1 Introduction

To build machines for automation of what we today call artificial intelligence (AI) is an idea that has been around for a long time. As early as in the fourteenth century Ramon Llull described his concept Ars Magna, a concept for implementing thought and reasoning processes in an intelligent machine. Ars Magna was built around the idea of combining logical system to evaluate if postulates were true or false, something that could be seen as the very origin of what we today call symbolic AI [19]. Ideas that later inspired scientists such as Giordano Bruno, Athanasius Kirchner, Agrippa of Nettesheim and Gottfried Wilhelm Leibniz [32]. What we today consider to be the modern computational model for intelligent reasoning was presented by Alan Turing [70], a machine model that is the foundation of computer science, and the idea of using computers to solve complex problems.

The emerging field of Artificial Intelligence in Education (AIED) is a younger one, with its roots in the 1970s [60, 65]. In the twenty-first century, AI have in various ways and frequently been suggested to enhance educational activities and processes [22, 36]. AIED started out as a playground and research field for computer scientists [15], but a field that gradually made a strong impact on education and become a cross-disciplinary phenomenon [22, 36]. Some examples of how AIED might facilitate learning and teaching activities are the potential to support student collaboration, and to enable mass individualisation in large student groups [43]. A prioritised research area in AIED has been the attempt to reach the same quality and efficiency as in traditional one-to-one human-tutoring in technology enhanced learning environments [72].

This article updates and extends a literature study that previously has been presented by the authors [28]. Authors have taken courses on AI at university level and participated in specific research seminars on AIED. Outcomes from one of the research seminars have been published in Hrastinski et al. [27]. The aim of the study is to identify potential aspects of threat, hype and promise in artificial intelligence for education. With the SWOT model as the analytic lens the research question to answer was: Which are the strengths, weaknesses, opportunities, and threats of the ongoing implementation of artificial intelligence in education?

2 Theoretical background

An important concept in this study is 'hype', which has been based on the descriptions provided by the Gartner Hype Curve introduced in 1995 [58]. The hype curve was designed to support the decision of when to invest in a novel technology and consists of five stages: technology trigger, peak of inflated expectations, trough of disillusionment, slope of enlightenment, and plateau of productivity [58]. Fenn [17] describe these stages as:

  1. 1.

    Technology trigger: Industry interest and significant press is generated through, for example, a public demonstration, a breakthrough, or product launch.

  2. 2.

    Peak of inflated expectations: Limits of the technology is pushed with unrealistic projections and overenthusiasm as more failures than successes are achieved.

  3. 3.

    Trough of disillusionment: The technology becomes unfashionable and interest wanes as the overinflated expectations cannot be met.

  4. 4.

    Slope of enlightenment: A true understanding of the risks, applications, and benefits of the technology is reached through solid hard work and focused experimentation by organisations of a diverse range.

  5. 5.

    Plateau of productivity: The reduced risk of the technology and the demonstrated benefits makes more organisations comfortable, and a phase of growing adoption and acceptance begins.

2.1 Artificial intelligence

When Turing [71] extended his ideas in 'Computing Machinery and Intelligence', he created a bedrock for AI, and the start of what is called the first wave, or the first spring of AI. Turing's ideas in the 1950s are today seen as the foundation for both computer science and AI, despite the fact that Turing never used the term AI [11]. The following hypes and declines of AI has been described as springs and winters of AI [47], or as the three waves of AI [35]. There are today, in the third wave of AI, high expectations of a future in which AI systems act as partners to humans rather than as tools [35]. At the same time the new AI spring shows a rapid progress of sub-symbolic AI, which also brings a number of caveats to consider if AI systems should be classified as beneficial and human compatible [47].

To further formalise logic and intelligent reasoning and to refine and reinforce AI, has been an ongoing process since Turing built the first computer chess engines in the 1950s. As suggested some years later by Claude Shannon, computer chess is an interesting testbed for computer science and AI [34]. A series of successful computer programs have challenged chess grandmasters, and today chess engines such as Stockfish and Komodo outplay the strongest grandmasters. With the use of machine learning and deep learning technology, computer software such as Alpha Go has also mastered the complex Go game [66]. However, these strong and specialised AI systems have not reached a lobster’s level of social skills and cannot handle general real-world problems. This comparison that was presented by John Searle [62], is still valid today. Searle was also the researcher who coined the terms weak AI and strong AI.

Today, weak AI and strong AI have been renamed as narrow AI and artificial general intelligence (AGI). The term AGI was coined and spread by the AI-researchers Shane Legg, Mark Gabrud and Ben Goertzel with AGI defined as a general artificial intelligence on the level of human intelligence [67]. Strong AI or AGI can also be illustrated by the Turing test, where true AGI is accomplished when it is no longer possible to tell the difference between a natural language conversation between humans, and the one between a human and an AI system [71]. Narrow AI can be exemplified by specialised AI applications communicating via Bluetooth or other protocols. More impressing examples of narrow AI on superhuman level, is when the heuristics-based chess engine Deep Fritz beat the grandmaster Vladimir Kramnik [25], or when the machine learning based AlphaGo defeated the Korean Go grandmaster Lee Sedol [77].

Machine learning with the use of deep neural networks is the AI field that has made a fast progress in the third wave of AI. A deep neural network is a neural network with more than one hidden layer between the input layer and the output layer, where complex tasks can be divided between the different layers [47]. However, as pointed out by Korteling et al. [37], all developed AI systems must be classified as weak or narrow AI, since they all lack the general problem-solving skills that exist in biological intelligence. As an example, the answer to the question if AI systems might replace human doctors in the treatment of patients is still an undoubtful no. There are today no deep learning algorithms or combination of heuristics that can understand human emotions at the deeper level necessary for treating patients with severe diseases [77].

2.2 SWOT

A common part of quality control and investigation of strategic planning in an enterprise system is to carry out an analysis of the potential strengths, weaknesses, opportunities, and threats (SWOT) [51]. A SWOT analysis can be a powerful tool for identifying potential influences of a system on an organisation, and to identify skills, support, attitudes, knowledge, and abilities that are needed in an organisation [39]. However, there are also critique of how SWOT analysis are interpreted. SWOT analyses are typically used for identifying and naming strengths, weaknesses, opportunities, and threats and does not include the impact of individual factors, or the potential impact, on desired outcomes [39]. Furthermore, traditional SWOT analyses are mainly concerned with the strengths, weaknesses, opportunities, and threats of a specific technology or implementation and therefore provide little guidance for alternative decisions [39]. To enhance the SWOT analysis conducted in this study, the identified factors of strengths, weaknesses, opportunities, and threats are discussion in “Results and discussion” section to address these limitations of the SWOT analysis.

The conventional approach in conducting a SWOT analysis is to generate SWOT categorisation in a 2 × 2 matrix, containing internal factors of strengths and weaknesses and external factors of opportunities and threats [39]. Typically, there are two questions being asked related to internal and external factors when conducting a SWOT analysis: Which are the benefits and costs? And are these occurring inside or outside of the organisation? [39] In this study, this has been translated to:

Which are the benefits (strengths and opportunities) and costs (weaknesses and threats) of artificial intelligence in education?

Are these benefits and costs internal (strengths and weaknesses within the AI-technology) or external (opportunities and threats for education)?

In previous research, SWOT analyses have been used for examining the strengths, weaknesses, opportunities, and threats of reforming higher education with artificial intelligence, machine learning, and extended reality; with a focus on how these techniques can be used to support the development of a new strategy for higher education [31]. The SWOT analysis approach has further been used for investigating the strengths, weaknesses, opportunities, and threats in strategic documents for the development of AI. This was conducted with specific attention for aspects concerning science, technology, engineering, and mathematics (STEM) education [2]. SWOT analysis has also been used for examining whether implementation of artificial intelligence in higher education would hinder or support educational processes and how teachers affected by the implementation would perceive the new technologies [9, 41].

3 Method

This study was conducted as a scoping literature review to provide an overview of a selected topic, as described by Munn et al. [49], with AIED as the selected topic. A scoping review can be an appropriate approach to use for studies with an aim of clarifying concepts and to identify knowledge gaps. Contrary to the systematic review, the aim of a scoping review is not to synthesise the results related to a specific research question but rather to map and provide an overview [49]. Furthermore, a scoping review is a method for finding key concepts in specific research fields, and to identify the main sources for further research [49]. Therefore, scoping reviews can serve as a precursor to further research and systematic reviews [49].

3.1 Data collection

The main keywords in the literature search were: 'artificial intelligence', 'artificial intelligence in education', AI, AIED, education, teacher, and 'teacher perspective'. These keywords were combined with the use of the Boolean operators 'AND' and 'OR'.

Query 1: "artificial intelligence in education" OR "AIED"

Query 2: ("artificial intelligence" OR "AI") AND ("education" OR "teacher" OR "'teacher perspective")

Google Scholar was used as the search engine to find research papers that had a potential to answer the aim and research question of the study. Results were filtered to only include papers with a publication year no older than 2015. Backward searches were also used to include papers of interest to the study’s aim and research question. All papers were not directly accessible through Google Scholar and were retrieved from the aggregation of research databases at the Mid Sweden University library.

The different combinations of keywords used in the searches resulted in many hits. However, the potential to contribute to answering the research question was often limited. Many papers did not primarily study AI concepts that could be classified to involve some degree of intelligence, but more of automatised systems. Similarly, the categories in the SWOT framework served as filter for the papers to be included. A total of 20 papers were selected in the first phase of this study conducted in 2019 [28]. In the second phase, conducted in 2022, the list of included papers was revised and expanded to 41 papers to be included in the study (Table 1). The new list includes publications between 2020 and 2022 to contribute to the description of a research field that has rapidly grown during recent years. All papers that were considered for inclusion have been discussed between the authors to ensure contribution to the results of the study.

Table 1 Overview of papers included in the study
Full size table

3.2 Data analysis

This study used thematic analysis to identify themes in the selected and included papers [46]. The analysis was inspired by the six phases for conducting a thematic analysis presented by Braun and Clarke [8]: (1) familiarizing yourself with the data, (2) generating initial codes, (3) searching for themes, (4) reviewing potential themes, (5) defining and naming themes, and (6) producing the report. However, with the use of the SWOT framework (Fig. 1) and a deductive-inductive approach, the analysis started with another initial phase. The initial phase consisted of defining the main categories for the deductive part of the coding in the analysis, which was decided to be strengths, weaknesses, opportunities, and threats, or the SWOT framework. A more detailed description of the SWOT framework is provided in “SWOT” section.

Fig. 1
figure 1

The SWOT analysis framework

Full size image

The deductive approach was used as a 'top-down approach' [8] to order data and identified themes in the SWOT framework. The inductive approach was used as a 'bottom-up approach' [8] to group codes and extracts from the included papers in potential themes based on similarities and differences in the data. Further, it should be noted that the thematic analysis presented in this study has been conducted at two points in time. The first analysis was conducted during autumn of 2018 and spring of 2019 and was presented as a shorter conference paper in 2019 [28]. The second analysis was conducted during 2022 and was used to revise and extend the previous analysis. A detailed description of the phases of analysis is provided below.

In the first phase of analysis, familiarisation with data were reached through thoroughly reading potentially relevant papers that could support in answering the study’s aim and research question and by adding to the SWOT categories for analysis. In parallel with the first phase, the second phase of analysis was carried out deductively by collecting extracts from the papers under the relevant SWOT category in a text document. In the third and fourth phase of analysis, these extracts where grouped and re-grouped inductively within the SWOT categories in search for potential themes. These themes were discussed and revised for consistency by both authors. In the fifth and sixth phase, authors had come to an agreement on the identified themes and proceeded to naming these and writing up the report. This concluded the analysis conducted for the first conference paper, which identified 17 themes divided between the SWOT categories [28].

For the second iteration of the analysis, conducted in 2022 and the main contribution of this study, the six phases were repeated. In the first phase of analysis, both previous included papers and new potential papers were considered and re-considered for inclusion in the analysis. In the second phase of analysis, both old and new extracts were collected in the SWOT categories. In the third and fourth phase of analysis, extracts were once again grouped and re-grouped within the SWOT categories in search for themes that could either be the same as before, revised, or new. In the fifth and sixth phase of analysis, authors had once again come to an agreement on the themes for the study and proceeded with naming these and writing up the report. This second iteration of analysis resulted in a revised and extended list of included papers (41) but a more condensed number of themes (10), which are divided between the SWOT categories and further presented in “Results and discussion" section.

4 Results and discussion

In this section the results from the study are presented and discussed. Results from the literature study are presented in sub-headings below. Themes in sub-headings 4.1 and 4.2 address internal strengths and weaknesses of artificial intelligence in education, that is, the strengths and weaknesses of the system for education. Themes in sub-heading 4.3 and 4.4 address external opportunities and threats of artificial intelligence in education, that is, the opportunities and threats that artificial intelligence brings to education. Further, themes addressed in sub-headings 4.1 and 4.3 can be viewed as aspects of artificial intelligence that could be helpful for achieving objectives of implementation artificial intelligence in education. While themes addressed in sub-headings 4.2 and 4.4 can be viewed as aspects of artificial intelligence that could be harmful for achieving objectives of implementation artificial intelligence in education (Table 2).

Table 2 Summary of themes related to the SWOT categories
Full size table

4.1 Strengths

A belief surrounding the development of AI is that AI will support in making computer-aided teaching and learning more efficient [42]. A common field for addressing teaching and learning with computer systems is science, technology, engineering, and mathematics (STEM). Previous research show that artificial intelligence in education often focuses on domain knowledge in STEM and computer science [61, 76]. With a step-based approach to well-defined problems in STEM, AIED has been successful in supporting teaching and learning of domain knowledge [61]. Examples of where AIED and intelligent tutoring systems have been applied in STEM are computer science education [6], and in computer programming education [14].

There are many sub-fields and branches of artificial intelligence, and one of these is natural language processing (NLP) [3, 13]. NLP can be described as computational techniques intended for producing, understanding, and learning human language [24] and is commonly applied in the context of education [13]. Although the use of NLP for educational purposes is often irregular and complex [56], previous research has highlighted use cases. An example of where NLP can be applied is to support development of students’ social, language and work skills [3]. Speech generation and translation of text can be performed by software-controlled AI assistants with NLP algorithms [20]. NLP can also support students in learning and work life-training by recording speech, provide feedback, and order and suggest steps of action [74]. Further, previous research has suggested to use NLP in combination with machine learning to aid in preparing texts of appropriate difficulty for reading comprehension [5].

Systems used in educational contexts can be described as an interconnected ecosystem where a change in one system can affect the whole [44]. It is therefore important that the implementation of AI-systems in education builds on supporting educational ecosystems [44, 54]. A benefit of applying artificial intelligence in education as an ecosystem is that each AI-system can target a specific purpose, building on highly specialised research, and keep a wide application through the connectivity to other systems in the ecosystem [55]. Specialised AI-systems that could be integrated in such ecosystems are intelligent tutoring systems, which can personalise tutoring, suggest learning paths, engage students, provide feedback, and improve learning experiences [18, 21, 38, 76].

4.2 Weaknesses

Artificial Intelligence techniques can be applied in educational contexts for adaptive tutoring of students in the form of intelligent tutoring systems (ITS) [16]. However, ITS often rely on AI-techniques that are much simpler than what was initial intended [4, 16]. As a part of this, Baker [4] labelled many ITS as 'stupid tutoring systems' and in need of collaboration with human intelligence to be amplified [4, 16, 21]. Other systems that draw on AI-techniques, such as intelligent decision support systems (IDSS), have been surrounded by similar AI hype, which unfortunately have fallen short of the expectations [26].

With the rise of artificial intelligence in education, the systems implemented in educational contexts is and will be built by people. Algorithms that are developed to process data are created by programmers with potential biases in the code [53, 59]. Moreover, training data and models used for machine learning are corrected and evaluated by humans [50]. Since there are no definite guidelines for ethics in either AI or AI-applications in education [53], the weaknesses of AI-systems may have real consequences for education due to the increased attention that AI researchers, product developers, venture capitalists, and advocates for educational technology are putting on the educational market [50]. The consequences of potential biases in AI-systems for education are further amplified by the marketing efforts to present AI-algorithms as value-neutral and objective to the public [1]. The use of AI-systems and technological solutions in education raises the question of "who sets the agenda for teaching and learning" [59].

4.3 Opportunities

Luckin et al. [43] states that they do not see that teachers will be replaced by AI-systems in future education. AI is instead highlighted as the possibility for enhancing education with high-quality education and assisting human teachers to make this widely accessible [10]. Previous research suggests that AI-systems should focus on assisting concrete pedagogical tasks that for a human teacher would be perceived as exhausting and time-consuming, for example assisting in constructing grade responses [52]. Another possibility with AI as an assistant to teachers is that it could free teachers to focus more on supporting students' development to independent collaborative thinkers, instead of possessing and transmitting relevant knowledge [61]. The AI-system could record and analyse students’ work and report back to the teacher with suggestions of which students might need extra attention, what is sometimes referred to as cobots (co-working robots) [69].

Similar to the idea of using AI to assist human teachers, is the idea of using AI to assist the students. A suggested application for AI in education is to use it for connecting and engaging students with other students and with their teachers to increase efficiency of learning [20, 40]. Further, AI can be used for adaptation of learning materials for students with special learning needs, and to provide timely support for those students [45, 53]. Furthermore, there are several examples of studies where the roles of students as tutees are shifted to tutors and the students learn by coding, teaching, or instructing an artificial tutee [48, 57, 63, 64]. However, previous research suggests that personalised learning with AI should not be superficial, but that a crucial aspect is that the AI-system provides depth in the customisation [73].

Related to the idea of both AI as an assistant to teachers and to students, is the idea of using AI for individualisation of education on a massive scale. Previous research has suggested the potential of using AI-systems to put the learners in the centre and tailor the learning according to their needs and preferences [20]. Previous research has suggested that future research should investigate the potential of personalising education and promoting learning precision with AI-techniques such as natural language processing [13]. Previous research has also suggested that AI systems can be effective tools for supporting students with neurodevelopmental disorders to address challenges in learning and to personalise education [7]. AI is further anticipated to enhance education with support to teachers and making education of high-quality accessible more widely [10].

4.4 Threats

A potential concern for educational personnel is the "extinction risk fears" brought by predictions for future AI technologies [75] and the real and psychological effects this can have for those effected [30]. Although AI is not advanced enough to be able to replace teachers, the technology has presented itself capable enough to replace other personnel roles in education such as teacher assistants and administrators [59]. A potential change of the teacher role with AI in education can be exemplified by the stereotypical learning design of MOOCs, with low-level multiple-choice questions, and the teacher taking on the role of a content developer [61]. Previous research has further warned that widespread application of AI-techniques in education could potentially harm the relationship between teacher and students and hinder development of students to become independent learners, capable without online platforms and artificial teaching assistants [75]. Another potential threat is that AI-techniques meddles with what students are expected to learn, the standardisation, through ‘too much’ individualisation of education [75].

As AI-researchers, product developers, venture capitalists, and advocates for educational technology are putting more of their attention on the educational market [50], the potential negative consequences of poorly implemented or poorly adapted AI-systems in education increases. Still many schools and teachers lack in knowledge and are not prepared for the integration of AI in education, which increases the risk of technology abuse when AI is implemented [33]. This could for example have negative consequences for privacy and surveillance, and personal data could be leaked and used to influence individuals within the educational system [1, 53, 59]. Previous research has pointed out the importance of adopting ethical frameworks for the use and development of artificial intelligence in education [12, 29, 53]. If an ethical framework is developed and applied for AI in education, it still needs to be continuously discussed and updated because of the fast development of AI techniques and its potential for wide application [53].

5 Concluding remarks

In the literature study it has been noted that there is an ambiguity in many of the papers concerning AIED, particularly regarding the concept of 'intelligence'. Many of the ideas and technologies that have been suggested and discussed could be questioned whether they contain any level of intelligence. This ambiguity in the concept of AI might be explained by that AIED is in an emerging stage of hype [17, 58], with over-optimism regarding the potential to transform existing education. The AIED hype could be caused by the hype of progress in AI, but it is still unclear how the development in machine learning and deep learning could be applied in AIED. Despite the fast progress in the fields of deep learning and NLP there are still issues to address regarding these techniques. As highlighted by Mitchell [47], both these techniques are stuck in the 90–10 phenomenon that is common in AI development. The techniques are solved to 90%, but the remaining 10% can make AI systems fail with severe consequences. Moreover, the remaining 10% often takes longer time to sort out than the first 90% [47]. It could be wise to wait with investments until all involved AI techniques are completely developed without surprising side effect, especially in areas where AI systems interacts with humans. The authors would like to emphasise the importance of traditional academic values such as scepticism, to maintain that the goal of education should be to foster responsible citizens and educated minds [59].

This study resulted in findings that can answer the research question about strengths, weaknesses, treats, and opportunities of the ongoing implementation of artificial intelligence in education. Strengths were found in the areas of STEM education and language learning. However, the fast progress in machine learning with deep neural network in AI, is not at the same level in AIED. Regarding weaknesses, many AIED system are built with a low degree of intelligence, and sometimes on the level of what can be called 'stupid tutoring systems' [4]. Moreover, AIED like AI, still have the issues of biased data and biased algorithms to handle. In the identified general AIED hype there are also several opportunities, such as the potential for supporting teachers and students in more individualised and adapted learning environments. One example is that there exist intelligent solutions which can facilitate for students with impairments and special needs.

The found threats identifies important aspects of AIED to be mindful of in the ongoing implementation. What will the future role of the teacher, and other school staff, be in education with AI systems? And how does this align with our pedagogical beliefs or theories? Do educational leaders and teachers have sufficient knowledge in the field of AI to distinguish a poorly developed system from a good one? Or how to adequately apply these in educational setting? Also, how do we protect students’ and teachers’ data when the skills and knowledge to develop AIED systems lies with for-profit organisation and not within the educational sector? These are all questions that need to be considered carefully and addressed thoroughly in the years to come. Especially the question regarding AI’s alignment to pedagogical theory should be emphasised in future research, since any new technology integrated in education should be designed to fill a pedagogical need. Authors' recommendation is to develop new courses on AI, that not necessarily have to handle the more technical parts of neural networks and NLP, but rather be designed for a broader audience. There are some open MOOC alternatives developed during the last years [23], but there is still a lack of practical training courses on AI and courses that discuss the ethical aspects of AI [68].

The conclusion of the study is that there are threats, hype, and promises for artificial intelligence in education and the teacher practice. For a successful implementation of AIED systems that are in line with educational goals and beliefs the development, use, and consequences of applying these systems must be part of ongoing discussions and future research. Finally, authors' recommendation for the near future is not to attempt in building AIED ecosystems that should match the idea of AGI, but rather to invest in a synchronisation of specialised narrow AI systems that could function as extra support tools for teachers and students. Still, the concept of 'intelligence' should be emphasised in these tools to avoid ambiguity in AIED. Intelligence is of course a complex concept to pin down, but a practical definition such as “Ability to accomplish complex goals” [67], p.39) could be sufficient.

6 Limitations and future research

This study was conducted as a scoping literature review and therefore have some limitations compared to a systematic literature review. For a more comprehensively examination of the AIED field as a whole, a systematic literature review would be suggested. Although the scoping review succeeded in identifying interesting research topics and gaps, all identified topics and issues were not discussed in detail. To dig deeper into the field, authors suggest a follow-up study were found issues are directly asked to teachers. Data could be gathered both quantitatively by a large-scale survey and qualitatively by interviews. This would further allow for a comparison of the fears and threats highlighted in this study with teachers’ perceptions of AI in education. Considering the current AIED hype, teachers’ opinions and perceptions should be considered to facilitate AIED investments that are truly based in pedagogical needs and would add value to teachers’ daily activities.