General purpose LLMs have revolutionized language comprehension and generation capabilities in the field of AI, emerging as powerful systems that enable highly versatile applications. At the core of these models lies the transformer architecture, introduced by Vaswani and colleagues, which is based on a self-attention mechanism. Unlike traditional sequential models, this architecture enables parallel data processing and effectively captures long-range dependencies, thereby introducing a new paradigm in NLP.

The evolution of OpenAI’s GPT series—from GPT-1 to GPT-4—has led to remarkable advances in contextual awareness, coherence, and nuanced understanding of language. ^,, Notably, GPT-3.5 and GPT-4 have demonstrated exceptional performance in complex language tasks, highlighting the immense potential of LLMs. A prime example of this progression is ChatGPT, built upon GPT-3, which offers an interaction-optimized structure and has been effectively utilized in medical education and scientific research.

In parallel, Meta AI’s large language model meta AI (LLaMA) models ^, and Google’s T5 and pathways language model (PaLM) series have introduced unified approaches such as text-to-text transfer, delivering robust solutions across diverse tasks. Additionally, the proliferation of open-source models like chat general language model (ChatGLM) and Alpaca has improved accessibility within the LLM community, accelerating the development of innovative applications.

Recently, the incorporation of external knowledge sources, automated feedback loops, and modular system architectures has significantly enhanced the accuracy and reliability of these models. Such advancements are particularly relevant in high-stakes domains like health care and dentistry, where precision is critical. These developments signal a shift from mere text generation to knowledge-based decision support systems. In this context, the evolution of general purpose LLMs represents a strategic inflection point, enabling interdisciplinary use of AI-driven solutions.

Recent advancements in AI technologies have enabled the expansion of LLMs beyond text-based applications to multimodal architectures. These next-generation models go beyond traditional language processing by also interpreting visual and auditory data, thereby offering more comprehensive analytical capabilities across various fields, particularly in health care. Multimodal LLMs such as GPT-4 stand out not only for their success in NLP benchmarks but also for their ability to integrate diverse data types to solve complex problems holistically.

By facilitating the analysis of radiographic images alongside textual input, these models have the potential to enhance diagnostic accuracy and support the development of more personalized and precise treatment plans. However, limitations in access to medical data and ethical considerations present barriers to the widespread integration of these technologies. Consequently, there is a growing need for further scientific research to explore the applicability of multimodal LLMs in dentistry.

GenAI models are capable of producing unstructured data such as text, images, and audio. Typically, they generate content in response to natural language prompts; however, they can also operate without explicit instructions by mimicking the style and structure of the data on which they were trained ( Fig. 2 ). Although the conceptual foundations of this technology date back to the 1950s, significant progress has been made through advancements in deep learning (DL). The flexible architecture of DL has enabled more accurate modeling of complex processes, while the increased availability of large datasets and powerful computing resources has accelerated the development of generative models.

Examples generative AI models.

Among the most commonly used GenAI techniques are variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models.

VAEs are an advanced form of deep generative models known as autoencoders (AEs), capable of learning complex patterns within training data. AEs consist of 2 main components: an encoder and a decoder . The encoder transforms high-dimensional input data into a simplified, lower dimensional representation—referred to as the latent space —while preserving the most meaningful features. The decoder then reconstructs the original data as closely as possible using this latent representation.

However, traditional AEs focus solely on reconstructing the input data and therefore cannot generate new data instances. In contrast, VAEs incorporate probabilistic distributions within the latent space, enabling them to learn a range of possible values for each input. As a result, VAEs can perform both data reconstruction and the generation of new, synthetic data samples.

GANs consist of 2 neural networks: a generator , which produces synthetic data from random noise (typically Gaussian), and a discriminator , which evaluates whether the generated data are real or fake. The generator aims to create data that resemble real samples, while the discriminator compares the generated data to actual data and classifies it accordingly. Through this adversarial training process—where the 2 networks are pitted against each other—the generator progressively improves, eventually producing outputs that closely mimic real data.

In recent years, diffusion models have attracted significant attention in the field of GenAI due to their ability to produce more realistic and coherent outputs compared to GANs and VAEs. These models generate data through a 2 phase process known as the forward pass and the backward pass . In the forward pass, real images are gradually degraded by iteratively adding random Gaussian noise until the image is completely distorted. In the backward pass, the model denoises the corrupted image step by step, reconstructing it into a coherent and realistic representation.

During this denoising process, the model learns the pixel-level dependencies within the image, enabling it to both remove noise and accurately recreate the underlying structure. As a result, diffusion models can not only reconstruct training data but also generate entirely new synthetic images based on user-defined prompts.

GenAI enhances the learning experience in dental education by enabling students to simulate complex clinical scenarios and receive rapid feedback. This supports the development of clinical decision-making skills. New strategies powered by GenAI range from the creation of diverse learning experiences to the generation of realistic simulations based on clinical variability.

The personalization of learning processes enables the structuring of educational content in alignment with students’ diverse cognitive profiles and learning preferences. Adapting instruction to individual needs allows personalized learning approaches to enhance student engagement, support motivation, and improve academic achievement.

One study reported that AI-supported learning platforms can dynamically adjust the difficulty level of lessons based on students’ previous interactions. This capability enables high-achieving students to access more advanced content while providing additional support to those facing learning challenges. The results revealed a significant 25% improvement in grades among students using the AI platform compared to those in traditional learning environments.

In a mixed-methods study conducted by Kavadella and colleagues, dental students who used ChatGPT for their assignments outperformed their peers—who relied on conventional literature review methods—on a knowledge-based assessment.

Research at Meharry Medical College explored the integration of ChatGPT into dental education, highlighting both the potential benefits and challenges of incorporating this AI tool into academic settings. The study evaluated various questions derived from the dental school curriculum and found that ChatGPT could play a supportive role, particularly in academic writing and content development.

Thorat and colleagues emphasized ChatGPT’s value as a significant digital tool in modern dental education, owing to its ability to foster diverse learning experiences and its multilingual capabilities.

Nguyen and colleagues compared different LLM versions in answering multiple-choice questions in dentistry, including both text-based and image-based formats. While recent models such as Copilot, Claude, and ChatGPT achieved high accuracy rates—exceeding 80%—in text-based questions, their performance dropped to approximately 60% for image-based items. These findings underscore the need for continuous updates and improvements in LLMs to better handle complex and domain-specific challenges.

LLM-based chatbots offer students opportunities to practice clinical scenarios, receive immediate feedback, and reinforce theoretic knowledge. Student–LLM interaction contributes to the development of digital health literacy by fostering critical thinking, particularly in diagnostic reasoning, treatment planning, and clinical decision-making. In this context, chatbots play a significant role not only in information delivery but also in creating active learning environments.

In one study, a virtual patient was created using an AI chatbot to assess the diagnostic skills of fourth-year and fifth-year dental students regarding pulp pathology. The findings indicated that the majority of students was generally satisfied with the interaction. Notably, fifth-year students reported more positive evaluations and higher levels of satisfaction. Students expressed support for integrating such technology into the dental curriculum, emphasizing that it could enhance educational processes and facilitate adaptation to digital innovations.

At the University of Illinois, a custom-developed chatbot was compared with the traditional Blackboard online platform for addressing clinical questions in implantology by predoctoral dental students. While there were no significant differences in perceived educational value between the 2 platforms, the chatbot received higher satisfaction ratings. This study highlighted that integrating chatbot technology into dental clinical education significantly increased student engagement and learning outcomes, demonstrating the strong potential of such digital tools to enrich the future of dental education.

Dental education requires students to accurately learn complex anatomic structures and effectively apply this knowledge in clinical practice. In this context, the quality and interactivity of visual materials are critical factors that directly influence learning outcomes. GenAI technologies—particularly in the field of image generation—go beyond traditional educational resources by enabling the creation of high-resolution, personalized, and dynamic visual content.

In a study aiming to provide a future perspective on dental imaging and growth prediction, GANs were used to generate images that visually represent the transition from primary to permanent dentition using panoramic radiographs. Kokomoto and colleagues employed Progressive Growing GANs to generate new, full-color intraoral synthetic images from real intraoral scans; pediatric dentists were reportedly unable to distinguish the generated images from real ones.

LLMs and integrated AI systems can play a key role in developing educational materials to improve dental students’ abilities to identify pathologic lesions in radiographic images. In this regard, AI applications supported by LLMs can generate synthetic pathologic lesions of varying size, density, and localization for use in training environments. This allows students to engage in hands-on learning through virtual scenarios that closely mimic real cases, thereby enhancing diagnostic proficiency.

Mehandru and colleagues used GANs to produce synthetic panoramic radiographs both with and without radicular cysts to support studies in cyst detection. Another study employed GANs to generate synthetic panoramic radiographs containing sinus pathologies for the diagnosis of maxillary sinusitis. In addition to supporting a more standardized and accessible learning process, GANs hold potential for training AI models that require large datasets and for facilitating the detection of rare pathologies.

Recently developed models, such as DALL·E 2—which integrates GPT-3.5 as the encoder and a diffusion model as the decoder—have demonstrated the capability to generate synthetic medical images. These systems can also manipulate image characteristics such as noise, contrast, or resolution to create visuals tailored to specific medical conditions.

GenAI technologies are driving transformative changes across numerous clinical domains in dentistry by contributing significantly to diagnostic and therapeutic processes. These advanced AI tools enable dentists to assess patients’ health conditions in a more comprehensive and detailed manner, facilitating more accurate diagnoses and the development of personalized treatment plans.

The effectiveness of GenAI and LLM technologies in providing accurate and reliable information is critical for their adoption in clinical decision-making. LLMs can analyze dental images provided as input, identify dental problems, and generate recommended action plans for the dentist. This process illustrates how AI can support clinical decision-making. Fig. 3 exemplifies how an LLM can assist with such analysis.

LLM-based workflow for processing the input image of a tooth decay to produce a versatile output for dentistry applications.

In a study evaluating the clinical decision support potential of LLMs, responses generated by ChatGPT to 243 dental questions were reviewed by 27 experts from various dental specialties. The findings indicated that ChatGPT was generally able to provide accurate and precise responses. Notably, the model performed particularly well in the areas of oral medicine and radiology. However, the study emphasized the importance of human oversight and warned that AI should not replace professional judgment.

Mago and Sharma posed questions to ChatGPT-3 covering anatomic landmarks, various pathologies, and corresponding radiographic findings within the field of oral and maxillofacial radiology. Their analysis showed that the model generally provided effective and supportive information, suggesting its potential as an auxiliary tool when additional insight is needed by oral radiologists.

Rewthamrongsris and colleagues examined the accuracy of LLMs in recommending prophylaxis for infectious endocarditis during dental procedures such as extractions and intraoral surgeries. Using 28 binary clinical questions based on the 2021 American Heart Association guidelines, 7 LLMs were evaluated. The highest accuracy rate was observed in GPT-4o (80%), followed by Gemini 1.5 Pro (78.57%) and Claude 3 Opus (75.71%).

Dental trauma is highly prevalent in pediatric populations, particularly in cases of avulsion where timely intervention is critical. This highlights the need for effective, evidence-based clinical decision support tools in emergency situations. Based on the guidelines of the International Society of Dental Traumatology, a total of 33 questions—including multiple-choice, binary, technical, and patient-based items—were used to compare ChatGPT and Gemini. Pediatric dentists evaluated the answers, and Gemini achieved a statistically significantly higher average score than ChatGPT, though both models showed promising potential.

Özdemir and colleagues evaluated responses from different LLM-based chatbots to questions in the field of restorative dentistry and reported that ChatGPT-4o and Chatsonic produced promising results. Similarly, in a study by Makrygiannakis and colleagues, responses from 4 LLMs—Google’s Bard, OpenAI’s ChatGPT-3.5 and ChatGPT-4, and Microsoft’s Bing—were compared on clinically relevant orthodontic questions. The findings underscored the evidence-based potential of GenAI LLMs in orthodontic care.

However, Eggmann and colleagues cautioned that LLMs like ChatGPT are limited by their training on relatively recent data and are not designed for direct medical consultation. Nonetheless, they acknowledged that LLMs trained on extensive datasets possess significant potential to support clinical workflows.