific problems. While AI-generated submissions consistently identified foundational frameworks and produced well-structured outputs, they lacked the depth and contextualization required to develop effective, patient-specific solutions. Weak alignment between identified gaps and proposed interventions reflects broader challenges in maintaining contextual coherence. This is consistent with prior findings across disciplines that report similar shortcomings despite fluent language, structural coherence, and high grammatical accuracy [16,44]. The absence of visual artifacts, such as patient journey maps or flowcharts, further constrained the communication of complex ideas, underscoring limitations in GenAI’s ability to produce meaningful visual representations [45]. These findings reinforce the essential role of human expertise in refining and contextualizing solutions. For DIGHIM assessments, tasks should prioritize deep contextual engagement, patient-specific scenarios, and dynamic problem-solving. Requiring the integration of real-world data, graphical visualizations, and detailed case-based justification can better assess applied competence while reducing overreliance on AI tools. This approach aligns with constructivist learning theory, which emphasizes learning through authentic engagement rather than surface-level reproduction [46,47], and with higher-order cognitive processes in Bloom taxonomy that current GenAI systems cannot reliably replicate [48,49].

Reflective assessments demonstrated ChatGPT’s strongest performance, with later versions, particularly GPT-4 (V3), achieving the highest score at 80% and surpassing student submissions in structure, coherence, and adherence to reflective frameworks. Similar patterns have been reported in dental education, where ChatGPT performed strongly in structured-portfolio tasks [17]. However, despite polished language and systematic organization, AI-generated reflections often lacked genuine depth, contextual nuance, and personalized insight, which are central to HIM professional development. These results indicate a need to redesign reflective assessments to preserve authenticity. This aligns with social cognitive theory, which emphasizes authentic self-reflection, mastery experiences, and the development of self-efficacy as central to professional learning, all of which are elements that GenAI cannot meaningfully replicate [50]. Emphasizing personalized insights, unique contextual connections, and critical evaluation of lived experiences, alongside dynamic scenarios, peer interaction, or real-time contextual observations, can reduce AI dependence while reinforcing academic integrity and professional skill development [51,52].

Objective assessments, including multiple-choice and true or false questions, showed consistently high AI performance with minimal error rates. This reflects known strengths of AI in structured, rule-based tasks requiring factual recall and logical reasoning and aligns with prior evaluations of GenAI performance on objective assessments [53-56]. However, this raises concerns regarding the validity of such tasks in measuring individual learning, particularly in online contexts. To maintain assessment integrity, objective questions should be supplemented with higher-order components such as justification, explanation, or reasoning. Embedding real-world context and multistep problem-solving can further enhance authenticity and alignment with learning outcomes while limiting AI dominance [57].

Clinical coding scenario-based assessments revealed a clear performance gap between AI and students, particularly in tasks requiring nuanced contextual interpretation and application of coding rules. While AI performed comparably on objective components, it struggled with complex scenarios, resulting in significantly lower scores than high-performing students. This aligns with US-based findings showing that ChatGPT performs adequately for simple, single-diagnosis coding tasks but struggles with complex patient data [58,59]. Soroush et al [59] reported that GPT-4 achieved only a 34% exact match rate for ICD-10-CM (International Classification of Diseases, Tenth Revision, Clinical Modification)codes and concluded that LLMs lack a complete internal representation of medical coding rules, rendering them unsuitable for clinical coding tasks. These findings underscore the need for assessments that emphasize contextual reasoning and real-world coding application, reinforcing the importance of developing independent clinical coding proficiency.

SQL programming assessments highlighted further limitations in ChatGPT’s ability to perform technically precise, context-dependent tasks. Although later versions showed incremental improvement, persistent issues included incorrect table and column usage, overreliance on prompts, limited understanding of health data conventions, and failure to execute complex multistep queries. Errors such as misspellings and the absence of real-time testing and debugging capabilities further undermined output reliability. These findings contrast with prior research reporting strong SQL generation performance by LLMs [60], suggesting that health database contexts demand deeper domain knowledge and contextual awareness. Previous work has emphasized the importance of iterative testing and contextual understanding in SQL code generation [61]. For educators, this highlights the value of assessments that extend beyond syntax to include debugging, real-world application, and analytical interpretation of health data, thereby fostering job-ready competencies that AI cannot fully replicate.

In communication, writing, and referencing tasks, ChatGPT demonstrated strong performance in structure, clarity, and adherence to academic conventions, particularly in later versions such as V3. Nonetheless, limitations in depth, contextual relevance, repetition, and critical engagement persisted, especially in longer writing tasks. Similar findings have been reported in business education, where AI-generated responses for extended academic writing scored between 40% and 77% due to insufficient depth and analysis despite high grammatical quality [16]. Issues such as hallucinated references and inconsistent citation formatting further undermined credibility, consistent with prior research [62,63]. These results indicate that while AI can support foundational academic writing, it struggles with nuanced argumentation and original critical thought. Assessment designs should therefore require critical engagement with sources, precise referencing, and personalized application of concepts to ensure authentic student learning.

Chain prompting, involving iterative refinement of prompts, significantly improved AI performance, particularly in descriptive and reflective tasks, with later versions benefiting from rubric-aligned guidance. Comparable improvements have been observed in pharmacy education, where structured prompting enhanced performance in knowledge recall tasks [56]. This demonstrates the effectiveness of multistep prompting in optimizing AI outputs for well-defined assessment formats.

Overall, o1 Preview (V4) underperformed in descriptive tasks relative to V3, although performance was comparable or improved in simpler objective assessments. This suggests that newer AI versions do not necessarily yield better outcomes for complex, context-driven tasks. While earlier studies reported GPT-4 superiority in medical licensing examinations [64] and basic sciences [65], such improvements were not consistently observed in this study. There was no clear evidence that o1 Preview enhanced performance in tasks requiring critical thinking or reflection. These findings highlight the importance of assessment designs that prioritize originality, contextual understanding, and higher-order cognition over tasks that can be easily completed using AI assistance.

The findings from this pilot study highlight the importance of designing assessments that align with the evolving roles in HIM and DH while fostering the ethical and practical use of GenAI. Health information managers play a crucial role in health data management, analytics, and health information and communication technology [33], and as DH increasingly integrates disciplines such as clinical care, data science, and IT [66], it is essential for educational assessments to reinforce interdisciplinary skills that align with the evolving demands of these domains [67,68]. Recent evidence highlights that traditional assessment methods in higher education are increasingly ineffective in GenAI-facilitated learning environments, necessitating redesigned assessments that promote critical thinking, creativity, and lifelong learning skills [28]. For instance, practical redesign approaches, such as “flipped assessment,” can be implemented, where students cocreate and then critically validate AI-generated items under educator supervision, improving confidence while maintaining expert oversight [31]. This type of redesign aligns with this study’s recommendation for tasks that encourage students to critique, refine, and contextualize AI outputs rather than merely generate them.

Assessments should emphasize tasks that require contextual understanding, technical proficiency, and decision-making. Furthermore, these assessments should integrate AI as a tool to augment, rather than a substitute for critical thinking. For example, students could refine AI-generated clinical coding or SQL outputs to align with standards-based frameworks, critically evaluate AI-generated business proposals for depth and contextual relevance, or integrate visualizations such as patient journey maps into health information reports. These approaches not only build essential competencies such as critical analysis, programming, and communication but also encourage students to navigate AI’s limitations and responsibly incorporate its outputs into their work, promoting the ethical use of GenAI to support, rather than replace, student efforts.

In DIGHIM education, AI use in assessments should be carefully structured to align with professional competencies while maintaining academic integrity. Some tasks should avoid reliance on AI, particularly those requiring personal reflections, ethical decision-making, and critical thinking in complex DH scenarios. Other tasks can adopt AI for efficiency, such as assisting in summarizing health policies, generating structured reports, or organizing data-driven insights. Finally, AI-generated content can be adapted in assessments that require students to critically refine, validate, and contextualize AI outputs, such as evaluating AI-generated clinical coding recommendations, refining SQL queries for health databases, or assessing the applicability of AI-generated DH solutions within industry frameworks. Additionally, educators and students must ensure privacy by refraining from supplying GenAI tools with copyrighted or sensitive health data and should receive structured training on the responsible and ethical use of AI in DH and HIM to support professional readiness.

Taken together, current 2024-2025 evidence converges on a dual imperative: redesign assessments to cultivate critique, verification, and reflection skills, and establish clear governance mechanisms, including policy guidance, training programs, and disclosure norms, proportionate to the evolving capabilities and risks of GenAI systems [31,32,38].

These approaches prepare students for real-world applications of GenAI in HIM and DH roles, where such tools may be used to assist with technical tasks such as data analysis, report generation, and strategic planning. By designing assessments that both test and build skills in these areas while fostering responsible AI use, educators can equip graduates to navigate the interdisciplinary and technology-driven landscape of modern DH care effectively.

Table 8 provides a summary of important recommendations.

This study had several limitations. First, as a pilot study, the findings are preliminary and intended to provide directional insights rather than definitive conclusions; however, the use of multiple assessment types and structured AI submission versions was intended to strengthen internal comparison. Larger-scale studies are needed to confirm these patterns. Second, the quasi-experimental design lacked random assignment, which may limit generalizability. This was partially mitigated by comparing AI-generated submissions with authentic, previously graded student work under blinded conditions. Future studies could strengthen external validity through multiinstitutional or randomized designs. Third, assessments were marked by discipline-specific academic staff, and individual grading variability may have influenced results despite mitigation strategies, including standardized rubrics, blinded review, and second-marker consensus when remarked scores differed by more than 10 marks. Future work could further enhance reliability by incorporating multiple independent markers per assessment. Fourth, the number of assessment samples within each category was limited, which may reduce the robustness of task-specific conclusions. This was addressed by selecting assessments across multiple levels and disciplines, though larger samples are required in future studies. Fifth, the study could not fully replicate real-time student interactions with GenAI, such as iterative prompting or blended human-AI workflows. The structured versioning approach (V1-V4) was used to approximate common usage patterns, but longitudinal studies are needed to capture authentic use over time. Sixth, while all student submissions underwent formal academic integrity checks and no concerns were identified, it was not possible to determine with absolute certainty whether any student work involved AI assistance, which may have influenced comparisons. Finally, the study did not systematically assess the detectability of AI-generated content. Although markers occasionally inferred AI authorship, these judgments were informal and sometimes inaccurate, consistent with prior reports of false positives [16,30,69] and highlighting the unreliability of human detection methods.

Future research should explicitly investigate the accuracy of markers in distinguishing AI- from student-generated work and consider how assessment design can emphasize originality, critical thinking, and personalized insights to reduce reliance on detection and strengthen academic integrity. Another area that future studies could explore is expanding the scope beyond ChatGPT to include other advanced LLMs (eg, Gemini, Claude, and AI-assisted coding tools such as GitHub Copilot), to allow broader comparison of capabilities and provide a more comprehensive understanding of how different AI systems perform across assessment types in DIGHIM.

This pilot study has highlighted the potential and limitations of ChatGPT in HIM and DH assessments. While GenAI demonstrates strengths in structured tasks and foundational content generation, it struggles with contextualization, depth, and the technical precision required for assessments requiring independent critical thinking. These findings emphasize the need for carefully designed assessments that integrate AI ethically and prioritize tasks that require human judgment and contextual understanding to ensure meaningful learning outcomes and academic integrity.