The purpose of this investigation is to explore the use of ChatGPT as a resource to assist instructional designers and educators in selecting the most suitable media for delivering their instructional content. This inquiry is structured as follows: First, a brief introduction to ISD alongside a review of media selection is presented, and in doing so, the challenges are discussed. Next, research and development of AI as it applies to enhancing learning outcomes are explored, with emphasis on the use of AI-driven technology in ISD. The method employed to investigate media selection is subsequently described. Then, the findings are examined, accompanied by a detailed discussion. Finally, this work considers the future of AI-driven technology as a resource in ISD.

Altogether, this research examines the potential advantages and implications of using ChatGPT in instructional design. The aim is to encourage instructional designers and educators to explore and adopt generative AI-driven tools to streamline what are often demanding and time-consuming ISD practices. Finally, the results of this study were disseminated among educational professionals whose constructive feedback contributed to the refinement and improvement of the research.

Several terms that require clarification are found in this work. Historically, ISD has been recognized by various monikers. Terms like “educational technology” (Seels & Richey, 1994) , “instructional design” (Martin, 2011; Reiser, 2001b) , “instructional design and technology” (Reiser, 2001a, 2001b) , “instructional development” (Reiser, 2001b) , “instructional systems development” (Merrill & ID2 Research Group, 1996) , “instructional technology” (Seels & Richey, 1994) , and “the systems approach” (Reiser, 2001b) are all present in scholarly literature and have been used synonymously in the context of ISD. This diversity might partly stem from the field’s evolution. For example, Funaro and Mulligan (1978) discussed “systems analysis” as an antecedent to ISD and a “systems approach to training” in the context of developing and managing U.S. Naval training programs. Then, scholars have denoted differences among these terms (e.g., Seels & Richey, 1994 ), including the distinction between ISD and instructional design (e.g., Nolin, 2019 ). For the sake of simplicity, this investigation uses the term “instructional design” (ID).

Moreover, there have been many attempts to define the field of ID. Merrill and the ID2 Research Group (1996) described instructional systems development as a “set of procedures for systematically designing and developing instructional materials” (p. 30), whereas Merrill et al. (1996) described ID as “the use of these scientific principles [i.e., natural principles involved in instructional strategies] to invent instructional design procedures and tools” (p. 5). Reiser (2001a, 2001b) described the field of ID and technology as encompassing the “analysis of learning and performance problems, and the design, development, implementation, evaluation and management of instructional and non[-]instructional processes and resources intended to improve learning and performance in a variety of settings, particularly educational institutions and the workplace” (p. 53 and 57 respectively). While Martin (2011) defined ID as the “science of creating detailed specifications for the design, development, evaluation, and maintenance of instructional material that facilitates learning and performance” (p. 956).

The current work adopts the 1994 definition attributed to the Association for Educational Communications and Technology (AECT), which defined instructional technology as “the theory and practice of design, development, utilization, management and evaluation of processes and resources for learning” (see Seels & Richey (1994) for a comprehensive analysis of the definition, to include its origin). This definition was chosen due to the AECT’s (2023) longstanding dedication to the field.

Artificial intelligence is an area of study that has also undergone many attempts to be defined and succinctly explained. Citing others, Pokrivcakova (2019) explained that AI has been defined in the context of machines, computers, or systems (e.g., Russell & Norvig, 2010 ) (although it is said not to describe any one technology; Baker & Smith, 2019 ), as a specific set of skills (e.g., Baker & Smith, 2019 ), and more broadly, as a science (e.g., Stone et al., 2016 ). Pokrivcakova (2019) adopted the definition of Luckin et al. (2016) , who borrowed from all three contexts, defining AI “as computer systems that have been designed to interact with the world through capabilities (for example, visual perception and speech recognition) and intelligent [behaviors] (for example, assessing the available information and then taking the most sensible action to achieve a stated goal) that we would think of as essentially human” (p. 14). Chen et al. (2020) went on to remark that although definitions deviate in language, many have the common theme of “mimicking human cognition by computers” (Wartman & Combs, 2018: p. 1107) , formulating their broader definition from the literature as AI encompassing “the development of machines that have some level of intelligence, with the ability to perform human-like functions, including cognitive, learning, decision making, and adapting to the environment” (p. 75267). Given the focus on ChatGPT, a large language model dependent on advancements within subfields of AI, this work adopts the explanation by Gimpel et al. (2023) , who characterized AI as a broad field that includes the creation of intelligent machines that can perceive their environment while making decisions with machine learning serving as a subfield enabling computers to enhance task performance through pattern recognition and data-driven predictions.

Finally, an “instructional designer” is defined as someone who uses a systematic approach grounded in ID theory and practices to develop materials for learning experiences. This individual may be a formally trained professional in ID or someone with practical, real-world experience.

Dating to the Second World War (Dick, 1987; Funaro & Mulligan, 1978; Reiser, 2001a, 2001b; Seels, 1997) , ID emerged from military and academic initiatives during the late 1940s (Dick, 1987; Funaro & Mulligan, 1978) , continuing through the 1950s (Dick, 1987; Funaro & Mulligan, 1978; Reiser, 2001b; Wilson & Jonassen, 1990) and 1960s (Funaro & Mulligan, 1978; Reiser, 2001b; Wilson & Jonassen, 1990) . The aim was to establish systematic approaches for developing instructional materials (Reiser, 2001b; Seels, 1989; Wilson & Jonassen, 1990) , drawing on concepts from general systems theory (Seels, 1997; Wilson & Jonassen, 1990) , psychology, management science (Wilson & Jonassen, 1990) , instructional theory, and communications theory (Seels, 1997) . The thought was that system design methods and principles used in the advancement of weapons and manufacturing could be used to develop complex training systems (Wilson & Jonassen, 1990) . That is, it was proposed that something could be learned by examining design models found in engineering and other similar design fields ( Wilson & Jonassen, 1990; see Reiser (2001b) for a historical account of ID and technology).

In the years that followed, ID has been rooted in dozens of models and frameworks to guide the creation of instructional and training materials (Nolin, 2019; Reiser, 2001b; Wilson & Jonassen, 1990) . Reiser (2001b) notably highlighted Gustafson and Branch’s (1997) detailed examination of the many models formulated during the 1980s and 1990s. Contrary to being a singular theory (Merrill & ID2 Research Group, 1996) , ID encompasses diverse principles (Martin, 2011; Nolin, 2019) developed with different emphases, such as learner participation, interaction, and engagement, as well as motivation, and student success (Nolin, 2019) . Therefore, it is common for instructional designers to rely on various models, frameworks, and theories to steer them through the phases of ID (Martin, 2011; Nolin, 2019; Reiser, 2001b) .

Among these models, Gagné (1965) categorized learning outcomes into five domains—verbal information, intellectual skills, psychomotor skills, attitudes, and cognitive strategies—each of which necessitated unique conditions for promoting learning (Reiser, 2001b) . He also outlined nine events of instruction—representing essential teaching activities to facilitate the achievement of diverse types of learning outcomes—detailing which instructional events were critical for certain types of outcomes and providing insights into the situations where certain events could be omitted (Reiser, 2001b) . Dick and Carey (1990) presented nine consecutive steps intended to achieve an instructional objective, beginning with the establishment of goals and concluding with a summative evaluation. Merrill et al. (1991) introduced the concept of knowledge objects to represent knowledge that can be used to build transactional shells (or instructional algorithms) to teach the knowledge or skill. Kemp et al. (1994) described a circular ID framework with nine interconnected components, presenting a non-linear, ongoing cycle of implementation and evaluation, contrasting with the Dick and Carey model by offering greater flexibility in navigating its stages. Then there is the work of Van Merriënboer (1997) , who posited that complex learning programs aimed at teaching professional competencies that integrate knowledge, skills, attitudes, and skill coordination can be deconstructed into four components: learning tasks, supportive information, procedural information, and part-task practice (Frerejean et al., 2019) .

Altogether, there is no shortage of models, frameworks, and taxonomies to aid in the ID process. The outpouring, however, has made it difficult for instructional designers to identify the most appropriate model to use for a given situation, further clouding the decision-making process (Slagter van Tryon et al., 2018) . This is problematic, given that one of the skills instructional designers are said to need is the ability to select an ID process (Lenane, 2022) . It also highlights the broader challenges, as ID—and, by extension, education—is complex and multifaceted (Bond & Dirkin, 2020) . It has been asserted, for instance, that the emerging nature of the profession itself constrains a shared comprehension of its practices (Sharif & Cho, 2015) .

The literature abounds with opinions and positions offering ID guidance that speaks to the many challenges in the field. Even the longevity of models has been questioned (e.g., Dick, 1996 ). It is argued, in the context of the current work, that the largest issue in ID is the ever-changing landscape of educational technology and pedagogy. With the rapid advancement of technology, instructional designers and educators must continually adapt to incorporate new tools and platforms effectively (Gameil & Al-Abdullatif, 2023) . This requires staying abreast of technological trends and understanding how to leverage these tools to enhance learning experiences (Liu et al., 2002) without overwhelming learners or sacrificing didactic integrity. However, finding the right balance between educational objectives and technical capabilities is difficult, as instructional designers must ensure that the technology supports learning (Holden et al., 2010) rather than driving the process (Martin, 2011) . This includes supporting the diverse needs of today’s learners, including accessibility matters (Pollard & Kumar, 2022; Office of Educational Technology, 2023) demanding technologies that are inclusive and adaptive.

The influence of technology on ID cannot be overstated. It speaks heavily to the importance of selecting the most appropriate media for instructional delivery.

Martin (2011) explained that a substantial aspect of the instructional process entails presenting learners with the essential information required for learning (Gagné, 1985; Reiser & Dick, 1996) . Emphasis should be placed on important (or characteristic) aspects of what needs to be learned (Gagné, 1985) , which should be chunked and organized in a meaningful way (Kruse & Keil, 1999) . This thinking further underscores the importance of selecting suitable media in that the selection supports the information and essential characteristics that learners need to grasp. Thus, in addition to focusing on effective representation and organization of the subject matter to facilitate learning (Merrill & ID2 Research Group, 1996) , it is equally imperative to integrate media that can support the desired learning goals.

The integration of media is not new. It has long been asserted that technology should be considered part of the instructional strategy (Martin, 2011; Merrill & ID2 Research Group, 1996) so that the selected media is aligned with the instructional objectives (Martin, 2011) . Seels (1997) stressed that delivery system selection (and development) is as essential as the other design stages, explaining that content, objectives, assessment, strategies, and delivery systems should be incorporated through systems and theory (i.e., the essential elements of effective instruction should be integrated into a cohesive whole). Reiser (2001b) echoed this, stating that the integration of media is one of the practices that, over the years, has formed the “core of the field” (p. 57), with media selection a recurring element found in many ID models.

Early academics stressed that instructional designers had yet to lay the groundwork for establishing robust approaches for choosing media suitable for specific learning tasks (e.g., Hawkridge, 1973 ). Hawkridge (1973) , for example, offered that his university’s choices of media were driven more by logistical, financial, and political factors rather than by well-founded, explicitly defined psychological and educational considerations. Scholars also highlighted the inefficiency of formal media selection procedures as being overly broad, too simplistic, or excessively complex (e.g., Braby, 1973 ). Upon conducting interviews with 29 instructional designers at four U.S. Army schools, Gagné et al. (1981) observed that the complexity of procedures often led to media selection decisions based on experience and intuition, therefore introducing the human factor into decision-making regardless of the model being used.

The difficulty of media selection is no secret (e.g., Braby, 1973; Gagné et al., 1981; Hawkridge, 1973; Sugrue & Clark, 2000 ). Sugrue and Clark (2000) referenced Heidt (1989) , who noted the inherent difficulty in handling media selection models, explaining that none of the examples reviewed were considered “easy-to-handle,” indicating a significant level of complexity (p. 397). This is arguably attributed to a multitude of factors.

Models have been accused of being more similar than different (Sugrue & Clark, 2000) . Sugrue and Clark (2000) spoke to the conclusions of Heidt (1989) , who found that of the models reviewed, some included identifiable factors that should be considered in the decision-making process, while others proposed procedures that typically involve elaborate steps culminating in a common-sense decision. Of the models comprising linear procedures, several incorporated the selection of media as an activity that occurs after initial design but before development (Sugrue & Clark, 2000) . So, as diverse as these models are, a consensus could be argued regarding the timing of media selection while highlighting the potential intricacies of processes that must weigh numerous factors.

Moreover, many models promote a two-stage process, first focused on instructional characteristics and then on practical considerations (Sugrue & Clark, 2000) . Seels (1997) offered the example of Reiser and Gagné (1983) , who grounded their work on the principle that ID factors should be prioritized when selecting the most suitable media. They centered their model on principles of human learning to guide selection, emphasizing the creation of optimal learning conditions and only then looking at practical factors (Seels, 1997) . This means that while there have been attempts to establish a scientific approach to media selection centered around instructional elements, models continue to account for practical considerations, including cost and other logistical aspects, which add a layer of subjectivity.

Compounding matters, Nolin (2019) , citing Reigeluth’s (1999) work, explained that ID is often described as prescriptive. It has been reasoned that because ID is not a theory, it stands as a series of procedural steps primarily emphasizing what to do rather than how to do it or why it is effective (Merrill & ID2 Research Group, 1996) . This only exacerbates matters, particularly for those who are not versed in ID theories and practices. Reiser (2001b) viewed individuals in the ID field as those who devote a substantial portion of their time to working with media and tasks related to systematic ID procedures or a combination of both. As defined in this study, however, instructional designers could be those who possess practical, firsthand experience. Thus, while these individuals are likely experts in their respective areas, they may lack essential ID knowledge (Martin, 2011) to navigate the complexity of these models.

Another challenge has been the lack of experimental evaluation (or the comparison of methods), with no formal efforts to determine their reliability and validity (Braby, 1973) . Sugrue and Clark (2000) maintained that justifying media selection rules on grounds other than practicality is challenging because there is no proof that any medium uniquely enhances learning, motivation, or the development of distinct or transferable cognitive skills ( Clark & Salomon, 1986; Clark & Sugrue, 1991; see Clark (1994) for a discussion of the counter-arguments, and Clark (1983) for a discussion of media comparison research). To support their claim, Sugrue and Clark (2000) referenced studies by Braby (1973) and Romiszowski (1970) , who reported conflicting findings in their examination of the comparative effectiveness of media selection models and intuitive methods. This supports Clark’s (1983, 1994) longstanding view that media serve primarily as conduits for instructional methods rather than directly influencing cognition.

Building upon this is the idea that no single medium is inherently superior or inferior to others. Ziagos (1991) explained that in the 1960s and 1970s, delivery media was traditionally seen as crucial for effective instruction, but by the 1990s, research suggested that if the media supports the instructional method to meet the learning objective, all delivery media could be equally effective. From this, Ziagos (1991) concluded that it is the method, not the medium of delivery, which determines the effectiveness of learning. In subsequent years, it has been proposed that the suitability of media elements depends on factors like the unique characteristics of each medium and the context in which it is used (e.g., Holden et al., 2010 ).

However, identifying the unique characteristics of media has its share of challenges. Sugrue and Clark (2000) cited Heidt (1975) , who highlighted issues about categorizing media to deliver learning, pointing to cases in which the classifications were too broad or offered little regarding their variances. Finally, there is the issue that although media may have been appropriate at their inception, they become less suitable over time (Sugrue & Clark, 2000) . This stresses the importance of prioritizing media that consistently aligns with and reinforces the intended learning outcomes (Holden et al., 2010) rather than focusing predominantly on the novelty or technological sophistication of the media (Martin, 2011) .

Altogether, despite the impetus to incorporate media selection into the ID process, choosing suitable media continues to be a challenge, in which selections are instead founded on pragmatic factors along with experience and intuition. Sugrue and Clark (2000) cited Heidt (1989) , who noted that apart from practical and quantifiable factors, the criteria for selecting media depend on the subjective judgment of the instructional designer or instructor, guided by potentially relevant considerations. Anderson (1983) perhaps best summarized the matter, asserting that charts (alongside other guides) found in models are aimed at structuring the process systematically and comprehensively. Consequently, media selection should not be viewed as an exact science (Anderson, 1983) .

An area of study that is making a profound impact on enhancing learning outcomes is AI (Luckin et al., 2022) . Artificial Intelligence has ushered in transformative advantages, marked by its ability to analyze large datasets efficiently, thus enhancing decision-making processes in a multitude of fields (Bahroun et al., 2023) ranging from healthcare to environmental science. For example, AI-based systems are being explored to create personalized solutions to improve patient outcomes using AI-driven diagnostics and treatment plans (e.g., Esteva et al., 2019 ). Although algorithms have long been used in anti-malware software, AI is revolutionizing cybersecurity in protecting Internet-connected systems through the application of machine learning, deep learning, and natural language processing, thereby automating and improving the intelligence of cybersecurity services and management (e.g., Sarker et al., 2021 ). Then, there are environmental conservation efforts, in which AI techniques are being used in the better sustainment of natural resources and the modeling of climate change scenarios, showing the benefits of AI-driven technologies in addressing global environmental challenges (e.g., Rolnick et al., 2022 ).

These examples, though limited, underscore AI’s multifaceted benefits, particularly its role in fostering innovation across diverse fields. This includes enhancing efficiency via quicker decision-making and problem-solving, automating both routine and complex tasks, analyzing vast data sets, and conducting predictive analytics. Furthermore, AI improves accessibility personalization by analyzing individual preferences, behaviors, and needs (Office of Educational Technology, 2023) .

It is this personalization, however, that is perhaps of most interest in the field of education. Using AI in education (AIEd) presents significant opportunities, enabling, for example, tailored learning experiences (Bahroun et al., 2023; Office of Educational Technology, 2023; Ruiz-Rojas et al., 2023) . Discussions on AIEd often touch upon personalized learning (e.g., Bahroun et al., 2023; Baidoo-Anu & Ansah, 2023; Bond et al., 2024; İpek et al., 2023; Luckin et al., 2016; Office of Educational Technology, 2023; Rudolph et al., 2023; Shum & Luckin, 2019; Zawacki-Richter et al., 2019; Zhai et al., 2021 ), which involves customizing content and instruction to suit the needs and preferences of learners (Atlas, 2023) . It is believed that AI’s ability to analyze and adapt content based on learner feedback and performance can lead to more effective and personalized experiences (Shum & Luckin, 2019) .

A representation of this is intelligent tutoring, which can be thought of as an innovative form of personalized learning and adaptive testing in which the technology provides customized, step-by-step guidance and real-time feedback personalized to each learner’s performance and needs (Atlas, 2023) . It is akin to having a personal tutor who can tailor their teaching to their student’s learning needs but is void of a direct human presence (Luckin et al., 2016) . As with personalized learning, intelligent tutoring is commonly discussed in the literature (e.g., Atlas, 2023; Bond et al., 2024; Luckin et al., 2016; Pokrivcakova, 2019; Rudolph et al., 2023; Stone et al., 2016; Zawacki-Richter et al., 2019; Zhai et al., 2021 ), with ample research that suggests intelligent tutoring systems enhance learning outcomes, whether used independently or in conjunction with traditional teaching methods (Weitekamp et al., 2020) .

Indeed, attempts at using intelligent computers in the design of instruction go back decades (Bond et al., 2024) , with Luckin et al. (2016) claiming AIEd has been of interest for over 30 years. However, within the past decade, the potential societal impact of AI has garnered considerable attention (Cooper, 2023) , indicating a growing interest in its broader implications. So, unsurprising, AI has been said to have far-reaching consequences in the fields of business (Atlas, 2023; Susnjak, 2022) and finance (Bahroun et al., 2023; Luckin & Cukurova, 2019) , science and engineering (Atlas, 2023; Bahroun et al., 2023; Susnjak, 2022) , the humanities (Atlas, 2023; Susnjak, 2022) , healthcare (Atlas, 2023; Bahroun et al., 2023; Luckin & Cukurova, 2019) , and law (Atlas, 2023) . This widespread applicability underscores its potential, reflecting its versatility and transformative power.

The recent emergence of generative AI holds possibilities within education (Cooper, 2023; İpek et al., 2023; Zawacki-Richter et al., 2019) . The same has been said about other technologies throughout the 20th century (e.g., computers, film, radio, television; Cuban, 1986 ), and as with other advancements, the use of generative AIEd (Gimpel et al., 2023; İpek et al., 2023) is intensely debated (Atlas, 2023; Baidoo-Anu & Ansah, 2023; Gimpel et al., 2023; İpek et al., 2023; Rudolph et al., 2023) . Plagiarism, for instance, alongside other forms of cheating, is commonly discussed in the literature (e.g., Atlas, 2023; Gimpel et al., 2023; İpek et al., 2023; Office of Educational Technology, 2023; Rudolph et al., 2023; Susnjak, 2022 ). However, there is ample research suggesting that AIEd has far-reaching benefits.

Luckin and Cukurova (2019) pointed out that research demonstrating well-designed AI that works in education is increasing, alongside the growing argument that AIEd contexts have yet to be fully realized (Celik, 2023; Luckin et al., 2022) . Atlas (2023) detailed the versatile use of generative AIEd, encompassing automated essay scoring, personalized tutoring, classroom assistance, language translation, and the development of writing, research, and communication skills. Additionally, this technology can aid in creating syllabi, quizzes, exams, summaries, reports, and other research documents while also offering support in email and chatbot communication, meeting and event organization, campus tours, policy guidance, and report generation (Atlas, 2023) .

ChatGPT is a large language model (Gimpel et al., 2023; Susnjak, 2022) developed by OpenAI that can generate human-like text responses and provide natural language understanding in a conversational context (ChatGPT-4, personal communication, December 15, 2023). Generative AI uses large language models, a type of machine learning model, to produce text, images, and other content with minimal human involvement (Gimpel et al., 2023) . GPT, short for Generative Pre-Trained Transformer (Gimpel et al., 2023; Tlili et al., 2023) , is OpenAI’s acronym for large language models that undergo pre-training on publicly available data (Gimpel et al., 2023) . OpenAI introduced its advanced models in November 2022, culminating in the release of ChatGPT-4 in March 2023 (Gimpel et al., 2023; İpek et al., 2023) . Cooper (2023) , citing Scharth (2022) , explained that these models were developed using machine learning algorithms trained on comprehensive text datasets comprising books, news articles, and websites. (See Rudolph et al. (2023) for an explanation of ChatGPT and its history.)

ChatGPT’s popularity has been attributed to its ability to produce high-quality text, its user-friendliness, as well as its free access (Atlas, 2023) (for ChatGPT-3.5; Gimpel et al., 2023 ), enabling new users to experience the possibilities of text generation firsthand. Its popularity may also come from its ability to be incorporated into existing applications. Gimpel et al. (2023) , referencing Salz (2023) , explained that ChatGPT is seeing rapid integration into Microsoft 365, making it increasingly challenging to evade the influence of AI. Moreover, it is believed that AI will continue to impact the development of technological tools for education and training (Luckin & Cukurova, 2019) .

Tools like ChatGPT could, therefore, be viewed as disruptive in education (Rudolph et al., 2023) . There is a concern that AI will assume roles performed by humans (Atlas, 2023; Office of Educational Technology, 2023) . From a historical standpoint, though, technological changes have brought new roles, creating new opportunities (Ch’ng, 2023; Luckin et al., 2016) . Gimpel et al. (2023) argued that ChatGPT (and generative AI tools like it) are supplementary; that is, used to enhance learning, teaching, and research (Atlas, 2023) . Thus, while it is conceivable that AI may assume some tasks currently performed by humans, generative AI is more likely to aid in curriculum development and similar tasks rather than entirely replace instructional designers and educators (Atlas, 2023) .

Generative AI can be used to free individuals from basic tasks, affording the time to focus on more advanced higher-order thinking (Ch’ng, 2023; İpek et al., 2023) . Instructional designers can use ChatGPT to perform laborious tasks so they can concentrate on other aspects of ID, reducing course content turnaround. Comparing traditional, current ID to AI collaborative practices, Ch’ng (2023) offered that content is often written by subject matter experts (SMEs) who conduct time-consuming research. Complicating matters, instructional designers must often deal with a general lack of understanding when it comes to their function and role, particularly as it relates to collaborating with others, including these experts (Pollard & Kumar, 2022) . This can create challenges because, at least in the case of SMEs, instructional designers are not necessarily authorities in an area (Pollard & Kumar, 2022) ; thus, their reliance on these experts is imperative. ChatGPT can be used to assist in identifying sources and rapidly summarizing articles (Atlas, 2023) , including translating content from different languages to help SMEs as well as instructional designers enhance collaboration efforts. Content can be logically chunked and organized, creating outlines that instructional designers can flesh out. It can also be used as a copyeditor to assist in producing polished text (Gimpel et al., 2023) . Then there is its ability to create introductions, objectives, main content topics, quizzes, assignments, and instructions (Ch’ng, 2023) .

Ch’ng (2023) —continuing the comparison of traditional ID to AI collaborative practices, but in the context of video, audio, and images—explained that the generation of multimedia can be a time-consuming and costly process as the production of video with audio typically involves a studio and recording equipment. However, ChatGPT can assist with aspects of multimedia generation. For example, it can be used to generate transcripts (based on text input) for closed captioning. ChatGPT can also aid in developing scripts to be used in the later creation of voice-overs. As for images, instructional designers must find copyright-free (or purchase stock) images that match the content and objectives (Ch’ng, 2023) , create the illustrations themselves, or rely on artists and graphic designers. Although the copyright of images generated by AI is currently evolving (Ch’ng, 2023) , tools like DALL·E can be used to create drawings, illustrations, and related graphics.

Overall, generative AI holds the potential to undertake lengthy and complex tasks, thus playing a supportive role in the ID process. This idea is perhaps best conveyed by the findings of İpek et al.’s (2023) systematic review of ChatGPT, which highlighted the evolving capabilities of generative AI in academic and research contexts. Some of these included summarizing and creating abstracts of articles, finding literature resources, generating literature, translating and paraphrasing, generating exam answers, identifying learners’ needs early in the learning process, personalizing learning experiences, grading and assessment, conducting data analysis, and helping with studying (İpek et al., 2023) . (See İpek et al. (2023) for other capabilities and an extensive list of research published in the examination of ChatGPT, and Bahroun et al. (2023) for a comprehensive analysis of generative AIEd.)

The authors and the GPT concurred in their identification of the central verb in all tasks except for 5.1. The authors identified the verb “conduct,” whereas the GPT chose “evaluate,” though recognizing the task involved “conduct[ing] a portfolio review.” The authors saw the task as performing the review, while the GPT likely focused on judging the content’s quality, as reflected in their respective category selections of “adaptation” and “evaluation,” respectively. This resulted in varied choices for the remaining instructional elements, which, in the context of each analysis, were aligned with the guide and knowledge source contents.

The authors determined the learning outcome to be “cognitive strategies,” while the GPT decided upon “intellectual skills,” reasoning that the task “requires learners to critically assess and provide feedback on portfolios, a task that involves higher-order cognitive processes like analysis, synthesis, and evaluation.” Despite these different interpretations of the task, both agreed that a facilitator-led “small group” was the optimal instructional method, subsequently suggesting a “peer/pressure group” instructional strategy. This led to distinct yet

The GPT generated the italicized text.

similar methods of delivery. While the authors favored a classroom environment, they also suggested online breakout rooms, which corresponded with the GPT’s recommendation of a learning management system (LMS) coupled with a synchronous online platform (e.g., Microsoft Teams).

Excluding 1.4, variations in the instructional elements were observed across all the other tasks. As shown in Table 3, tasks 1.1, 1.2, and 1.3 were dissimilar in learning level, category, learning outcome, LOI, and instructional method. Although “specify” was used for the verb, focusing on the cognitive (knowledge) domain, the authors viewed the task as learning “factual” information, whereas the GPT likely saw the matter as responding differently to specific (i.e., aperture, shutter, ISO), albeit similar stimuli. This thinking persisted in their subsequent choices, with the authors characterizing the outcome as “verbal” with an LOI 2, designating limited learner involvement, and the GPT viewing the matter as an “intellectual skill” necessitating a higher order of participation (LOI 3). Interestingly, even with deviating in their analysis, both opted for a “traditional” instructional strategy and proposed distinct yet analogous delivery systems. The authors advised self-study via physical or digital textbook, while the GPT suggested an interactive e-learning module or a digital text with integrated tutorials.

Not all the analyses resulted in differences across the same instructional elements. For 2.1, the authors and the GPT agreed upon the learning outcome and LOI but diverged regarding learning level, category, instructional method, and instructional strategy. Tasks 3.1, 3.2, and 3.3 revealed agreement on all the elements and conflicted on outcome, method, and strategy. Whereas 4.1 only showed a match with LOI and strategy. These differences do not suggest the analyses were flawed, but rather, the authors and GPT viewed the tasks differently, as was described above with tasks 1.1, 1.2, 1.3, and 5.1.

Reinforcing this point with another example, the authors and the GPT viewed tasks 3.1, 3.2, and 3.3 as a skill comprising a “complex overt response (mechanism),” which the GPT summarized as “performing complex and coordinated actions, such as manipulating camera settings (shutter speed) to achieve specific photographic effects.” However, the authors noted a learning outcome of “cognitive strategies,” whereas the GPT denoted a “motor skill.” That is, the authors saw a cognitive task, basing their conclusion, in part, on the work of Gagné (1985) , who characterized “cognitive strategy” as an internal regulatory process, allowing learners to choose and adjust their methods of focusing, acquiring knowledge, retaining information, and reasoning, as well as the insights of Bruner et al. (1956) on the operation and usefulness of cognitive strategies in problem-solving. Conversely, the GPT deduced a physical task, rationalizing that “it focuses on developing the physical ability to adjust camera settings effectively to achieve different effects in photographs.”

The most interesting of these findings is that, as with 1.1 through 1.4 and 5.1, the remaining tasks revealed different yet compatible delivery systems. In tasks 3.1, 3.2, and 3.3, the authors and the GPT advocated using the camera, albeit presumably influenced by the enabling objective. Additionally, the GPT recommended using AR and VR as auxiliary technologies, reasoning that they could assist in “demonstrating the effects of different aperture settings in various lighting conditions without the need for immediate access to indoor and outdoor settings.” The exception in delivery system selection was task 2.1. Even though the choices made sense in the context of their respective analysis, the authors proposed a traditional classroom or online breakout room, while the GPT recommended interactive multimedia instruction (IMI).

Although the selections did not always align between the authors and the GPT, as presented, they were aligned within their respective individual analyses. Nevertheless, it is important to note that the authors and the GPT did go “off script,” which may help explain some of their choices beyond their initial interpretation of each task. It is recognized that the media selection approach provided to the authors and the GPT served as general guidance. Furthermore, the contents of the MSG and knowledge source were in no way exhaustive; they only included a few examples. Therefore, a certain level of deviation was anticipated.

Based on this, the selection of a combination of instructional methods, instructional strategies, and delivery systems showed a departure from the prescribed guidelines. For tasks 1.1, 1.2, and 1.3, the authors noted “assigned reading” as the method, “traditional” as the strategy, and “self-study” as the delivery system. Though not explicitly one of the methods in the MSG, the authors based their “assigned reading” decision on guidance from the MIL-HDBK-29612-2A (DoD, 2001) , which underpinned the MSG. Moreover, “assigned reading” is contextualized in the MIL-HDBK within lecture and traditional classroom learning, which influenced the authors’ decision to adopt a “traditional” strategy. Meanwhile, “self-study” via textbook was based on the instructional elements alongside the authors’ experiences.

For these tasks, the GPT also departed from its training, choosing a “reference-based” instructional method, which it explained encompassed “utilizing the text as a primary source of information.” In the knowledge source, “reference-based” is not linked to “intellectual skills”; however, it is connected to the “traditional” instructional strategy. Furthermore, the choice of employing an interactive e-learning module or digitally enhanced text with tutorials was comparable with the authors’ preference for a digital textbook. Thus, the GPT’s analysis is arguably aligned with the authors, albeit departing from the examples in its method selection, but presumably a result of its exposure to the enabling objective.

Of the remaining tasks, the authors diverged from the MSG in their choice of instructional method and instructional strategy for tasks 3.1, 3.2, and 3.3, while the GPT departed from its training regarding the method and strategy decisions of 4.1. For tasks 3.1 through 3.3, the authors concluded on “practical application” for the method and “traditional” as the strategy. Although not explicitly among the examples, the authors decided upon the method of “practical application” because they viewed the tasks in the context of individuals performing them and having their specific environments (i.e., indoor and outdoor settings) that required camera configurations unique to each situation. The authors identified that trial and error learning would be enhanced to meet the photo portfolio requirement. Otherwise, a small group method could have been selected with the same element alignment but might allow an individual to lean heavily on teammates as opposed to putting in the individualized effort to learn and practice.

As for task 4.1, the GPT noted an instructional method of “project-based learning” and the instructional strategy, “Exercise, Experiential, or Experimental (E3)”. “Project-based learning” was also not part of the training examples but was instead derived by the GPT, demonstrating its heuristic capability to make independent choices. It could also be said that—although project-based learning has been described in the context of ID—its selection might have been influenced by the task’s language, which referenced the use of a portfolio. This rationale was further articulated by the GPT, which noted that “project-based learning is particularly relevant as it involves completing a real-world task (here, the creation of a photo portfolio) that allows learners to apply and demonstrate their understanding in a practical context.” Finally, its decision to adopt an “E3” strategy was based on the rationale that the task “emphasizes learning through doing and experimenting, which is essential for understanding various photographic techniques.”

Finally, using the MST revealed additional observations beyond those previously mentioned. The GPT consistently selected instructional elements across tasks 1.1, 1.2, and 1.3, as well as 3.1, 3.2, and 3.3. However, for tasks 1.1 through 1.4, a secondary instructional strategy aligned with the secondary instructional method was chosen. This behavior strayed from its training, as the model was directed to “choose a primary and secondary instructional method” and then “emphasize a single strategy as the most suitable.” Furthermore, for tasks 1.1 through 1.4, the GPT offered a rationale for the chosen secondary strategies. While consistent in meaning, the language varied, demonstrating its non-deterministic behavior. This means ChatGPT may produce diverse responses to identical prompts (ChatGPT-4, personal communication, January 22, 2024).

Building upon ChatGPT’s non-deterministic nature, the GPT also exhibited various levels of detail in explaining its selections. That is, despite instructions to “[provide] detailed yet easy-to-understand explanations” and “[offer] rationale behind selections,” the GPT exhibited inconsistency in its verboseness while explaining its decision-making rationale. For instance, the GPT provided detailed explanations for its choices in tasks 3.1 and 3.2, offering much less in 3.3. Altogether, the GPT’s performance indicates a tendency towards variability and unpredictability, not only in its strategic selections but also in the level of its explanations.