Embracing the future of Artificial Intelligence in the classroom: the relevance of AI literacy, prompt engineering, and critical thinking in modern education

In the evolving landscape of education, the integration of Artificial Intelligence (AI) represents a transformative shift, stipulating a new era in learning and teaching methodologies. This article delves into the multifaceted role of AI in the classroom, focusing particularly on the primacy of prompt engineering, AI literacy, and the cultivation of critical thinking skills.

The advent of AI in educational settings transcends mere technological advancement, reshaping the educational experience at its core. AI's role extends beyond traditional teaching methods, offering personalized learning experiences and supporting a diverse range of educational needs. It enhances educational processes, developing essential skills such as computational and critical thinking, intricately linked to machine learning and educational robotics. Furthermore, AI has shown significant promise in providing timely interventions for children with special educational needs, enriching both their learning experiences and daily life (Zawacki-Richter et al., 2019). However, integrating AI into education is not without its challenges. It requires a systematic approach that takes into account societal structural conditions. Beyond algorithmic thinking, AI in education demands a focus on creativity and technology fluency to foster innovation and critical thought. This requires a paradigm shift in how education is approached in the AI era, moving beyond traditional methods to embrace more dynamic, interactive, and student-centered learning environments (Chiu et al., 2023).

This article sets the stage for a comprehensive exploration of AI's role in modern education. It underscores the need for an in-depth understanding of prompt engineering methodologies, AI literacy, and critical thinking skills, examining their implications, challenges, and opportunities in shaping the future of education. Whereas previous papers have already hinted at the importance of recognizing the relevance of AI in the classroom and suggested preliminary frameworks (Chan, 2023), the present discussion claims that there are three prime skills necessary for the future of education in an AI-adopted world. These three skills are supplanted with practical application advice and based on the experience of lecturers at a University of Applied Sciences. As such, the present paper is a conceptual discussion of how to best integrate AI in the classroom, focusing on higher education. While this means that it may predominantly be relevant for adult students, it is believed that it may be useful for children as well.

The current paper entails a conceptual discussion about the proper use of AI in terms of the necessary skillset applied. It is based on a two-step approach:

The informal discussions with students and personnel were unstructured and collected where feasible in these early days of AI use to gather a holistic and trustworthy picture as possible about the explicit and implicit attitudes, fears, chances, and general use of the technology. Hence, this included teacher-student discussions in classroom settings with several classes where students were asked to voice their ideas in the plenum and in smaller groups, individual discussions with students during the breaks, lunch talks with professors and teachers, as well as gathering of correspondence about the topic in the meetings that were held at the university. Taken together, this provided enough information to weave together a solid understanding of the present atmosphere concerning attitudes and uses of AI.

The introduction of ChatGPT (to date one of the most powerful AI chatbots by OpenAI) in November 2022 is significantly transforming the landscape of education, marking a new era in how learning is approached and delivered. This advanced AI tool has redefined educational paradigms, offering a level of personalization in learning that was previously unattainable. ChatGPT, with its sophisticated language processing capabilities, is quickly becoming a game-changer in classrooms, to provide tailored educational experiences that cater to the unique needs, strengths, and weaknesses of each student. This shift from traditional, uniform teaching methods to highly individualized learning strategies will most likely signify a major advancement in educational practices (Aristanto et al., 2023). ChatGPT's role in personalizing education is particularly noteworthy. By analyzing student data and employing advanced algorithms, GPT and other Large Language Models (LLMs) can create customized learning experiences, adapting not only to academic requirements but also to each student's learning style, pace, and preferences. This leads to a more dynamic and effective educational environment, where students are actively engaged and involved in their learning journey, rather than being mere passive recipients of information (Steele, 2023). Furthermore, LLMs have shown remarkable potential in supporting students with special needs. They provide specialized tools and resources that cater to diverse learning challenges, making education more accessible and inclusive (Garg & Sharma, 2020). Students who might have found it difficult to keep up in a conventional classroom setting can now benefit from AI’s ability to tailor content and delivery to their specific needs, thereby breaking down barriers to learning and fostering a more inclusive educational atmosphere (Rakap, 2023). In all of this, the integration of language models like GPT into educational systems is not just a mere enhancement but has the potential to become an integral part of modern teaching and learning methodologies. While adapting to this AI-driven approach presents certain challenges, the benefits for students, educators, and the educational system at large are substantial (for in-depth reviews, see Farhi et al., 2023; Fullan et al., 2023; Ottenbreit-Leftwich et al., 2023). ChatGPT in education can be a significant stride towards creating a more personalized, inclusive, and effective learning experience, preparing students not only for current academic challenges but also for the evolving demands of the future.

However, the many precious possibilities in positively transforming the education systems through AI also comes with some downsides. They can be summarized in several points (Adiguzel et al., 2023; Ji et al., 2023; Ng et al., 2023a, 2023b, 2023c; Ng et al., 2023a, 2023b, 2023c):

In order for all parties to be best prepared for using AI in education, based on a case study and a subsequent literature analysis, there are three necessary skills that can remedy these problems, which are AI literacy, knowledge about prompt engineering, and critical thinking. A more detailed analysis of the challenges is discussed, followed by suggestions for practical applications.

The concept of AI literacy emerges as a cornerstone of contemporary learning. In its essence, it deals with the understanding and capability to interact effectively with AI technology. It encompasses not just the technical know-how but also an awareness of the ethical and societal implications of AI. In the modern classroom, AI literacy goes beyond traditional learning paradigms, equipping students with the skills to navigate and harness the power of AI in various aspects of life and work. It represents a fundamental shift in education, where understanding AI becomes as crucial as reading, writing, and arithmetic (Zhang et al., 2023).

The current state of AI literacy in education reflects a burgeoning field, ripe with potential yet facing the challenges of early adoption. Educators and policymakers are beginning to recognize the importance of AI literacy, integrating it into curriculums and educational strategies (Casal-Otero et al., 2023; Chiu, 2023). However, this integration is in its nascent stages, with schools exploring various approaches to teaching this complex and ever-evolving skillset. The challenge lies in not only imparting technical knowledge but also in fostering a deeper understanding of AI's broader impact – be this on a social, psychological, or even economic level. Due to its importance, there are first AI-Literacy-Scales emerging using questionnaires that can be handed to students (Ng et al., 2023). Although to date there is no stringent consensus on the full scope of the term, it may be argued that AI literacy consists of several sub-skills:

It was shown that early exposure to technology concepts can significantly influence students' career paths and preparedness for the future (Bembridge et al., 2011; Margaryan, 2023). By introducing AI literacy at a young age, students develop a foundational understanding that paves the way for advanced learning and application in later stages of education and professional life. This early adoption of AI literacy is crucial in preparing a generation that is not both adept at using AI as well as capable of innovating and leading in a technology-driven world. This makes the development of AI literacy at schools and universities an important feature of every student. Furthermore, its role extends beyond academic achievement; it is about preparing students for the realities of a future where AI is ubiquitous. In careers spanning from science and engineering to arts and humanities, an understanding of AI will be an invaluable asset, enabling individuals to work alongside AI technologies effectively and ethically. As such, AI literacy is not just an educational objective but a vital life skill for the twenty-first century.

One concrete suggestion is to provide “AI literacy courses” that have the deliberate intent to foster the associated skills in students. In order to have a well-rounded and holistic class, an AI literacy program should entail several key components (Kong et al., 2021; Laupichler et al., 2022; Ng et al., 2023c):

At the moment, there are specific projects that attempt to implement AI literacy at school (Tseng & Yadav, 2023). The deliberate goal is to eventually lead students towards a responsible use of AI, but to do so, they need to understand how one can “talk” to an AI so that it does what it is supposed to. This means that students must become effective prompt engineers.

Prompt engineering, at its core, involves the strategic crafting of inputs to elicit desired responses or behaviors from AI systems. In educational settings, this translates to designing prompts that not only engage students but also challenge them to think critically and creatively. The art of prompt engineering lies in its ability to transform AI from a mere repository of information into an interactive tool that stimulates deeper learning and understanding (cf. Lee et al., 2023). The relevance of prompt engineering in education cannot be overstated. As AI becomes increasingly sophisticated and integrated into learning environments, the ability to effectively communicate with these systems becomes crucial. Prompt engineering empowers educators to guide AI interactions in a way that enhances the educational experience. It allows for the creation of tailored learning scenarios that can adapt to the needs and abilities of individual students, making learning more engaging and effective (Eager & Brunton, 2023). One of the most significant impacts of prompt engineering is its potential to enhance learning experiences and foster critical thinking. By carefully designing prompts, educators can encourage students to approach problems from different perspectives, analyze information critically, and develop solutions creatively. This approach not only deepens their understanding of the subject matter but also hones their critical thinking skills, an essential competency in today’s fast-paced and ever-changing world. As one particular study showed, learning to prompt effectively in the classroom can even help students realize more about the limits of AI, which inevitably fosters their AI literacy (Theophilou et al., 2023). Moreover, AI has the potential to lead to highly interactive and playful teaching settings. With the right programs, it can also be implemented in game-based learning through AI. This combination has the potential to transform traditional learning paradigms, making education more accessible, enjoyable, and impactful (Chen et al., 2023).

Just recently, there are a handful of successful prompting methodologies that have emerged, which are continuously being improved. Prompt engineering is an experimental discipline, meaning that through trial and error, one can slowly progress to create better outputs by revising and molding the input prompts. As a scientific discipline, AI itself can help to find new ways to interact with AI systems. The most relevant prompting methods are summarized in Table 4 and are explained thereafter.

There are two major forms of how a language model can be prompted: (i) Zero-Shot prompts, and (ii) Few-Shot prompts. Zero-Shot prompts are the most intuitive alternative, which most likely all of us predominantly use when interacting with models like ChatGPT. This is when a simple prompt is provided without much further details and then an unspecific response is generated, which is helpful when one deals with broad problems or situations where there is not a lot of data. Few-Shot prompting is a technique where a prompt is enriched with several examples of how the task should be completed. This is helpful in case one deals with a complex query where there are already concrete ideas or data available. As the name suggests, these “shots” can be enumerated (based on Dang et al., 2022; Kojima et al., 2022; Tam, 2023):

These prompting methods have gradually developed and became more complex, starting from Input–Output Prompting all the way to Tree-of-Thought Prompting, which is displayed in Table 4.

When people usually start prompting an AI, they begin with simple prompts, like “Tell me something about…”. As such, the user inserts a simple input prompt and a rather unspecific, generalized output response is generated. The more specific the answer should be, the more concrete and narrow the input prompt should be. These are called Input–Output prompts (IOP) and are the simplest and most common forms of how an AI is prompted (Liu et al., 2021). It has been found that the results turn out to be much better when there is not simply a straight line from the input to the output but when then AI has to insert some reasoning steps (Wei et al., 2023). This is referred to as Chain-of-Thought (CoT) prompting where the machine is asked to explain the reasoning steps that lead to a certain outcome. The framework that historically has worked well is to prompt the AI to provide a solution “step-by-step”. Practically, it is possible to give ChatGPT or any other LLM a task and then simply add: “Do this step-by-step.” Interestingly, experiments have further shown that the results get even better when at first the system is told to “take a deep breath”. Hence, the addendum “Take a deep breath and do it step-by-step” has become a popular addendum to any prompt (Wei et al., 2023). Such general addendums that can be added to any prompt to improve the results are sometimes referred to as a “universal and transferrable prompt suffix”, which is frequently employed as a method to successfully jailbreak an LLM (Zou et al., 2023).

Yet another prompt engineering improvement is the discovery that narrative role plays can yield significantly better results. This means that an LLM is asked to put itself in the shoes of a certain person with a specific role, which then usually helps the model to be much more specific in the answer it provides. Often, this is done via a specific form of role play, known as expert prompting (EP). The idea is that the model should assume the role of an expert (whereas first the role of the expert is explained in detail) and then the result is generated from an expert’s perspective. It has been demonstrated that this is a way to prompt the AI to be a lot more concrete and less vague in its responses (Xu et al., 2023). Building explicitly on CoT-prompting, yet a further improvement was detected in what has come to be known as Self-Consistency (SC) prompting. This one deliberately works with the CoT-phrases like “explain step by step…”, but it adds to this that not only one line of reasoning but multiple of them should be pursued. Since not all of these lines may be equally viable and we may not want to analyze all of them ourselves, the model should extend its reasoning capacity to discern which of these lines makes the most sense in light of a given criterion. The reason for using SC-prompting is to minimize the risk of AI hallucination (meaning that the AI might be inventing things that are not true) and thus to let the model hash out for itself if a generated solution might be potentially wrong or not ideal (Wang et al., 2023). In practice, there may be two ways to enforce self-consistency:

Generalized Self-Consistency: The model should determine itself why one line of reasoning makes the most sense and explain why this is so.

Criteria-based Self-Consistency: The model is provided with specific information (or: criteria) that should be used to evaluate which line of reasoning holds up best.

Sometimes, one may feel a little uncreative, not knowing how to craft a good prompt to guide the machine towards the preferred response. This is here referred to as the prompt-wise tabula-rasa problem, since it feels like one is sitting in front a “white paper” with no clue how to best start. In such cases, there are two prompt techniques helping us out there. One is called the Automatic Prompt Engineer (APE) and the other is known as the Generated Knowledge Prompting (GKn). The APE starts out with one or several examples (of text, music, images, or anything else the model can work with) with the goal to ask the AI which prompts would work best to generate these (Zhou et al., 2023). This is helpful when we already know how a good response would look like but we do not know how to guide the model to this outcome. An example would be: “Here is a love letter from a book that I like. I would like to write something similar to my partner but I don’t know how. Please provide me with some examples of how I could prompt an AI to create a letter in a similar style.” The result is then a list of some initial prompts that can help the user kickstart working on refinements of the preferred prompt so that eventually a letter can be crafted that suits the user’s fancy. This basically hands the hard work of thinking through possible prompts to the computer and relegates the user’s job towards refining the resulting suggestions.

A similar method is known as Generated Knowledge (GKn) prompting, which assumes that it is best to first “set the scene” in which the model can then operate. There are parallels to both EP and APE prompting, where a narrative framework is constructed to act as a reference for the AI to draw its information from but only this time, as in APE, the knowledge is not provided by the human but generated by the machine itself (Liu et al., 2022). An example might be: “Please explain what linguistics tells us how the perfect poem should look like. What are the criteria for this? Can you provide me with three examples?”. Once the stage is set, one can start with the actual task: “Based on this information, please write a poem about…” There are two ways to create Generated Knowledge tasks: (i) the single prompt approach, and (ii) the dual prompt approach. The first simply places all the information within one prompt and then runs the model. The second works with two individual steps:

Although AI systems are being equipped with increasingly longer context windows (which is the part of the current conversation the model can “remember”, like a working memory), they have been shown to rely stronger on data at the beginning and et the end of the window (Liu et al., 2023). Since hence there is evidence that not all information within a prompt is equally weighed and deemed relevant by the model, in some cases the dual prompt or even a multiple prompt approach may yield better results.

To date, the perhaps most complicated method is known as Tree-of-Thought (ToT) prompting. The landmark paper by Yao et al. (2023) introducing the method has received significant attention in the community as it described a significant improvement and also highlights shortcomings of previous methods. ToT uses a combination of CoT and SC-prompting and builds on this the idea that one can go back and forth, eventually converging on the best line of reasoning. It is similar to a chess game where there are many possibilities to make the next move and in ones head the player has to think through multiple scenarios, mentally going back and forth with certain figures, and then eventually deciding upon which would be the best next move. As an example, think of it like this: Imagine that you have three experts, each having differing opinions. They each lay out their arguments in a well-thought-through (step-by-step) fashion. If one makes an argumentative mistake, the expert concedes this and goes a step back towards the previous position to take a different route. The experts discuss with each other until they all agree upon the best result. This context is what can be called the ToT-context, which applies regardless of the specific task. The task itself is then the query to solve a specific problem. Hence a simplified example would look like this:

The present author’s experiments have indicated that GPT-3.5 provides false answers to this task when asked with Input–Output prompting. However, the responses turned out to be correct when asked with ToT-prompting. GPT-4 sometimes implements a similar method without being prompted, but often it does not do so automatically. A previous version of ToT was known as Prompt Ensembling (or DiVeRSe: Diverse Verifier on Reasoning Steps), which worked with a three-step process: (i) Using multiple prompts to generate diverse answers; (ii) using a verifier to distinguish good from bad responses; and (iii) using a verifier to check the correctness of the reasoning steps (Li et al., 2023).

Sometimes, there sems to be a degree of arbitrariness regarding best practices of AI, which may have to do with the way a model was trained. For example, saying that that GPT should “take a deep breath” in fact appears to result in better outcomes, but it also seems strange. Most likely, this may have to do with the fact that in its training material (which nota bene incorporates large portions of the publicly available internet data) this statement is associated with more nuanced behaviors. Just recently, an experimenter stumbled upon another strange AI behavior: when he incentivized ChatGPT with an imaginary monetary tip, the responses were significantly better – and the more tip he promised, the better the results became (Okemwa, 2023). Another interesting feature that has been widely known for a while now is that one can disturb an AI with so-called “adversarial prompts”. This was showcased by Daras and Dimakis (2022) in their paper entitled “Discovering the Hidden Vocabulary of DALLE-2” with two examples:

To us humans, nothing in the letters “turbo lhaff✓” has anything to do with a dog. Yet, Dall-E always generated the picture of a dog and transformed, for example, the mountain into a dog. Likewise, there is no reason to assume that “Apoploe vesrreaitais” has anything to do with birds and that “Contarra ccetnxniams luryca tanniounons” would have anything to do with berries. Still, this is how the model interpreted the task every time. This implies that there are certain prompts that can modify the processing in unexpected ways based on the procedure of how the AI is trained. This is still poorly understood since to date there is yet no clear understanding how these emergent properties awaken from the mathematical operations within the artificial neural networks, which is currently the object of research in a discipline called Mechanistic Interpretability (Conmy et al., 2023; Nanda et al., 2023; Zimmermann et al., 2023).

Critical thinking, in the context of AI education, involves the ability to analyze information, evaluate different perspectives, and create reasoned arguments, all within the framework of AI-driven environments. This skill is increasingly important as AI becomes more prevalent in various aspects of life and work. In educational settings, AI can be used as a tool not just for delivering content, but also for encouraging students to question, analyze, and think deeply about the information they are presented with (van den Berg & du Plessis, 2023). The use of AI in education offers unique opportunities to cultivate critical thinking. AI systems, with their vast databases and analytical capabilities, can present students with complex problems and scenarios that require more than just rote memorization or basic understanding. These systems can challenge students to use higher-order thinking skills, such as analysis, synthesis, and evaluation, to navigate through these problems. Moreover, AI can provide personalized learning experiences that adapt to the individual learning styles and abilities of students. This personalization ensures that students are not only engaged with the material at a level appropriate for them but are also challenged to push their cognitive boundaries. By presenting students with tasks that are within their zone of proximal development, AI can effectively scaffold learning experiences to enhance critical thinking (Muthmainnah et al., 2022).

As such, the integration of critical thinking in AI literacy courses is an important consideration. As students learn about AI, its capabilities, and its limitations, they are encouraged to think critically about the technology itself. This includes understanding the ethical implications of AI, the biases that can exist in AI systems, and the impact of AI on society. By incorporating these discussions into AI literacy courses, educators can ensure that students are not only technically proficient but also ethically and critically aware (Ng et al., 2021). There are a number of challenges that students face in a rapidly evolving world under the influence of Artificial Intelligence and critical thinking skills seem to be the most successful way to equip them against the problems at hand. Table 5 sketches out some of the major problems students face and how critical thinking measures can counteract them.

The idea of teaching scaffolding helps to foster students in their critical thinking skills in a digital and AI-driven context. There are several forms of scaffolding that lecturers, teachers, supervisors and mentors can apply (Pangh, 2018):

This goes to show that critical thinking is a key resource for dealing adequately with an AI-driven world and that educators play a vital role in leading students into digital maturity.

AI in education presents significant challenges and opportunities. Key challenges include the need for ongoing professional development for educators in AI technologies and pedagogical practices. Teachers require training in prompt engineering and AI integration into curricula, which must be restructured for AI literacy. This multidisciplinary approach involves computer science, ethics, and critical thinking. Rapid AI advancements risk leaving educators behind, potentially leading to classroom management issues if students surpass teacher knowledge.

Equitable access to AI tools is crucial to address the digital divide and prevent educational inequalities. Investment in technology and fair access policies are necessary, especially for underprivileged areas. Another challenge is avoiding AI biases, requiring diverse, inclusive training datasets and educator training in bias recognition. Additionally, balancing AI use with human interaction is vital to prevent social isolation and promote social skills development.

Opportunities in AI-integrated education include personalized learning systems that adapt to individual student needs, accommodating various learning styles and cognitive states. AI can assist students with special needs, like language processing or sensory impairments, through tools like AI-powered speech recognition. Ethical AI development is essential, focusing on transparency, unbiased content, and privacy-respecting practices. AI enables innovative content delivery methods, such as virtual and augmented reality, and aids in educational administration and policymaking. It also fosters collaborative learning, connecting students globally and transcending cultural barriers.

The AI in the realm of education marks a transformative era that is redefining the teaching and learning methodologies fundamentally. This paper has critically examined the expansive role of AI, focusing particularly on the nuances of AI literacy, prompt engineering, and the development of critical thinking skills within the educational setting. As we delve into this new paradigm, the journey, although filled with unparalleled opportunities, is fraught with significant challenges that need astute attention and strategic approaches. One of the most compelling prospects offered by AI in education is the personalization of learning experiences. AI's capacity to tailor educational content to the unique learning styles and needs of each student holds the potential for a more engaging and effective educational journey. Moreover, this technology has shown remarkable promise in supporting students with special needs, thereby enhancing inclusivity and accessibility in learning environments. Additionally, the focus on AI literacy, prompt engineering, and critical thinking skills prepares students for the complexities of a technology-driven world, equipping them with essential competencies for the future. However, these advancements bring forth their own set of challenges. A primary concern is the preparedness of educators in this rapidly evolving AI landscape. Continuous and comprehensive training for teachers is crucial to ensure that they can effectively integrate AI tools into their pedagogical practices. Equally important are the ethical and social implications of AI in education. The integration of AI necessitates a critical approach to address biases, ensure privacy and security, and promote ethical use. Another significant hurdle is the accessibility of AI resources. Ensuring equitable access to these tools is imperative to prevent widening educational disparities. Additionally, developing a critical mindset towards AI among students and educators is fundamental to harness the full potential of these technologies responsibly. The perhaps most significant danger is that both students and educators use AI systems without respecting their limitations (e.g. the fact that they may often hallucinate and provide wrong answers while sounding very authoritative on the matter).

Looking towards the future, several research and development avenues present themselves as critical to advancing the integration of AI in education:

The future of education, augmented by AI, holds vast potential, and navigating its complexities with a focus on responsible and ethical practices will be key to realizing its full promise. The present paper has argued that this can be effectively done, amongst others, through implementing AI literacy, prompt engineering expertise, and critical thinking skills.