Does higher education properly prepare graduates for the growing artificial intelligence market? Gaps’ identification using text mining

2.1Origins and evolution of AI

The term Artificial Intelligence (AI) includes all the theories and techniques used to build machines capable of simulating human intelligence [16]. Table 1 describes the most important breakthroughs in AI research over time that we detail in the following paragraphs.

The birth of AI coincides with the appearance of computers in the 1940s and 1950s, the emergence of the theories of connectionism and cognitivism, and the discussions about the possibility of creating an artificial brain. Indeed, a memo by Warren Weaver [31] on automatic translation of languages suggested that a machine could very well perform a task that falls within the scope of human intelligence. Similarly, Alan Turing’s article [18] explored the issue of determining whether a machine is conscious or not. The Turing Test, which evaluates the ability of a machine to hold a human conversation, is derived from this article. The formalization of AI as a true scientific field dates back to 1956, during a conference organized in the United Sates at Dartmouth College [32]. Subsequently, this field reached prestigious universities such as Stanford and MIT.

In the mid-1960s, AI research was mainly funded by the US Department of Defense and concentrated in a few specialized laboratories around the world. In 1974, an “AI Winter” occurred as many experts failed to deliver their promises on AI development and applications. As a result, the British and American governments reduced their funding for research in this field [22]. In the 1980s, the success of the expert systems led to the revival of AI research projects. An expert system is a software that is capable of reproducing the cognitive mechanisms of a human expert in a particular field based on a set of rules. Due to this success, the AI market reached a value of one billion dollars, which has motivated different governments to financially support academic projects again [33]. The exponential development of computing performance according to Moore’s Law made it possible, between 1990 and 2000, to exploit AI in previously uncommon areas such as mineral research or medical diagnosis. It was not until 1997 that a real media release took place when the famous AI Deep Blue created by IBM beat Garry Kasparov, the world chess champion [34].

Between 2000 and 2010, AI continued to grow with the development of connectivity and mobility, and the increased accessibility to personal computers, Internet, smartphones, etc. Since 2010, Moore’s Law has been guiding the progress of AI, supported by data processing. Indeed, to perform a task, a system only needs rules: when it comes to thinking or delivering the right answer, it needs to learn. The problem of AI was no longer to have the brains for systems’ development, but to have data to process. Thus, researchers developed new processes for Deep Learning which, fed by the data, achieved exceptional performance and enabled the launch diversified projects [35]. Today, data management allows AI to be applied to understand X-rays better than doctors, drive cars, do translation, play complex video games, create music, see through a wall, imagine a missing part of a photograph, etc. The fields where AI performs well are constantly growing, which raises many questions about the professional role of human beings in the future [36].

2.2AI market

It is difficult to understand and grasp all areas of application for AI, because of their diversity and their overlapping. We therefore list the most common applications of AI today. This technology can be found in expert systems capable of simulating the behavior of a human expert who performs a specific task. AI is also widely used for the symbolic representation of knowledge in order to manipulate it through software. A third common application is Natural Language Processing (NLP) to understand a text independently of the language in which it is written [37]. AI additionally allows the resolution of complex problems found for example in games, and the recognition of speech, handwriting, or faces, for which neural networks are particularly efficient [38]. Finally, robotics is a dazzling use of AI with robots that react according to their environment, learn to recognize faces, follow moving objects and respond to sound and visual stimuli [39]. All of these applications are based on learning, as an algorithm or software should have autonomous learning capabilities to be considered intelligent [35].

The presence of these AI applications varies according to the company’s sector of activity. Obviously, the sectors and processes that intensively use AI are those that generate the most data. We find it in the field of Marketing, especially for customer targeting, improving customer experiences using chatbot, and for e-commerce recommendation engines. AI is also used in Finance to perform high frequency trading, market analysis or even to develop advising robots. More specifically, in banking institutions, it can detect fraudulent transactions and calculate credit points to determine a customer’s financial profile [40]. In addition, AI is strongly mobilized in social media to train neural networks to recognize faces and to moderate audiovisual content and comments [41]. Moreover, the medical sector increasingly integrates AI technologies for diagnosis by taking into account patients’ history and allowing a better cross-referencing of medical specialties, or even for the automatic calculation of bone age from the X-ray of a hand [42]. The entertainment industry is also one of the biggest users of AI, which is integrated in gaming robots, augmented reality, and media content recommendation processes [38]. Finally, the recent emergence of IoT opens up the prospects for new uses of AI in home automation, farming, drones and autonomous cars [3].

Thus, AI techniques and their applications are constantly evolving and therefore require an evolution of professions. Indeed, several job positions are being transformed or completely replaced by AI technologies [9]. These include the professions of the past (e.g. assembly line workers, cashiers), those that are currently being automated (e.g. professional drivers, offshore professions) and those that will be automated later (health, auditing, etc.). In addition, the advent of AI is leading to the creation of new jobs related to algorithmic, emotional or data processing, such as Data Scientist, Big Data Developer or Data Manager [43]. Consequently, a major question arises concerning the readiness of academic institutions to train students in line with these rapid changes in the job market. However, no research performed an exhaustive analysis of the potential gap that exists between the needs of the recruiting firms and the skills developed in academic training focused on AI.

Few previous studies provided a partial answer to this question. On the one hand, the authors who performed a gap analysis focused on the IT sector as a whole including AI, but without providing a refined analysis of AI needs and skills. [11] studied the demand of IT professionals by exploring job offers and identified the most taught IT skills such as programming languages, databases and networks. [10] compared IT professional skills found in academic and professional literature, while [44] determined IT skill gaps from candidate profiles on social networks and job offers. On the other hand, studies that looked at skill requirements for AI in particular analyzed the needs of the job market without comparing them with the content of academic training. [45] used Latent Semantic Indexing to construct a taxonomy of Business Intelligence and Big Data skills from job offers. Other researchers used interviews and questionnaires to identify the skills required to be an AI professional ([13, 14]). They pointed out the soft skills needed, discussed the role of AI in IT education, and presented recommendations for the integration of AI in academic programs. The present research therefore complement these previous studies that provided a relative reading of AI skills needed by the job market. We propose to use a text mining approach that is independent of researchers’ interpretation biases, to draw up an overview of recruiting firms’ requirements in terms of AI competencies. We also carry out a comparison between these professional requirements and the content of academic training to formulate concrete proposals for adjusting the pedagogic offer in light of the emerging professions and the subsequent skills needed.

3.1Data collection

To extract the content of courses in France integrating AI, we identified the French engineering and business schools that offer such training. In this respect, we considered the Figaro rankings (Ranking of the best business schools in 2020 and the 2020 ranking of French engineering schools) and looked at whether each of these schools offered courses in AI. We then retrieved the brochures of the AI specialized curricula of these engineering and business schools. To allow a more refined analysis of training content in France, we also retrieved the syllabi of AI courses in these schools aside of specialized AI curricula. This process resulted in 33 extracted documents. Finally, we formatted them in.txt to enable their analysis by an algorithm.

For the extraction of the job offers, we chose the website www.indeed.fr as it is considered as the most complete and up-to-date database for job advertisements in France. Other platforms such as LinkedIn are often used by some firms as an additional means to broadcast the same offers posted on Indeed. We employed a script developed with Python to scrap the content of the selected website. The script allows to perform a job search with chosen keywords, in our case “Artificial Intelligence; AI; and/or Data Science” in French and English. To form an accurate dataset of job offers with the least possible redundancies, we scanned the entire site and made an extract for each month starting with June 2020 and comparing it with the preceding month. This approach showed that companies tend to repost the same job offers every one to three months in order to always appear in the first search results on the website and ensure better visibility to candidates. Therefore, we restricted the considered time span to March to June 2020, as it was the most appropriate interval to take into account the job market dynamics and avoid redundant job postings. The extracted data progressively fed a table with a job offer Id, its title and a description that concatenates all the available information about the offer. By going through the 307 job postings resulting from this process, we established that some of them corresponded to positions citing AI as a skill and not to AI occupations per se. As we are concerned by the second category in our research, we manually filtered the job offers as an automated analysis algorithm would have been inefficient given the small size of our sample. This process finally resulted in 281 job offers centered on AI.

3.2Data analysis

To ensure consistency in the analysis, the content of the job and training offers in French has been translated into English via DeepL, a highly recommended translator using convolutional neural networks [49]. We relied on keyword extraction to analyze these contents. In this respect, we developed a Python code to process English text and identify its keywords based on NLP (Natural Language Processing). The NLP vectorizes the words and makes them understandable by an algorithm that can decipher, interpret and give meaning to human language. Our code analyzed all the words in the text, their frequency, their environment, their nature and function, and was able to produce the most relevant keywords as a result. This process allowed us to identify, within the training brochures and the job offers, elements associated with the level of training, its nature, and the listed skills. Finally, we manually sorted the extracted keywords because some of them remained incoherent, despite our efforts to optimize the code. Indeed, the use of NLP is not perfect and sometimes leads to errors that must be managed humanely in post-processing.

In addition, we analyzed the titles of the job offers to identify the emerging professions in AI. We used an automated learning approach to filter the titles and classify them into categories. This approach consists of a textual categorization process, which automatically constructs a classifier by learning the characteristics of the content categories from a set of pre-classified documents [50]. We experimented several classifiers and concluded that the Support Vector Machine classifier was the most accurate.

4.1Distribution of AI professions

Our resulting classification of job offers (Fig. 1) on AI shows that Data Scientist is the profession that recruits the most (25%). The latter designs programs to acquire, record, process and exploit massive data by crossing several sources thus making artificial learning effective. Then, two professions focusing on the upstream phases of the data value chain arrive at the second place (∼17%): the Data Analyst who has a more restricted role than the Data Scientist because he is often limited to a certain type of data; and the Data Engineer who focuses on building the architecture and infrastructure for data generation. In the third place, we find the Data Manager (12%) who has in-depth knowledge of the company’s sector and the competition, and participates in strategic decision-making. Then, the Data Architect, who is responsible for making the choice between the different storage technologies, is required in 7% of the job offers. Finally, four other AI professions with strong technical connotations are very rarely found in job postings. These are the Big Data Developer who is proficient in various programming languages particularly Java and Python; the AI Engineer who focuses on the design of neural networks; the Data Quality Engineer whose role is to design and verify infrastructures of data quality control; and the Data Curator who focuses on metadata management and on identifying best practices for working with data.

Fig. 1

Categories of AI professions.

4.2Comparison in terms of training peculiarities

As depicted in Table 2, based on the keywords extracted from the job offers, we characterized the market prerequisites in terms of the nature of training. We concluded that firms most often require a candidate with a purely technical background (77%). In a very rare cases, the advertisement stresses the need to have a dual engineer/manager background (1%). No job offer is primarily focused on the candidate’s managerial skills. The rest of the offers emphasizes the need for the candidate to have an engineering or a business school background (20%). These results are aligned with the characteristics of the academic market. Indeed, our analysis of the training content showed that curricula centered on AI are present in 75% of engineering schools while such diplomas exist in 30% of business schools. This can be accounted for the fact that AI professionals must primarily possess technical skills to properly handle data and algorithms. Additionally, AI courses for curricula not specialized on AI are proposed in 20% of engineering and 45% of business schools. The keen interest of management institutions in the AI thematic can be explained by the need to familiarize their future alumni with use cases of this technology and its jargon in order to promote their dialogue with different professions within firms.

As for the level of training, the vast majority of job postings request a Master degree (82%) due to the highly specific nature of the knowledge underlying the AI professions. Some job offers even require a PhD often in computer science (11%). In a very rare cases (6%), the level of training is not specified. However, the content of these offers does not suggest that candidates may have a level below the skills associated with at least a Master degree. These findings are overall consistent with the academic offers. Indeed, the majority of specialized AI curricula are Master degrees (85%). Also, 33% of the analyzed schools include AI courses in Bachelor and Master levels’ curricula not centered on AI.

4.3Comparison in terms of skills

Figure 2 summarizes the most frequent skills in the AI job advertisements while the overall skills that emerged from our analysis are introduced in Table 3. These skills can be divided into three categories namely Technical, Soft and Interdisciplinary skills. We analyze below the composition of these categories and compare them with the content of academic trainings.

Fig. 2

Word frequencies in AI job offers.

First, Fig. 2 shows that technical skills are the most recurrent in job postings. They include the following five sub-categories:

In addition to technical skills, the majority of job offers require soft skills that can be divided into two sub-categories:

Finally, our analysis highlighted the need for three sub-categories of interdisciplinary competencies with respect to AI:

5.1Theoretical implications

From a theoretical standpoint, we complement prior studies that examined some skills acquired by students enrolled in AI [53] and IT programs ([54, 55]) by providing an overview of emerging professions centered on AI and a categorization of the skills required to operate in this field. Our results would guide graduates towards some of the most critical competencies in the 21^st century [56], particularly by helping them build an AI capability [15] at the individual level. This study also supports the establishment of such capability at the organizational level by raising companies’ awareness of AI professions and the combination of skills that should be developed internally through employees’ engagement and reciprocity [57] as well as top management support in order to capitalize on this emerging technology.

5.2Practical implications

From a practical standpoint, this research first helps higher education institutions to align academic training on AI with the job market needs [58] and provides educators with an overview of the critical skills that they should strive to reach as learning outcomes of their courses [59]. Our results would also promote the integration of graduates in the job market by sensitizing them to the AI related competencies they need to develop even autonomously to fill potential gaps in their academic curricula. In addition, through this interdisciplinary study, companies can be informed on the important AI skills and occupations regardless of the operating sectors, thus allowing a cross-fertilization of good practices. At the country level, the present research may help French institutions to develop constructive partnerships for AI programs with other developed countries whose AI labour markets possess similar dynamics. These partnerships should be established to allow students acquire the skills lacking in their training within their original institutions. By applying a similar analysis at the level of each partnering country, it would be possible to build an educational ecosystem around AI among developed countries that promotes their complementarity by enabling students access country-specific sets of skills through academic exchanges. Our results can also be beneficial to developing countries to optimize their construction of school programs in AI and the organization of the labour market in order to capitalize on the contributions of this technology.

5.3Limitations and future research avenues

This study has several limitations that pave the way for future research. In fact, it would be interesting to conduct a refined analysis of our sample in order to establish a breakdown of skill requirements by level of experience and by profession category. Then, our sample can be expanded in three ways to yield new results. Firstly, the analysis can cover international AI courses that constitute an input of competencies for students in France as several schools established academic partnerships internationally. Secondly, the scope can be extended by including university Master and Doctorate level programs that represent a significant part of future AI recruits. Thirdly, it would be interesting to include in the sample of job offers those in the IT field integrating AI but that do not constitute jobs centered on this technology. This will raise awareness of other academic training in IT to the interdisciplinary nature of AI and promote its integration into these curricula. A last limitation concerns our translation from French to English of job offers and academic training content to ensure a consistent analysis. This step, although essential, can lead to errors and thus to misunderstandings or inconsistencies in the analysis algorithm. We used DeepL for translation which relies on convolutional neural networks and is highly recommended for this kind of application. It would be interesting to deepen the work by mobilizing other translators and comparing the results.