Active recall is a highly effective learning strategy and significantly outperforms passive restudying when it comes to certain learning outcomes, such as conceptualization and long-term retention [21]. It has yielded better evaluation testing performance than traditional studying or rereading. Kornell et al [22] reviewed recall when participants were presented with fictional and nonfictional information, modifying the time for pretesting and read-only strategies. The testing strategy yielded a greater amount of correct answers than the read-only strategy, with statistical significance when equal or more time was allocated to the testing condition when the final test was performed more than 24 hours after the learning exercise, as well as in the fictional topic scenarios (P<.01). Other studies supporting the use of pretesting have been reported [22,40,41]. This highlights the role of pretesting in learning new information.

The benefits of active learning through testing have also been supported by other authors. Butler [25] tested students’ recall ability when they were either passively restudying or studying via repeated testing. Butler [25] found that repeated testing resulted in better performance on a recall test than passive learning strategies and concluded that repeated test-taking increases the transfer of learning. In another study performed by Karpicke and Blunt [26], when retrieval practice (testing) was evaluated against passive learning strategies and even concept mapping, it proved to be better for verbatim and inference question answering, resulting in an improvement of about 50% in long-term retention scores (d=1.50; F_1,38=21.63; η_p²=0.36). Additionally, the superiority of retesting over passive restudying for long-term retention has even been proven in a randomized controlled trial, wherein pediatric and emergency medicine residents were randomized to study the same text passages either via testing or repeated studying (ie, rereading). They were then tested on day 1, week 2, week 4, and month 6. The test results showed that the scores of participants who studied via testing were, on average, 13% higher than the scores of participants who performed repeated studying (P<.001), with an effect size of 0.91 [28]. Spaced testing (taking tests on different days between study sessions) has an even better effect on retention, long-term memory, and evaluation performance than repeated test-taking [27]. On the other hand, research indicates that spaced repetition (regardless of whether studying is done actively or passively) promotes more efficient and effective learning [42].

The previously mentioned studies highlight the importance of leveraging evidence-based techniques for studying rather than passive learning strategies. As we exemplified, these strategies can be coupled with AI. This approach addresses the limitation of relying solely on teacher-provided tests or textbook tests [20]. Moreover, ChatGPT is available on different platforms (web application and mobile app), and it can save chats (interactions) across these platforms. Therefore, students can easily access ChatGPT wherever it is needed and space their study sessions. Further, self-testing with AI reduces the pressure of graded assessments and leverages errors as learning opportunities [23,24], which research has shown to be particularly effective when the learner is confident in their incorrect answers [23]. This could be related to the effect of feedback.

Several articles on active recall and learning emphasize the role of feedback in enhancing learning processes, which is a characteristic that passive studying lacks. Roediger and Butler [27] compared traditional studying with test-taking studying without feedback and test-taking studying with differently timed feedback to determine whether test-taking studying results in better retention and whether retention is enhanced by feedback. They tested participants at different times and highlighted the efficiency of test-taking as a studying strategy, which was superior to that of traditional reading and restudying (22%, 32%, and up to a 43% difference between traditional studying and test-taking studying without feedback, test-taking studying with immediate feedback, and test-taking studying with delayed feedback, respectively). Since feedback has a clear impact on learning, especially when coupled with active recall strategies, non–AI-mediated active learning strategies could be limited by the lack of opportunities to offer feedback. Feedback generally comes from a reliable source, such as a teacher or an expert, or is obtained through an appropriate literature search, which can be time consuming. Sometimes, it is not possible to have the timely intervention of a teacher or an expert if there is a lot to study, and it may not be possible for a student to conduct a proper literature search if the student is new to a given topic. Furthermore, there are different types of feedback; some feedback is self-directed (ie, obtained through an introspective process). Feedback in the learning process can be used to enhance or develop skills for setting goals, monitoring one’s own learning process, and assimilating input (feedback) toward enhancing performance [43]. All of these are important skills to have when one attempts to obtain feedback on their own, such as when using LLMs for feedback.

AI can enhance medical education by offering feedback and explanations to clarify incorrect responses, thereby increasing study efficiency. By using AI tools like ChatGPT, students can receive detailed feedback on their answers, including the identification of errors and the provision of correct information, as we have shown. ChatGPT presents promising implications in providing technically accurate medical feedback, given the exceptional knowledge it has exhibited, as we previously described. This process, however, should not replace thorough literature research or foundational knowledge acquisition. AI models can also serve as tutors to facilitate discussions on specific knowledge areas, similar to existing models in other fields, such as Khan Academy’s Khanmigo, which serves as a fully personalized tutor [44]. One limitation of AI systems like ChatGPT is their character limit for inputs, which can be managed by breaking text into sections or using multiple prompts. Additionally, web searching is only available with ChatGPT’s paid subscription; for the free version, one should provide ChatGPT with the reference text by copying and pasting it. Further research is needed to explore the potential of AI-assisted tutors in medical education, especially in education on basic subjects.

Memory cues can be used as effective learning tools that ease the studying experience for a student attempting not only memorization but also mastery of complex concepts and new vocabulary. Evidence has described the superiority of self-generated memory cues over cues created by others [30,31]. The explanations behind the superiority of self-generated memory cues are (1) the generation effect of creating such cues, that is, the act of generation requires significant cognitive effort, which boosts memory, and (2) the cue selection process itself and the consequential metamnemonic effect. This means that students who identify the learning formats that work best for them are able to create their own memory cues by using a modality that is tailored to their needs [31]. For example, if a student prefers to have a visual representation of the ideas that they are attempting to memorize, they might lean toward the creation of mnemonics that create a mental picture to integrate information. Tullis and Fraundorf [31] proposed evidence that the benefits of self-generated cues come in great part from the correct selection of a cue from a list of candidates. If students can create multiple cues, they can, with greater effectiveness, select the cue that best benefits retrieval. Tullis and Fraundorf [31] further suggested that allowing a learner to select from multiple options of cues requires less cognitive work, takes less time, and may not hinder memorization.

As we described, the act of generation is effortful and may be time consuming. AI tools like ChatGPT can help students as a result of their seemingly tireless and effortless generative capacity. In addition, these tools can create multiple cues with different modalities (textual-based cues and pictorial cues) when prompted to do so, thereby allowing learners to focus on understanding the material and selecting the most appropriate cue that fits their educational needs. The downside to the use of this method is that multiple attempts may be required for ChatGPT to produce a mnemonic that is subjectively good or fitting enough for a particular student. In addition, it is our opinion that these tools are best used when the user has an idea of what they should learn or memorize, and the user should prompt the AI tool to create a mnemonic device that facilitates the recall of the information they wish to encode. This is because there is abundant evidence of ChatGPT not only making errors but also blatantly providing false information [45], which is known as “artificial hallucination.”

Previous research has explored the effectiveness of visual mnemonics in improving learning outcomes. An experimental study that compared pictorial mnemonic use to traditional study methods found that pictorial mnemonics aid in learning from text passages by improving the recall of factual knowledge and long-term memory retention in college students [46]. Additionally, a randomized trial compared audiovisual mnemonics against traditional text-based learning for retaining medical knowledge; participants who used mnemonics demonstrated significant improvements in free-recall tests, with scores improving by 65%, 161%, and 208% immediately, after 1 week, and after 1 month, respectively, when compared to those who used text materials (P<.001). Moreover, the group that used mnemonics outperformed the group that used text materials by 55% in a 1-week–delayed multiple-choice test that focused on higher-order thinking (P<.001) [47]. In a comparative study of visual mnemonics versus traditional lectures for learning the porphyrin pathway, there was no significant difference in quiz scores immediately or 1 week after the intervention; however, the mnemonic group exhibited a 20% higher score 3 weeks later (P=.02) [48]. In another randomized trial that compared story-based audiovisual mnemonics with traditional text reading for memory retention among medical students, the audiovisual mnemonics group demonstrated significantly better performance in multiple-choice tests immediately after the intervention (P=.04), as well as at 1 week, 2 weeks, and 4 weeks after the intervention [49]. These results underscore story-based mnemonics’ superior effectiveness in enhancing immediate and long-term memory retention in medical education. Although there is some variation in the visual mnemonic techniques across studies (eg, the studies by Yang et al [47] and Abdalla et al [49] used some audiovisual mnemonics), they consistently demonstrated that factual knowledge can be represented visually and that the use of this type of mnemonic enhances both the recall and long-term retention of knowledge, with large effect sizes.

The visual mnemonic proposed in our study highly resembles the strategy used in the experiment by Rummel et al [46], in which visual mnemonics were created from texts about psychologists, incorporating elements for recalling both the psychologists’ names and the key aspects of their theories. In our mnemonic, “long hair” aids in recalling the name of the disease (hairy cell leukemia), and the other elements in the image are used to help recall the disease’s main features. The Picmonic System, which uses mnemonics from a web-based educational platform [50] that was used in the studies by Yang et al [47] and Abdalla et al [49], also adopts the visual mnemonic approach by combining visual elements and storytelling to enhance the recall of information; this is also highly similar to our approach. Thus, using DALL·E 3 for mnemonic generation shows promise for improving different learning outcomes, such as test performance, long-term retention, and free recall. Future studies should experimentally investigate the effectiveness of visual mnemonics generated by text-to-image models in learning processes. A significant limitation of using DALL·E 3 for medical mnemonic generation is its restriction on explicit content, prohibiting prompts with terms like “blood.” By recognizing this limitation, knowledge area–specific text-to-image models can be developed to more accurately describe the information needed and enable the use of words that are commonly used in a knowledge area but are censored in current models. Another limitation is that creating stories that accurately reflect the intended factual knowledge for mnemonic cues can be complex, particularly for certain subjects. Effective prompt engineering techniques could help in creating more relevant and coherent visual mnemonics.