ChatGPT in Nuclear Medicine Education

DISCUSSION

The performance of ChatGPT in nuclear medicine assessment was enlightening. ChatGPT performed well when brief and shallow insights were required, typical perhaps of first-year subject topics and early or general aspects of second-year subject topics. Third-year topics are far from the shallows, and as a result, the expectations of depth and insight were beyond the capabilities of ChatGPT, even when prompted by a specific word count. Importantly from an education perspective, ChatGPT provided answers with no evidence, with outdated evidence, or with fabricated evidence. Such answers not only are devoid of current insight into the fluid nuclear medicine environment but also would represent academic misconduct if used by students in their submissions. ChatGPT was particularly poor for calculations, despite providing convincing working and justification for the same. Our findings are consistent with those reported for medical examinations, for which ChatGPT scored 43%–68% in open-ended questions and 40%–65% in multiple-choice questions (7,8). The authors similarly reported lower scores correlating with more complex questions.

Alarmingly, one of the chief benefits (and risks) of ChatGPT is in written tasks, for which ChatGPT was shown to perform well short of expectations across all levels and courses. The depth of research and writing, the insight and understanding demonstrated in the writing, and the writing style itself (e.g., professional language and tone, integration across the piece, and integration with practice) not only penalized scoring but could not be used to positively impact skill development in students. This shortfall raises serious questions about using ChatGPT to generate questions (revision or assessment) when it lacks the insight to answer them. The performance of ChatGPT on very general topics or at a lower level of education (e.g., high school) might be better, and in the current evaluation, ChatGPT performed well on topics requiring shallow information and on topic that were widely evidenced before September 2021 (e.g., the role of bone scans in prostate cancer). Nonetheless, nuclear medicine education requires depth and specific detail in a rapidly evolving domain that confounds ChatGPT. It is possible for students to plan sections or topics within a written response and ask ChatGPT more targeted questions to produce a higher-scoring paper; however, it will remain constrained by lack of depth and insight, language that is less than professional, incorrect information, and inadequate or fraudulent referencing. Regardless of well-constructed responses or poorly constructed responses, Turnitin software confidently predicted when responses were AI-generated.

In the hands of a student already performing at a passing level or better, ChatGPT is unlikely to boost grades; indeed, it may reduce grades and risk academic misconduct. In theory, when norm-based referenced testing is used, the class mean improves and those using ChatGPT have an increased representation in the higher grades, potentially relegating non-ChatGPT users to the lower grades. For criterion-based referenced testing, it could allow student performance to improve independently of the class performance. Those using ChatGPT are advantaged but not at the expense of the grades of those not using ChatGPT. In reality, the current version of ChatGPT (GPT 3.5) does not provide that capability, and students relying on ChatGPT for either cheating or enhancing responses are likely to be penalized in their scores, independently of academic misconduct issues.

On the basis of the evidence in this evaluation, ChatGPT does not pose a risk of masking the students’ shortcomings against the learning outcomes. Among nuclear medicine students, ChatGPT also does not appear to enhance grades by honing skills in writing and expanded learning. The significant risk to academic integrity is raised by poor, absent, or fraudulent referencing in ChatGPT responses. ChatGPT does not raise concerns about students graduating without the requisite knowledge and skills for safe clinical practice because reliance on ChatGPT will not allow the student to accrue a passing grade. It is appropriate to reevaluate the ChatGPT claimed benefits to student learning and assess whether these are conceivable (Table 1). Among the potential roles of ChatGPT for nuclear medicine student learning, the following would likely be the most appropriate: use of ChatGPT as a language assistant or language practice tool for those for whom English is a second language (for this role, ChatGPT would be developing basic English language skills, not professional language or writing skills); use of ChatGPT to support accessibility for students with disabilities, such as through assistive technology for text-to-voice conversion; use of ChatGPT as a training tool in which ChatGPT responses are applied in clinical, theory, or research learning domains to have students critique, refine, and correct; use of ChatGPT to proof, test, and refine assessments and scoring rubrics as part of moderation; and use of ChatGPT to simulate conversations with patients or caregivers as a form of authentic assessment.

Academic misconduct concerns for ChatGPT have been related largely to the potential capability to generate examination answers or responses to written tasks; that is, cheating for advantage. For nuclear medicine subjects, the risk of this use appears to be low, with limitations to the ChatGPT capability. This use could be further limited by structuring assessments that target student insight and understanding at a deeper level and by setting the minimum standards for a passing score in the rubric more rigorously against learning outcomes. That is, if a learning outcome for a subject requires students to be able to demonstrate their understanding by explaining a concept, then a student response that does not explain that concept (lists or outlines key information or facts) or does not show understanding (has errors or lacks integration with practice) should receive a failing score for that question or task. Typically in a rubric, there is some wiggle room that would recognize some knowledge and produce a passing grade with a credit, perhaps representing what should be the minimum standard (Table 2). ChatGPT will be able to produce those passing grades in which the credit expectations in a rubric represent what should be a minimum passing-grade requirement. Adjusting this approach would minimize the risk that ChatGPT will be used to navigate through subject assessment and would better align assessment with learning outcomes. Perhaps the biggest academic integrity issue for ChatGPT use is the potential plagiarism or fraud that students using generated responses could confront. Writing and referencing are typically below the expected standards; a paucity of citations would represent plagiarism, and the tendency for ChatGPT to add citations that cannot be verified is potential fraud. Indeed, ChatGPT can simply fabricate answers.

Although ChatGPT has surprising accuracy for some topics, it is prone to several different types of errors, and these were apparent throughout this investigation when the specific nature of nuclear medicine confounded ChatGPT responses. The term hallucination has been used widely in AI to refer to false or misleading information, yet the term is more specific, referring to a plausible response that is incorrect (it seems correct to ChatGPT but is not—identifying a stick that is not there) (8). Other types of AI errors have also been named according to this original psychiatry analogy:

• Illusion is like a hallucination except it is an error of similarity (mixing up similar items)—mistaking a piece of rope for a stick.

• Delusion is false belief or error (wrong information)—after correction and examination, insisting the rope is a stick.

• Delirium is either a sophisticated or a nonsensical answer because the algorithm is overwhelmed or confused—describing a ball when asked to describe a stick.

• Confabulation or lying is fabrication of information—photoshopping images of sticks in a scene to provide evidence that sticks were there without actually checking the scene for real sticks.

• Extrapolation or interpolation is a logical, although incorrect, extension of known information—based on 3 dogs carrying sticks, declaring a fourth dog also is carrying a stick without seeing the dog.

• Miscalculation or blunder is a computational error despite the correct equation and data—there are 5 sticks but 6 are counted.

All of these error types were evidenced through the scoring of examinations and written assignments. Students leaning on ChatGPT are ill-equipped to identify such errors. These students risk undermining their understanding and ongoing learning. Information is not knowledge.

The GPT 4–powered chatbot could be a bigger problem, with easier, faster, and more accurate responses. GPT 4 allows voice-to-text conversion, which would enable students to simply read the question to ChatGPT. GPT 4 will allow importing of images for analysis and interpretation, a task that would previously have confounded ChatGPT. ChatGPT is prone to hallucinations (false or misleading information), which a student using ChatGPT will not be aware enough to correct. GPT 4 is 60% more accurate with answers and, in particular, with interpreting context and reasoning. GPT 4 is trained on 3 times more data with a 500-fold increase in capacity, which will allow greater originality and accuracy of responses, confounding even the best plagiarism and AI detection software.

For nuclear medicine courses, there appears to be no advantage to students who misuse ChatGPT; indeed, they will be disadvantaged. It is important, therefore, to educate students about use, misuse, risk versus absence of benefit, professional responsibilities to their future patients, and the consequences, now or in the future, of cheating (e.g., if new technology is developed in 10 y that allows retrospective detection of ChatGPT, degree disqualification and deregistration could and should be a consequence). Education of students about ChatGPT would allow it to be integrated into the learning environment to support students when appropriate and to be used as a learning tool. This use is best supported by reengineering assessments and by recrafting learning outcomes to be both authentic and capability-focused, independently of whether ChatGPT is used.

Although not available at the time of writing, Google plans to release its AI chatbot, Bard, which would be in competition with ChatGPT. Although the chief comparisons relate to their chatbot functions for Google and Bing search engines, respectively, both have competing capabilities as a text generator. It is unrealistic to compare Bard with GPT 3, with Bard having capabilities mirroring the recently released GPT 4; these include context interpretation, image analysis, and mathematic problem solving. There are similar issues with accuracy and bias that need to be interrogated in the public user arena. Unlike GPT 4, Bard lacks plagiarism detection or prevention, accesses the Internet in real time, and updates the corpus of knowledge, providing currency at the expense of increased misinformation and bias. It is a reasonable prediction that ChatGPT will be the preferred tool for generative text for academic and scientific writing whereas Bard may emerge with broader applications in text generation (e.g., list generation, agenda production, and scheduling tasks) and image or video creativity for personal and general professional purposes.