Abstract
The purpose of this literature review is to provide an in-depth analysis of the transformative potential of artificial intelligence tools in the elementary classroom. It explores the rapid evolution of technology in classrooms, from the minimal use in the late 1980s to the integration of advanced devices like Chromebooks and smartphones, and the challenges these pose for teachers unprepared by traditional training programs. The review delves into the intricacies of Large Language Models (LLMs) and Natural Language Processing (NLP), underlining their critical role in modern AI tools used in education. It examines the current use of AI in elementary education, emphasizing the need for updated research given the fast-paced advancements in AI technology. Key themes such as AI’s role in enhancing learner motivation, providing timely feedback, and facilitating customization and individualized learning are discussed. The review also touches upon the use of AI for students with special needs, underscoring AI’s capacity to support diverse learning requirements. Through a comprehensive analysis of existing literature and case studies, the review critically examines the relevance of previous research in light of the latest AI developments and identifies gaps that need addressing. This review ultimately underscores the significant potential of AI tools in revolutionizing elementary education by catering to individual learning styles and increasing student engagement, while also acknowledging the need for continuous research and adaptation in teacher education programs.
Introduction
While the field of education is forced to constantly navigate evolving technologies and the impact they have on teaching and learning, in recent years it has become increasingly more challenging to do so. As the rate at which these technologies are being released and what they are capable of accomplishing is nothing short of astonishing. In classrooms of the late 1980’s and early 1990’s one may have seen the singular computer sitting on the teacher’s desk, or gathering dust in the back of the classroom. Fast forward to 2011, when Chromebooks, inexpensive, fast, and nimble devices designed to navigate the web alongside other education focused tasks, began to infiltrate classrooms across the world (Ackerman, 2021). Paring these tools next to smartphones, the learning curve for educators became quite steep quite quickly. In addition to this, most teacher preparation programs only require one course on technology integration and most teachers have little to no formal computer science education in their pre-service teaching program at all (Ozogul et al., 2018).
The lack of understanding on how to integrate these technologies, specifically smartphones, into the classroom setting as learning and instructional tools has intensified even more recently as the third largest school district in the country has recently completely banned students from using their phones throughout the duration of the school day (Singer, 2023). This reaction to student phone use mirrors the initial response of the New York City Public School System, in which they decided to put a complete ban in place on the newly released, artificial intelligence (AI) model, ChatGPT, which was growing in both popularity and notoriety worldwide in January of 2023 (Elsen, 2023). Since the initial response, this ban has been lifted, but many questions remain amongst educators as to how artificial intelligence technology can effectively be used in the K-12 classroom, and if so, what does that look like?
This literature review will analyze potential opportunities for teachers at the elementary level to leverage AI tools in their classroom instruction in order to support individual learning needs and will explore ways in which these tools might be used to create higher levels of student engagement. This review will analyze and define: large language models and natural language processing, the current state of AI tools being utilized in the elementary classroom, and finally the impact these tools have on addressing the needs of individual learners.
Large Language Models: A Definition
In order to understand tools such as ChatGPT, it is imperative to have background knowledge on the way in which they were engineered and part of that is defining the models upon which they are built and how they function. Many of the current AI tools emerging are being developed on large language models (LLM). These programs are essentially what the name states. They are fed, or input, large quantities of data, and in this case text. Earlier AI models had more of a focus on understanding humans. LLM’s however, utilize this text to develop an understanding of patterns in which the data presents itself in order to replicate, or generate, responses that resemble that of a human (What Are Large Language Models?, 2023). Beyond this text generation feature, these tools can also be used to translate text into a variety of languages, summarize and classify data and information.
Natural Language Processing: A Definition
Another key component of many AI applications currently being pushed to the public, is known by the term, Natural Language Processing (NLP). To understand the functionality of an AI that functions with NLP, it is necessary to have working knowledge of why NLP is so important. Many of us have been interacting with technology for more than a decade that has been embedded with NLP abilities. Anytime one uses a smartphone or smart device and verbally provides this device with commands which the device can interpret and translate into an executable command, they are utilizing the natural language processing feature that is housed within their device.
Essentially, what is occurring is the device is able to act as a receiver of information, which then decodes the information into an executable function. Once the command is carried out, the human user should be encountered with some form of output or response. NLP does not only refer to verbal commands, and it can in fact work in reverse. One may see someone interacting with an online chatbot in which the chatbot is providing text-based content and the human user then will need to receive, decode, and determine the appropriate application for the information that was received. Figure 1 below demonstrates the linguistic steps that occur through natural language processing (Sanagapati, 2020).
Figure 1
In addition to defining these terms, it is important to note that an AI model that is an LLM, can also have NLP capabilities. These are not exclusive to one single form of AI, and many AI models will incorporate a plethora of other features such as GANs generative adversarial networks. GANs are dual purpose AI which have the ability to generate, and also discriminate. They can generate content, but then they are also able to use their discriminative functionality to evaluate the quality of the output that was generated. In addition to GANs, there are AI tools available that are considered multimodal. This refers to the AI’s ability to interpret multiple forms of input allowing it to solve more complex tasks (Dwivedi, 2022).
All of this information is important in order to have the appropriate background knowledge to, not only understand the research that is being carried out currently involving AI, but also to recognize, and appreciate the rapidity with which the field is being molded, shaped, destroyed, and reshaped again. One year ago at this time, ChatGPT was just taking off as a celebrated chatbot. Two weeks ago, with its most recent update, ChatGPT is now a multimodal, natural language processing, large language model, generative, pre-trained transformer, artificial intelligence tool that has advanced so far, that you no longer need to understand any form of coding to create an individualized version of itself. Anyone can do so now. Everyone now has the power to customize this tool in order to assist and generate content, responses, or feedback in any specific area of interest. This brings us to one of the conundrums regarding the current state of research in the field. As there is a solid body of knowledge from which to glean information from, however much of this research was conducted prior to the inundation which has occurred over the past 12 months with the release of the various LLMs and other generative AI tools. Therefore, there is an even more urgent need that should be placed in furthering our understanding of how the current generation of AI tools impact teaching and learning.
Current Knowledge Claims/Themes
As we continue to analyze artificial intelligence and new methods of implementation emerge, it is imperative to have an understanding of what the existing body of knowledge has already taught us. In familiarizing oneself with these case studies, one must now critically ask the question as to whether or not the results presented from studies carried out prior to this AI emergence still hold as much validity today as they did when carried out. If there is lingering doubt as to their value since the currently being used were not available at the time of these studies, one may potentially consider the possibility of replicating one of these studies in order to create a scenario of comparison. This would allow them to not only determine the impact of AI tools currently available, but also to conduct a comparative analysis between the two studies.
For example, had a research study been carried out in the year 2018 that focused on the research question, “How is AI perceived and utilized in the elementary school classroom?” the responses elicited would appear completely different than if the question were asked today. One influence that should not be overlooked in this case is the impact of mass and social media as each of these mediums have been flooded with information about artificial intelligence tools, something that, although often superficial, still builds some understanding in the collective audience.
AI and Learner Motivation
It may be possible that the number one factor that artificial intelligence currently has in its favor in terms of potential classroom use, is that of learner motivation. First, learner motivation can often be increased through topics of interest to students, timely feedback, as well as curiosity. Do not be confused here, the AI is not curious. The teachers and the students are curious. Though some may be fearful, there is often still an underlying curiosity as to the capabilities of AI and what it may help one to accomplish. However, this sense of wonder can only go so far as the novelty of any technology can potentially wear off. Therefore, learning with and utilizing this technology must be embedded within structured curriculum and instructional design (Lin et al., 2021).
While we have focused predominantly on LLMs throughout this analysis, there are additional ways in which educators have been known to employ AI technology in order to increase learner motivation. While curiosity is one factor that increases learner motivation, feedback is another key area that can help drive learners to deeper levels of learning and content engagement. As far back as 2018, studies were conducted to determine the impact of quality, timely feedback from AIs in large scale educational settings known as MOOCs. As students participate in these major open online courses (MOOC), they often struggle to maintain their motivation to learn throughout the duration of the course. While there are a plethora of factors that could impact the level of drive demonstrated by students in these settings, such as the quality of instruction and the lack of opportunities for social interaction, one factor often cited is the lack of quality and timely feedback due to the vast number of students that one instructor is responsible for. To better understand the impact feedback has on students in this setting, Goel and Polepeddi deployed an AI tool called Jill Watson to provide discussion board feedback to their students. Because this was an AI, the responses, though simple, were received with very short turnaround time (2018). This timely and responsive feedback encouraged learners to continuously engage in the conversation, leading to higher levels of overall engagement and motivation to finish the course.
While the outcome of this study conclusively demonstrates that feedback leads to increased and sustained motivation, one must still ask certain questions to determine whether the feedback provided through AI is as meaningful to that received from an instructor. Is it possible that students who know their feedback is being derived through an artificial intelligence tool may initially view this as a novelty that could potentially wear off over time? Would learners eventually feel as though they are missing out on the humanity aspect attributed to instructor generated feedback? Would it be possible to replicate this study in a way that would allow for a placebo test to occur in which one group of students was provided feedback from an AI tool, and another from their human instructor? Whatever the outcome, we know that AI can be leveraged to provide timely feedback and at least for now, the curiosity that AI drives can lead to higher levels of learner motivation.
Customization and Individualized Learning
In addition to learner motivation, it’s also important to consider an additional subfactor that can lead to increased learner motivation, however it should be discussed in isolation as well as this subfactor speaks to the potential of AI use in a variety of areas. One challenge that many educators, especially at the elementary level are faced with on a daily basis, is that their students come to them with a wide range of abilities and experiences. While diversity in nature is good, this can be challenging to address the needs of a large group of students with the small amount of preparation time that teachers have. With the utilization of AI technology, particularly LLMs, teachers now have the opportunity to produce individualized learning opportunities and experiences that meet the needs of their individual students.
It has been well-documented that one of the main challenges posed by student-centered learning methods such as project-, problem-, and inquiry-based learning is that students often lack the background knowledge to connect with material in their specific challenges. Simultaneously, it can be difficult for instructors to connect content to students’ lives making it both relevant and accessible to students (Nariman & Chrispeels, 2016). AI can help address this in a systematic way by allowing the students and the instructor to collaborate using an AI tool that would produce specific content the student could engage with, allowing them to efficiently fill in knowledge gaps that were previously identified.
Leveraging AI to create customized pieces of content for students does not have to end there though. Teachers at the elementary level are also tasked, especially at the lower grades, with teaching students to learn to read, something more easily said than done. LLMs can, “assist in the developing of reading and writing skills (e.g., by suggesting syntactic and grammatical corrections), as well as in the development of writing style” (Kasneci et al., 2023, p. 2).
With recent updates to LLMs such as ChatGPT, the tool has now become what is known as multimodal, a term here meaning the application can now accept multiple forms of input such as .pdf documents, standard text, images, spreadsheets, and with the inclusion of certain plugins and extensions, video content as well. In addition to this, it can also generate multiple mediums as well. While the text generation has been discussed in great detail to this point, it can also generate realistic images through a technology known as generative adversarial network, or GAN.
Essentially, a GAN works as two separate entities. The first function is that of a forger or a generator which produces forgeries of images. The second function is that of a discriminator. The discriminator challenges the content that is generated by the forger, comparing it to real or authentic images. The idea is that working simultaneously in competition with one another, the discriminator will challenge the generator to create something realistic, a term here that is simply put, relative to the prompt information that has been input by the subject. Why are we discussing GANs at this point in this literature review? Because, students do not simply learn through text. Many learners build skills and understanding through graphical representations and videos as well. These new updates allow teachers to generate images and in some cases videos as well that can be used to customize the learning experience for their students.
While these tools are readily available to many educators across the globe now, it would also behoove schools to get them into the hands of their students as well and educate them on best practices. This would allow students to have at their fingertips a tool that would allow them to ask questions, provide guidance and feedback, challenge their critical thinking skills, formulate high level questions, and provide answers. Essentially, the tool becomes a teaching assistant in the classroom. While there has been considerable conversation in the field of education amongst teachers as to whether their role as the teacher will still be necessary in the future as AI flows in from society to the classroom, this will never be the case. The nuances that each student brings to the classroom does not lend itself to something that can be successfully navigated by a machine, and teaching requires the flexibility to maneuver between a variety of student personalities, life experiences, and emotional challenges and for these reasons these fears, while appreciated, are unwarranted (Douali et al., 2022).
Support, Remediation and Students with Special Educational Needs
As previously mentioned, students at the elementary level arrive on the first day of school from a variety of backgrounds, and oftentimes they present themselves with a number of learning needs that can be challenging for one single classroom teacher to meet. In general, disabilities can be classified into four main categories, mobility impairment, hearing impairment, visual impairment, and cognitive impairment. Each category previously listed can be at least partially addressed with the use of AI to better meet the needs of students with special educational needs (Garg & Sharma, 2020).
Although the main AI focus of this review to this point has been on LLMs and GANs, it is important to understand that these are simply the trending tools of the times. AI can take the form of other tools and models as well. For example, in order to identify students presenting with dyslexic tendencies, machine learning algorithms and software can actually be used to accomplish this. When students present with hearing impairments, common software suites such as Microsoft Translator can be used to assist students in converting speech to text, allowing them to communicate more effectively with their peers and instructors (Garg & Sharma, 2020). Finally, in order to assist learners with visual impairments, students can utilize screen readers, and AI powered braille converters, effectively making the idea of an inclusive classroom a reality.
Current Gaps in the Field
Though there is a plethora of research that has already been done connecting artificial intelligence and machine learning to the field of education, as previously mentioned, there is a problem with this information. One might erroneously state that it is irrelevant. While much of the current data that has been collected has been presented in a way that supports our growing body of knowledge in understanding how AI can be used in classrooms, the fact remains that things have changed, and they have changed dramatically.
I am certain that I am not meant to do this, but I would like to speak from the heart for a moment and share some personal thoughts about educators. This conversation is necessary at this juncture in this review as it clearly explains why the current body of research that we have regarding artificial intelligence in education is outdated. Elementary teachers always have and always will take on all challenges presented to them and address them with the care, kindness, and thoughtfulness they deserve and creativity that can not be understood by those outside the field. They will come up with methods of maximizing the potential for these tools in ways that cannot be quantified through any form of research methodology. The bond they form with their students will spark conversations and inspirational ideas that with the integration of these technologies will solve problems most would have assumed could not have been addressed. For these reasons, educators should have no fear regarding their craft. Also, for these reasons, the body of knowledge must be updated, because what we are looking at right now, is simply one large gap in our collective understanding.
That being said, the ensuing conversation will focus predominantly on studies conducted more recently. One such study was conducted to determine the value of formative and summative quizzes generated using an end-to-end GPT-3 model titled EduQuiz. In this study, the researchers found that EduQuiz was able to generate quality assessment items, effectively limiting the amount of time an instructor would spend on developing these themselves. In the discussion of this study however, the researchers, though optimistic by the overall results, still conveyed the need to vet the information generated for the assessments as there remained instances of questionable content generated (Dijkstra et al., 2022). While this study focused solely on content developed for reading comprehension assessments, one could argue that a similar study could be conducted for assessment purposes in virtually every content area. In fact one of the areas in which many individuals have noticed AI struggling most frequently is mathematics, so if teachers intend on leveraging AI to develop assessment items for a math assessment, they should most definitely be on the lookout for potential errors and even hallucinations of the content. In an interview with Ethan Dyer, a machine learning specialist at Google, he stated, “humans give inconsistent answers, make errors, and fail to apply core concepts, too. The borders at this frontier of machine learning, are blurred” (Garisto, 2022).
In another recent study, GPT-3 was used specifically to build higher levels of curiosity among children through challenging the depth of questioning students were providing (Abdelghani et al., 2023). While the authors concluded the overall results presented through the prompt-based method were positive, they also noted the need for a future study to charge the LLMs “to analyze childrens’ questions and give them real-time feedback about their relevance, their divergence level and their syntactic construction” (Abdelghani et al., 2023, p. 29).
This was interesting as this need for feedback reflects what Paul Kim and the Stanford Mobile Inquiry-based Learning Environment or the SMILE Project at Stanford University, are currently exploring with their program. The SMILE project is currently leveraging the ChatGPT API to develop a tool in which learners can type in their questions and receive feedback based on the depth of their question. However in this instance, they will receive feedback in two formats. First, the AI will provide them a scaled rating from 1-5 and verbal feedback as well. Secondary, their classmates have the opportunity to provide them with human generated feedback. This form of hybrid design serves to hold the attention of the learners through challenging them to create the most in-depth question, but it also allows students to recognize the value of multiple forms of feedback, leading to higher levels of learner motivation (SMILE: Stanford Mobile Inquiry-Based Learning Environment | Office of Innovation & Technology, n.d.).
While the SMILE Project and the Abdelghani study demonstrate potential uses for these tools in the realm of student feedback, it’s important to recognize these studies are still quite limiting in potential use cases of LLMs in an educational capacity. With the addition of new LLM and NLP models being released on the regular, it’s necessary to conduct studies that will allow us to better understand their use, application, and effectiveness. A specific example of this is the inclusion of ChatGPT’s API by the non-profit organization, Khan Academy in the development of their new AI tutoring tool, Khanmigo. While there have been news articles written about the impact these tools have had on high school students (Carlson, 2023), and likewise studies have been written about the impact of these tools in the realm of higher education (Vasarhelyi et al., 2023), there has been little research conducted at the elementary level and the impact tools such as this can have on the learning of our youngest students.
Conclusion
The analysis presented in this literature review underscores the imperative need to continuously grow the body of research concerning the application and impact of AI-driven tools, particularly Large Language Models (LLMs), in elementary education. As technology evolves at an astonishing pace, it becomes increasingly crucial to understand how these advancements can be harnessed to enhance teaching and learning processes, foster learner motivation and engagement, and address the unique learning needs of students.
The dynamic nature of AI, especially in the form of LLMs and other emerging technologies like GANs and multimodal AI, offers unprecedented opportunities for personalized education. However, the rapid advancements in these technologies also mean that much of the existing research may soon become outdated or insufficient to fully grasp their potential and implications. There is a pressing need for ongoing, up-to-date research to keep pace with these advancements and to provide educators with the knowledge and tools they need to effectively integrate AI into their teaching practices.
Furthermore, this review has highlighted the diverse applications of AI in supporting students with special needs, emphasizing the role of AI in creating inclusive educational environments. However, to realize the full potential of AI in this regard, more targeted research is needed, particularly in exploring and validating new approaches and tools.
In conclusion, this literature review serves as a call to action for researchers, educators, and policymakers to invest in and prioritize the exploration of AI in education. As we stand at the cusp of a new era in educational technology, it is vital to foster a deeper understanding of how AI can transform elementary education. This involves not only keeping abreast of technological advancements but also actively participating in a dialogue that shapes how these tools are developed and implemented in educational settings. By doing so, we can ensure that AI serves as a powerful ally in the quest to meet the diverse needs of learners and to prepare them for a future where technology and human ingenuity coalesce to create limitless possibilities for learning and growth.
References
Abdelghani, R., Wang, Y.-H., Yuan, X., Wang, T., Lucas, P., Sauzeon, H., & Oudeyer, P.-Y. (2023). GPT-3-driven pedagogical agents for training children’s curious question-asking skills. International Journal of Artificial Intelligence in Education, 1-36.
Ackerman, D. (2021, May 11). The Chromebook at 10: How this ‘browser in a box’ became the perfect pandemic laptop. CNET. Retrieved November 26, 2023, from https://www.cnet.com/tech/computing/the-chromebook-at-10-how-these-stripped-down-computers-went-mainstream/
Carlson, C. (2023, June 05). Brave new world: Hobart students, teachers cite gains with new AI pilot. Chicago Tribune.
Dijkstra, R., Genc, Z., Kayal, S., & Kamps, J. (2022). Reading comprehension quiz generation using generative pre-trained transformers.
Douali, L., Selmaoui, S., & Bouab, W. (2022). Artificial intelligence in education: Fears and Faith. International Journal of Information and Education Technology, 12(7), 650-657.
Dwivedi, D. (2022, March 31). What is multimodal AI and some interesting applications. Medium. https://dwivedy.medium.com/what-is-multimodal-ai-and-some-interesting-applications-d15c32663b8a
Elsen, M. (2023, January 3). NYC bans access to ChatGPT on school computers, networks. Chalkbeat. https://www.chalkbeat.org/newyork/2023/1/3/23537987/nyc-schools-ban-chatgpt-writing-artificial-intelligence/
Garg, S., & Sharma, S. (2020, July). Impact of artificial intelligence in special need education to promote inclusive pedagogy. International Journal of Information and Education Technology, 10(7).
Garisto, D. (2022, June 12). AI langauge modles are struggling to “Get” math > should this be telling us something? IEEE Spectrum. https://spectrum.ieee.org/large-language-models-math
Kasneci, E., Sessler, K., Kuchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Gunnemann, S., Hullermeir, E., Krusche, S., Kutyniok, G., Michaeli, T., Nerdel, C., Pfeffer, J., Poquet, O., Sailer, M., Schmidt, A., Seidel, T., … Kasneci, G. (2023, April 1). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences: Journal of Psychology and Education. https://doi.org/10.1016/j.lindif.2023.102274
Lin, P.-Y., Chai, C.-S., Jong, M. S., Dai, Y., Guo, Y., & Qin, J. (2021). Modeling the structural relationship among primary students’ motivation to learn artificial intelligence. Computers and Education: Artificial Intelligence, 2.
Nariman, N., & Chrispeels, J. (2016). PBL in the era of reform standards: Challenges and benefits perceived by teachers in one elementary school. Interdisciplinary Journal of Problem-Based Learning, 10(1), 1-16. : https://doi.org/10.7771/1541-5015.1521
Ozogul, G., Karlin, M., & Ottenbreit-Leftwich, A. (2018). Preservice teacher computer science preparation: A case study of an undergraduate computer education licensure program. Journal of Technology and Teacher Education, 26(3), 375-409.
Sanagapati, P. (2020). Knowledge Graph & NLP Tutorial-(BERT,spaCy,NLTK). Kaggle. Retrieved November 29, 2023, from https://www.kaggle.com/code/pavansanagapati/knowledge-graph-nlp-tutorial-bert-spacy-nltk
Singer, N. (2023, November 6). This Florida school district banned cellphones. Here’s what happened. The New York Times. https://www.nytimes.com/2023/10/31/technology/florida-school-cellphone-tiktok-ban.html
SMILE: Stanford Mobile Inquiry-based Learning Environment | Office of Innovation & Technology. (n.d.). Office of Innovation & Technology. Retrieved December 3, 2023, from https://gse-it.stanford.edu/smile
Vasarhelyi, M. A., Moffitt, K. C., Steward, T., & Sutherland, D. (2023, August). Large language models: An engineering technology in accounting. Journal of Emerging Technologies in Accounting, 20(2), 1-10.
What are Large Language Models? (2023). NVIDIA. Retrieved November 26, 2023, from https://www.nvidia.com/en-us/glossary/data-science/large-language-models/
Leave a Reply