Enterprise AI Analysis

In the 2024 article [35], the applications of AI in higher education are grouped into four categories, viz. profiling and prediction, intelligent tutoring, assessment and evaluation, and adaptive personalisation:
– Profiling and prediction is focused on employing data-driven approaches to make informed decisions regarding students' academic pathways. It includes the use of AI tools for optimising admission-related decisions and course scheduling, for predicting and improving dropout and retention rates, as well as for recognising patterns in students' data.
– Intelligent tutoring includes the use AI tools for providing customer-fit instructional interventions, in particular – for teaching course content, for diagnosing students' strengths and weaknesses and offering personalised feedback, for developing appropriate learning materials, as well as for providing insights from the teachers perspective, useful for improving pedagogical strategies.
– Assessment and evaluation comprises the use of algorithms for automated grading, providing immediate and personalised feedback to students, evaluating student understanding and engagement, as well as implementing mechanisms for the evaluation of teaching methodologies and effectiveness.
– Adaptive personalisation consists in using AI tools for fitting educational experience to the individual needs of students. It includes shaping course contents, recommending personalised content and learning pathways, utilising academic data to monitor, guide, and support students.
The following journals, by virtue of their mission, are most predestined to publish articles devoted to the use of AI tools in higher education: Active Learning and Higher Education, Assessment & Evaluation in Higher Education, Internet and Higher Education, Higher Education, Higher Education Research and Development, Journal of Higher Education, Research in Higher Education, Review of Higher Education, Studies in Higher Education, and Teaching in Higher Education. Many important contributions to the topic under consideration appear, however, elsewhere – in the journals of more technical or interdisciplinary nature. Syntheses of practical guidelines and recommendations concerning the use of AI tools in education may be found in [36] and [37].

The development of AI has been an issue of philosophical, especially ethical, debate for more than 80 years. The implementation of AI tools in education has enhanced this debate with a didactic and psychosociological perspective [38–41]. During the recent two decades, numerous attempts have been made to synthetise the multidimensional experience of using AI tools in education. Here is a package of review papers, published in the 2020s, which characterise the state of the art in this domain: [42–54] and six papers on pages 69–172 of the book [55].

The protagonists of using AI tools in education point out that these tools (1) can help increase students' engagement by enabling them a more personalised and interactive learning experience, (2) may improve learning outcomes by providing them with access to personalised tutoring and immediate feedback, (3) can reduce teachers' workload by automating tasks such as grading home works [56]. What are the main concerns about using AI tools in education? Roughly speaking, they are the same as those related to other uses of AI tools, viz.: possible infringements of privacy, new opportunities for dishonesty, information uncertainty and disinformation; there is, moreover, a concern that overreliance on AI tools can reduce human interaction between teachers and students, and – consequently – hinder the social maturing of the latter [57].

By many academic teachers, AI tools are still perceived as the enemies of education: these tools are blamed for eroding critical thinking, enabling plagiarism, and displacing academic staff - cf. [58]. It is true that, when deployed without ethical foresight or pedagogical reflection, Al tools can be misused, but we should not assume that such academic values as curiosity, rigour and creativity cannot coexist with Al systems. We should not blame the latter for generating new unsolvable problems, but rather notice that they reveal weak points of our routine approach to higher education, such as the bureaucratisation of learning, the ritualisation of assessment, and the widening gap between curricula and real life experience. In response to this challenge, we should try to reconsider the assumptions our educational systems are based upon; in particular – to redefine such basic concepts as knowledge, thinking and teaching in a new technological context; without going through pertinent philosophical reflection, we will not be able to reasonably decide whether students should be allowed to use AI tools in class, and to adequately redefine the role of an academic teacher – cf. [59]. Institutions of higher education must reclaim the future – they must preserve the moral and intellectual purpose of education, but not necessarily cling to traditional approaches and solutions at any price. Since MS&T is an interdiscipline with enormous diversity of applications, ranging from sciences to arts, it is an appropriate area for pioneering the implementation of this teaching philosophy.

Among the inevitable problems related to the above-suggested implementation, the issue of plagiarism requires deeper consideration here despite its universal nature. Noam Chomsky, the doyen of world linguistics, formulated it in a very concise and emphatic way: 'ChatGPT is basically high-tech plagiarism' [60]. In fact, the training process, including both pretraining and fine-tuning, for AI tools like ChatGPT involves learning from a vast amount of both publicly available and licensed data, which enables those tools to generate responses based on patterns in those data. As a rule, however, the AI tools do not directly plagiarise existing contents they generate new responses based on learned patterns rather than exactly copy phrases from original sources. Today, many artists and intellectuals use AI tools to brainstorm ideas, refine concepts, or stimulate their creativity. What about students? There are two extreme approaches to answering this question, and the whole spectrum of more moderate approaches logically falling between them. One extreme approach is to entirely ban the use of AI tools in class-works and home-works, the second to tolerate such practice without any constrains. None of them can ignore the facts of academic life concerning actual use of AI tools by students. In December 2024, the Higher Education Policy Institute (London, UK) surveyed over 1 000 full-time undergraduate students of British institutions of higher education about the use of generative AI tools. The key finding reported is that 92% of them use such tools for various purposes: generating text (64%), enhancing and editing text (39%), summarising and note-taking (36%), translation or language support (35%), speech-to-text transcription (24%), generating images, videos or audio (19%), data analysis and presentation (15%) or writing computer code (15%) [61]. It is clear that any attempt to constrain this kind of students activities would be doomed to failure. A rational and constructive reply of academic staff and institutions should be rather an offer of diversified forms of education and training oriented on teaching safe and efficient use of AI tools without violating legal and ethical norms.

A global debate on the use of AI tools in academic education seams to give a slight preference to an intermediate solution based on somewhat nuanced ethical evaluation of the role of an Al tool in creative processes, viz. on the answer to the question whether this tool plays the role of an assistant to students or it replaces their genuine creativity. Of course, this demarcation may seem, very fuzzy, but it can be a good starting point for formulating more specific recommendations, in particular – the requirements concerning the role of an AI tool as an assistant. A tentative characterisation of that role could be as follows. A tool under consideration is an assistant if:
– it helps students to generate new ideas by providing inspiration, but not replacing them in making final decisions;
– it assists them in polishing a draft of a document, improving its logical coherence and suggesting editorial corrections, but not deciding about its core contents;
– it is a helpful collaborator providing 'raw materials' and suggesting options of their processing, but never dictating the final shape of an intellectual product.
The ethical evaluation of an AI-aided work is always less problematic if its author follows the rule that every 'borrowed' element of creative work is accompanied by a reference to its original source. This expectation may be met by appropriate prompting – a skill that should be developed by students under wise supervision of academic teachers. In this respect the AI tool called Perplexity (https://www.perplexity.ai/) is quite exceptional because, by principle, it draws on reliable sources from the internet and cites them for verification by its users [62].

Numerous methodologies of supporting educational processes with AI tools, overviewed in the literature referred to in this section, are potentially useful in teaching MS&T. Any adaptation of those methodologies to specific needs of teaching MS&T should, however, highlight ethical aspects of Al implementation – cf., for example, [63–68] – and serve enhancing students' skills of critical thinking – cf., for example, [69] and [70]. This recommendation is valid not only for the proponents of new courses in MS&T, but also for the academic teachers involved in the development of new curricula in other domains of technoscience. It will be, therefore, not considered here in more detail. The next section will focus on specific aspects of designing MS&T courses.

We - teachers of measurement science and technology – need some deep reflection on the educational implications of accelerated penetration of Al tools to all the domains of technoscience and socio-economic life. The current practice of teaching MS&T is still predominantly based on a traditional (naïve) understanding of measurement, guided by the following premises:
– measurement is the only reliable source of information about material reality;
– mathematical modelling of material entities, being the essence of technoscience, entirely relies on measurement;
– providing data for mathematical modelling is just one of the numerous applications of measurement;
– evaluation of measurement uncertainty consists in comparing the actual result of measurement with universally agreed etalons or standards.
There are still in use the textbooks on MS&T, such as [71], which follow approaches of the 1960s and 1970s focused on extensive catalogues of MS&T components, sometimes quite well classified. Such approaches were criticised by Ludwik Finkelstein already in 1983: he pointed out that teaching MS&T should be based on general principles and logical reasoning rather than on the catalogues of ready for use solutions [72]. Although, the search for unifying concepts and principles of MS&T was undertaken in the 20th century, the first systematic elaboration of this topic in the form of a monograph appeared only in 2021 [73].

There is no doubt that in the age of pervasive AI, we should put more emphasis on concepts, principles, logical reasoning, and critical thinking. Let us support this statement with an example. The three-dimensional taxonomy of measurement sensors includes their categorisation according to (1) the measured quantity, (2) the operating principle, and (3) the application. The first category contains at least 14 types of sensors (acceleration sensors, flow sensors, gas sensors, humidity sensors, level sensors, light sensors, magnetic-field sensors, nuclear sensors, position sensors, pressure sensors, proximity sensors, sound sensors, strain sensors, and temperature sensors). The second category contains at least 7 kinds of sensors (capacitive sensors, inductive sensors, magnetic sensors, optical sensors, piezoelectric sensors, resistive sensors, and ultrasonic sensors). The third category contains at least 7 groups of sensors (aerospace sensors, automotive sensors, biomedical sensors, consumer electronics sensors, environmental sensors, industrial sensors, and military sensors). Of course, the students of MS&T should be aware of this structured taxonomy of sensors, but not necessarily learn exhaustive specification of each type, each kind and each group of sensors. They will be able to autonomously acquire and understand the exhaustive specification of any type, any kind or any group of sensors, provided they are equipped with adequate patterns of their characterisation based on their mathematical models. This is just an example, but a very important one since sensors are the most distinctive parts of measuring systems. Convergent technologies contribute to relatively quick unification and standardisation of the other constitutive elements such systems, while the progress of unification and standardisation of sensors is much slower.

The protagonists of the traditional understanding of measurement, as a rule, ignore the dependence of measurement on mathematical models, more precisely – the roles of those models in defining measurands, calibrating measurement channels, estimating measurands and evaluating measurement uncertainty, although – as already mentioned in Section 1 – basic tools of AI, viz. universal approximators (such as neural networks), have been already for years used in modelling measurement channels, and consequently – in performing those operations.

Thinking in terms of semantic and mathematical models of material entities (phenomena, objects or processes) is a natural way of dealing with the problems of MS&T in the age of ubiquitous use of AI tools which are, as a rule, semantic or mathematical models: large language models or universal approximators. Mathematical modelling of a material entity is always preceded by its semantic modelling. The latter is an operation of fundamental importance from an epistemic point of view because it consists in crossing the boundary between the sphere of material entities and the sphere of their abstract images. Since mathematical modelling is indispensable for defining measurands, designing measuring systems and evaluating measurement uncertainty, it may be considered a meta-concept for building educational frameworks in the field of MS&T. Detailed descriptions of the instances of mathematical modelling in MS&T should be included in the corresponding undergraduate courses, and a general methodology of mathematical modelling – in the relevant graduate courses.

The meta-concept of mathematical modelling is closely related to another meta-concept of key importance, viz. understanding – in the sense of our ability to recognise the properties of a material entity and to use the knowledge thus obtained for solving related practical problems. The latter clarification is important because the word understanding may be ambiguous: its meaning significantly depends on the context [74], and is a subject of deep philosophical considerations – cf. [75] or [76]. In teaching MS&T, its interpretation as the ability to recognise causal relationships and to use the knowledge thus obtained for solving measurement-related problems of practical nature is satisfactory in the majority of cases; some weakly defined measurements [77], however, cannot be properly understood under this constraint imposed on the definition of understanding.

The art of modelling causal relationships is of key importance both for the methodology of designing measuring systems and for making measurement-based decisions of ethical nature (to be addressed in Section 4). The existence of cause-effect relationships in material reality cannot be proven, neither by logical nor by experimental means. However, taking into account the practical productivity of reasoning guided by such relationships, we accept a paradigm of causality and make endeavours to built semantic and mathematical models of material entities, which are assumed to represent those relationships. Discovering causal relations and making use of them is not only a distinctive feature of understanding but also of human intelligence in general. It should be noted that there is no closed-form definition of causality that would not provoke any logical or philosophical objections. There is, however, a vast practical experience concerning identification of simple networks of causal relationships, especially – the networks without loops. The currently available AI tools, although very useful as human assistants in this respect, fail to meet our expectations driven by the needs related to engineering design (in particular – the design of measuring systems and devices) and to the ethical decision-making (in particular - those related to measurement applications). Judea Pearl, an Israeli-American computer scientist and philosopher who in the 1980s pioneered the development of causal models in Al in the form of so-called Bayesian networks, concluded in 2018 that future AI systems must incorporate causal models to reach the level of human-like intelligence [78]. It seems that achieving this goal is harder than expected; therefore, causality continues to be an active and important topic in AI-related research. The identification of multidimensional causal networks with multiple loops is a challenge that should be urgently undertaken for the sake of safe use of AI tools, especially those based on large language models, in applications directly concerning human lives and wellbeing now and in the future. The functioning of large language models is based on correlation-type relationships among linguistic items; the supervision of their operation based on deductive reasoning and causal relationships is needed to make this functioning safe and reliable. This is why teaching students of MS&T the art of causal modelling is so critically indispensable.

A general methodology for mathematical modelling of measurement-related entities is systematically presented in Chapter 6 of the author's handbook Technoscientific Research: Methodological and Ethical Aspects [79]. It is a refined version of the methodology proposed by the author in 2013 [80] – the version that takes into account his later contributions, viz. the articles [81] and [82]. A meta-model of measurement described there includes generic mathematical models indispensable for defining measurands, calibrating measurement channels, and providing measurement results in the form of measurands' estimates accompanied by some indicators of their uncertainty. Each of those models is to be concretised in any specific case of measurement. Of course, the AI tools, such as universal approximators, can be used for this purpose. However, there are some limitations in their more-or-less routine application, viz. limitations due to the fact that measurand estimation and uncertainty evaluation are operations of abductive nature. Unlike deductive reasoning, abductive reasoning is highly uncertain, and its uncertainty cannot be mitigated solely by increasing the volume of data (like in the case of inductive reasoning), but additional information about the material (physical, chemical, biological...) background of measurement under consideration is necessary to sufficiently constrain the set of admissible estimates of a measurand [83]. Unfortunately, abductive reasoning is an underdeveloped aspect of the today's AI tools which are predominantly based on inductive reasoning, i.e. on data-voracious machine learning, and which are not creative enough to solve non-trivial problems of abduction as argued in Part II of the book [84]. The advancement in this area is a sine qua non condition of the full development of explainable AI, and the progress in the latter area is a sine qua non condition of the progress in the development of trustworthy AI, indispensable for predictable and controllable expansion of global-scale applications of AI.

For the time being, there is a lot of research activities oriented on the advancement of explainable AI. Enough to say that a 111-page article [85], providing a systematic review of the literature on applications of explainable AI, refers to 512 articles which were published in peer-reviewed journals only in the years 2021–2023. This is a fact of importance for the development of MS&T curricula, since the use of the tools of explainable AI for designing measuring systems, both in the role of the design tools and their elements, enables the reliable evaluation of the AI-generated quantitative components of measurement uncertainty. However, these tools may occur insufficient for assessing total measurement uncertainty, that includes also fundamental epistemic uncertainty, as this would require taking into account some qualitative - mainly psychosociological and humanistic – aspects of AI development towards trustworthy AI. Related topics should be urgently incorporated in MS&T curricula.

One more argument in favour of enhancing the intellectual toolbox of MS&T graduates with the skills related to abductive reasoning follows form the realisation of the importance of this type of reasoning in technoscience and everyday life: it is the basis of any non-trivial measurement [86], of scientific explanation, of medical and technical diagnostics, of criminal investigation, and of reliable communication between two human beings. Its creative potential cannot be fully developed without combining complementary strong features of human and artificial intelligence. The following example is quite indicative in this respect. In a study conducted at Stanford University and reported in the 2024 article [87], an AI system used as a diagnostic tool outperformed a group of 50 physicians. It turned out, however, that the availability of that system to the latter did not significantly improve their clinical reasoning. This finding indicates the need for significant advancement of research on the mode of 'collaboration' between human and artificial intelligence, which is impossible without fusion of technical and non-technical competences in the engineering curricula, in particular – in the MS&T-related curricula.

The situation outlined in the previous section may engender problems of ethical nature that were identified already 80 years ago, but have been generally recognised only recently in the context of the global discussion over intelligent text robots such as ChatGPT. In particular, this discussion has exposed some problems less realised in the past, viz. the ones related to legal protection of intellectual property. The 2024 Nobel Prize in Chemistry is an excellent example showing how insightful the Noam Chomsky's remark (quoted in Section 2) is since that prize was awarded for the computational design of proteins and the prediction of their structure, aided by an AI tool called AlphaFold2. From the ethical perspective, that decision may be perceived as problematic because the awarded achievement is based on an AI-model trained on the results of long-run research of numerous scientists all over the world not just on the research contributions of the awardees.

Many other ethical problems may directly affect MS&T – for example, when measuring instrumentation is applied in biomedical engineering. In the latter case, the evaluation of uncertainty must be precise enough to exclude the risk of harmful measurement-based decisions concerning human health or wellbeing. Less evident may seem the risks related to unreliable measurements embedded in such technical objects as autonomous vehicles, social robots or industrial control systems.

Over the 25 centuries of development of Western philosophy, many proposals for systematically solving moral problems have appeared. Fortunately, their deeper analysis leads to the conclusion that each of them can be viewed as a combination of three monistic ethics:
– virtue ethics which considers an act to be morally good if it is performed by a virtuous person;
– deontological ethics which considers an act to be morally good if it constitutes the fulfilment of a duty or law;
– consequentialist ethics which considers an act to be morally good if its consequences are good.
The importance of the above components differs from one ethical system to another, but regardless of their proportions a universal methodology of making morally significant decisions may be formulated – the methodology inspired by the practice of making engineering decisions based on mono-criterial optimisation. The key for its formulation is the analogy linking:
– the consequences of our actions with optimisation criteria;
– the relevant moral norms with optimisation constraints;
– the virtues of decision-makers with the numerical advantages of optimisation algorithms.
Using this analogy, we may postulate that – when faced with the need to undertake a morally significant decision one should make the best choice in terms of the anticipated consequences, but only within the space of admissible solutions, defined by relevant moral norms. The practical implementation of this pattern of decision-making requires the use of causal models for prediction of consequences of an act and a catalogue of norms generating constraints to be satisfied by that act. Numerous attempts to compile such a catalogue have been made – at least since 1942 when Isaac Asimov, in a short story published in Astounding Science Fiction magazine, proposed three laws of robot ethics. Over last eighty years, numerous organisations have proposed various sets of principles for ethical AI [88]; among them, a set of 23 recommendations for developers of AI systems in a document launched during the Al conference organised in 2017 in Asilomar (California, USA) by the Future of Life Institute, and called 'Asilomar AI Principles' [89]. The Canadian philosopher of science and medicine, Paul R. Thagard, has demonstrated [90] that they may be logically derived from four principles of bioethics, introduced in 1978 by the American ethicists Thomas L. Beauchamp and James F. Childress [91], viz.: (1) the principle of beneficence, which obliges us to maximise the health benefits of patients; (2) the principle of non-maleficence, which obliges us to minimise the harm associated with experimental activities with their involvement; (3) the principle of respect for autonomy, which obliges us to respect patients' decisions regarding their own lives; (4) the principle of justice, which obliges us to give patients everything they deserve, to treat them equally, fairly and impartially.

The main advantage of the above-outlined methodology of solving morally significant problems in AI-aided engineering, in particular – in engineering related to MS&T – consists in its high potential for automatic implementation. Moreover, this methodology is helpful in integrating ethical issues with technoscientific issues in the areas of morally sensitive applications of measurement, such as social robotics or biomedical monitoring. It enables a productive collaboration of engineers with professional ethicists, based on a common language resulting from the above-proposed analogy between constrained optimisation and making morally significant decisions.

The recognition of the need to include ethical considerations in engineering education seems to be a harbinger of the imminent recognition of the need to significantly enhance technoscientific curricula with psychosociological and humanistic contents – to overcome a divide between sciences and humanities that appeared in the 19th century and deepened in the 20th century as a result of overspecialisation. Already in the late 1960s, Charles P. Snow, in his lecture The Two Cultures and the Scientific Revolution [92], formulated the thesis about the incompatibility of two cultures having common historical roots in Western philosophy, viz. sciences and humanities. He argued that the representatives of these cultures had ceased to understand each other, and warned about the dangerous consequences of the deepening intellectual gap between them. Over the past six decades, Charles P. Snow's thesis has been the subject of critical discussion, which has made many thinkers aware of the need to reintegrate sciences and humanities. However, it was not until the age of ubiquitous use of AI tools that awareness of this need became common. The progressing convergence of technoscientific disciplines critically dependent on IT tools (including AI tools) requires an adequate convergence of technoscientific and humanistic education at all its stages – the introduction of selected elements of humanistic knowledge to technoscientific curricula and selected elements of technoscientific knowledge to educational programmes in humanities. Sustainable development and implementation of modern technologies is impossible without close interdisciplinary collaboration which involves not only experts in various domains of engineering, but also representatives of social sciences (especially psychologists and sociologists) and humanities (especially ethicists and philosophers of technoscience). In order to effectively communicate, they must master a kind of common language. This is only a very brief ad hoc justification of the need to reintegrate technoscience and humanities in academic education, but there are also some deeper reasons (of philosophical nature) supporting this idea.

The essence of the unity of sciences and humanities has been accurately grasped on the portal ScienceFather: a synergy resulting from the integration of humanities and sciences implies the fusion of the analytical rigor of scientific inquiry with the nuanced insights of humanistic inquiry, and in a holistic understanding of the world encompassing both empirical knowledge and human experience [93]. Albert Einstein justified this essence by referring to the Western intellectual tradition. According to him 'All religions, arts and sciences are branches of the same tree': their aim is to spread moral and cultural understanding without referring to force [94]. Further justification may be found in the preface to the 2018 report The Integration of the Humanities and Arts with Sciences, Engineering, and Medicine in Higher Education, where its authors recommend – despite limited causal evidence on the impact of integration on students – that further effort be made to develop and disseminate a variety of approaches to integrated education and to carry out research on the impact of such forms of education on students [95].

Examples of partial integration of technoscience and humanities in academic education may be already found in many Western institutions of higher education. Here are selected examples:
– Lehigh University (Bethlehem, Pennsylvania, USA) offers an undergraduate programme leading to the Integrated Degree in Engineering, Arts and Sciences. It is dedicated to students with high creative potential – future innovators able to use different styles of thinking ('renaissance thinking for the technological era') [96].
– The graduate programmes in Human-centred Artificial Intelligence are offered by several European universities (Göteborgs Universitet [97], Vrije Universiteit Amsterdam [98], Università degli studi di Milano [99], as well as by the consortium including Technological University Dublin, Università degli Studi di Napoli Federico II, De Hogeschool Utrecht, Budapesti Műszaki és Gazdaságtudományi Egyetem [100]); the corresponding curricula provide students with a deep understanding of how Al is transforming society, as well as with a broad theoretical preparation for analysing the consequences of the proliferation of Al tools.
Such educational offers are multiplying as IT employers increasingly value interdisciplinary competences in their employees, such as the ability to clearly communicate thoughts (both in written and oral form), productivity in teamwork, the ability to make morally significant decisions and the ability to apply knowledge for solving complex interdisciplinary problems – cf. [101]. They expect that graduates of engineering studies will not only be able to take up employment in a profession corresponding to the profile of their education, but also will manage to adapt their competences to changing needs in the following years of their professional careers – cf. [102].

According to the report about the global labour market, published in 2023 under the auspices of the World Economic Forum, over 60% of surveyed employers indicate the growing importance of the following competencies: creative thinking, analytical thinking, technological literacy, curiosity and lifelong learning and the broadly understood ability to respond flexibly (resilience, flexibility and agility) – cf. [103]. It seems that in today's approach to technoscientific education, including teaching of MS&T, we are inclined to overvalue purely technical skills and undervalue skills from the other four categories – the skills relying on the art of asking questions.

Why is asking questions so important today? In times of gigantic databases and intelligent search engines, in times of ubiquitous sensor networks – collecting measurement data that may be useful to someone somewhere-sometime – the ability to extract the useful and reliable information from these resources and to assess its credibility is becoming a key competence. Asking good questions is a form of creativity, the value of which we appreciate when using AI tools such as ChatGPT. The socio-economic importance of new professionals, called prompters, is growing rapidly. As a rule, holders of diplomas in humanities are better candidates for this profession than engineers, but this observation should not imply a conclusion that an engineer cannot acquire prompter qualifications by attending classes on logic, rhetoric, language culture or propaedeutics of philosophy.

Being aware that modern economies are based on knowledge, we must recognise that: (1) verified (or warranted) information and the methodology for obtaining it are crucial for the lifelong professional career of an engineer; (2) mastering this methodology is more important than learning detailed specialist contents that will, in practice, become outdated within three to five years after their appearance. The ability to learn independently and to define learning goals independently – these are the main competences of the future, with which we should equip our graduates. The implementation of this postulate requires very careful selection of the contents to be memorised a selection that gives priority to structural and methodological knowledge because this type of knowledge plays a fundamental role in finding and verifying information on any topic, and in particular – in asking questions skilfully.

In the 1990s, efficient use of computers and measurement instrumentation was impossible without certain hardware- and software-related skills. Today, such competences are not needed, user manuals have become obsolete; we expect that, when carrying out a specific task, we will communicate with a measuring system via an intelligent multimedia-based interface enabling us not only to operate this system, but also to compensate for our lack of precision in operation, to suggest a spectrum of alternative ways of operation, etc. Designing such an interface requires creative cooperation between an instrumentation engineer and a person whose competences include verbal and iconic communication, psychology of human needs and reflexes, as well as a methodology of evaluating new technologies from the point of view of ethics and ergonomics.

We - teachers of measurement science and technology – need some deep reflection on the educational implications of accelerated penetration of Al tools to all the domains of technoscience and socio-economic life. The current practice of teaching MS&T is still predominantly based on a traditional (naïve) understanding of measurement, guided by the following premises:
– measurement is the only reliable source of information about material reality;
– mathematical modelling of material entities, being the essence of technoscience, entirely relies on measurement;
– providing data for mathematical modelling is just one of the numerous applications of measurement;
– evaluation of measurement uncertainty consists in comparing the actual result of measurement with universally agreed etalons or standards.
There are still in use the textbooks on MS&T, such as [71], which follow approaches of the 1960s and 1970s focused on extensive catalogues of MS&T components, sometimes quite well classified. Such approaches were criticised by Ludwik Finkelstein already in 1983: he pointed out that teaching MS&T should be based on general principles and logical reasoning rather than on the catalogues of ready for use solutions [72]. Although, the search for unifying concepts and principles of MS&T was undertaken in the 20th century, the first systematic elaboration of this topic in the form of a monograph appeared only in 2021 [73].

There is no doubt that in the age of pervasive AI, we should put more emphasis on concepts, principles, logical reasoning, and critical thinking. Let us support this statement with an example. The three-dimensional taxonomy of measurement sensors includes their categorisation according to (1) the measured quantity, (2) the operating principle, and (3) the application. The first category contains at least 14 types of sensors (acceleration sensors, flow sensors, gas sensors, humidity sensors, level sensors, light sensors, magnetic-field sensors, nuclear sensors, position sensors, pressure sensors, proximity sensors, sound sensors, strain sensors, and temperature sensors). The second category contains at least 7 kinds of sensors (capacitive sensors, inductive sensors, magnetic sensors, optical sensors, piezoelectric sensors, resistive sensors, and ultrasonic sensors). The third category contains at least 7 groups of sensors (aerospace sensors, automotive sensors, biomedical sensors, consumer electronics sensors, environmental sensors, industrial sensors, and military sensors). Of course, the students of MS&T should be aware of this structured taxonomy of sensors, but not necessarily learn exhaustive specification of each type, each kind and each group of sensors. They will be able to autonomously acquire and understand the exhaustive specification of any type, any kind or any group of sensors, provided they are equipped with adequate patterns of their characterisation based on their mathematical models. This is just an example, but a very important one since sensors are the most distinctive parts of measuring systems. Convergent technologies contribute to relatively quick unification and standardisation of the other constitutive elements such systems, while the progress of unification and standardisation of sensors is much slower.

The protagonists of the traditional understanding of measurement, as a rule, ignore the dependence of measurement on mathematical models, more precisely – the roles of those models in defining measurands, calibrating measurement channels, estimating measurands and evaluating measurement uncertainty, although – as already mentioned in Section 1 – basic tools of AI, viz. universal approximators (such as neural networks), have been already for years used in modelling measurement channels, and consequently – in performing those operations.

Thinking in terms of semantic and mathematical models of material entities (phenomena, objects or processes) is a natural way of dealing with the problems of MS&T in the age of ubiquitous use of AI tools which are, as a rule, semantic or mathematical models: large language models or universal approximators. Mathematical modelling of a material entity is always preceded by its semantic modelling. The latter is an operation of fundamental importance from an epistemic point of view because it consists in crossing the boundary between the sphere of material entities and the sphere of their abstract images. Since mathematical modelling is indispensable for defining measurands, designing measuring systems and evaluating measurement uncertainty, it may be considered a meta-concept for building educational frameworks in the field of MS&T. Detailed descriptions of the instances of mathematical modelling in MS&T should be included in the corresponding undergraduate courses, and a general methodology of mathematical modelling – in the relevant graduate courses.

The meta-concept of mathematical modelling is closely related to another meta-concept of key importance, viz. understanding – in the sense of our ability to recognise the properties of a material entity and to use the knowledge thus obtained for solving related practical problems. The latter clarification is important because the word understanding may be ambiguous: its meaning significantly depends on the context [74], and is a subject of deep philosophical considerations – cf. [75] or [76]. In teaching MS&T, its interpretation as the ability to recognise causal relationships and to use the knowledge thus obtained for solving measurement-related problems of practical nature is satisfactory in the majority of cases; some weakly defined measurements [77], however, cannot be properly understood under this constraint imposed on the definition of understanding.

The art of modelling causal relationships is of key importance both for the methodology of designing measuring systems and for making measurement-based decisions of ethical nature (to be addressed in Section 4). The existence of cause-effect relationships in material reality cannot be proven, neither by logical nor by experimental means. However, taking into account the practical productivity of reasoning guided by such relationships, we accept a paradigm of causality and make endeavours to built semantic and mathematical models of material entities, which are assumed to represent those relationships. Discovering causal relations and making use of them is not only a distinctive feature of understanding but also of human intelligence in general. It should be noted that there is no closed-form definition of causality that would not provoke any logical or philosophical objections. There is, however, a vast practical experience concerning identification of simple networks of causal relationships, especially – the networks without loops. The currently available AI tools, although very useful as human assistants in this respect, fail to meet our expectations driven by the needs related to engineering design (in particular – the design of measuring systems and devices) and to the ethical decision-making (in particular - those related to measurement applications). Judea Pearl, an Israeli-American computer scientist and philosopher who in the 1980s pioneered the development of causal models in Al in the form of so-called Bayesian networks, concluded in 2018 that future AI systems must incorporate causal models to reach the level of human-like intelligence [78]. It seems that achieving this goal is harder than expected; therefore, causality continues to be an active and important topic in AI-related research. The identification of multidimensional causal networks with multiple loops is a challenge that should be urgently undertaken for the sake of safe use of AI tools, especially those based on large language models, in applications directly concerning human lives and wellbeing now and in the future. The functioning of large language models is based on correlation-type relationships among linguistic items; the supervision of their operation based on deductive reasoning and causal relationships is needed to make this functioning safe and reliable. This is why teaching students of MS&T the art of causal modelling is so critically indispensable.

A general methodology for mathematical modelling of measurement-related entities is systematically presented in Chapter 6 of the author's handbook Technoscientific Research: Methodological and Ethical Aspects [79]. It is a refined version of the methodology proposed by the author in 2013 [80] – the version that takes into account his later contributions, viz. the articles [81] and [82]. A meta-model of measurement described there includes generic mathematical models indispensable for defining measurands, calibrating measurement channels, and providing measurement results in the form of measurands' estimates accompanied by some indicators of their uncertainty. Each of those models is to be concretised in any specific case of measurement. Of course, the AI tools, such as universal approximators, can be used for this purpose. However, there are some limitations in their more-or-less routine application, viz. limitations due to the fact that measurand estimation and uncertainty evaluation are operations of abductive nature. Unlike deductive reasoning, abductive reasoning is highly uncertain, and its uncertainty cannot be mitigated solely by increasing the volume of data (like in the case of inductive reasoning), but additional information about the material (physical, chemical, biological...) background of measurement under consideration is necessary to sufficiently constrain the set of admissible estimates of a measurand [83]. Unfortunately, abductive reasoning is an underdeveloped aspect of the today's AI tools which are predominantly based on inductive reasoning, i.e. on data-voracious machine learning, and which are not creative enough to solve non-trivial problems of abduction as argued in Part II of the book [84]. The advancement in this area is a sine qua non condition of the full development of explainable AI, and the progress in the latter area is a sine qua non condition of the progress in the development of trustworthy AI, indispensable for predictable and controllable expansion of global-scale applications of AI.

For the time being, there is a lot of research activities oriented on the advancement of explainable AI. Enough to say that a 111-page article [85], providing a systematic review of the literature on applications of explainable AI, refers to 512 articles which were published in peer-reviewed journals only in the years 2021–2023. This is a fact of importance for the development of MS&T curricula, since the use of the tools of explainable AI for designing measuring systems, both in the role of the design tools and their elements, enables the reliable evaluation of the AI-generated quantitative components of measurement uncertainty. However, these tools may occur insufficient for assessing total measurement uncertainty, that includes also fundamental epistemic uncertainty, as this would require taking into account some qualitative - mainly psychosociological and humanistic – aspects of AI development towards trustworthy AI. Related topics should be urgently incorporated in MS&T curricula.

One more argument in favour of enhancing the intellectual toolbox of MS&T graduates with the skills related to abductive reasoning follows form the realisation of the importance of this type of reasoning in technoscience and everyday life: it is the basis of any non-trivial measurement [86], of scientific explanation, of medical and technical diagnostics, of criminal investigation, and of reliable communication between two human beings. Its creative potential cannot be fully developed without combining complementary strong features of human and artificial intelligence. The following example is quite indicative in this respect. In a study conducted at Stanford University and reported in the 2024 article [87], an AI system used as a diagnostic tool outperformed a group of 50 physicians. It turned out, however, that the availability of that system to the latter did not significantly improve their clinical reasoning. This finding indicates the need for significant advancement of research on the mode of 'collaboration' between human and artificial intelligence, which is impossible without fusion of technical and non-technical competences in the engineering curricula, in particular – in the MS&T-related curricula.

The situation outlined in the previous section may engender problems of ethical nature that were identified already 80 years ago, but have been generally recognised only recently in the context of the global discussion over intelligent text robots such as ChatGPT. In particular, this discussion has exposed some problems less realised in the past, viz. the ones related to legal protection of intellectual property. The 2024 Nobel Prize in Chemistry is an excellent example showing how insightful the Noam Chomsky's remark (quoted in Section 2) is since that prize was awarded for the computational design of proteins and the prediction of their structure, aided by an AI tool called AlphaFold2. From the ethical perspective, that decision may be perceived as problematic because the awarded achievement is based on an AI-model trained on the results of long-run research of numerous scientists all over the world not just on the research contributions of the awardees.

Many other ethical problems may directly affect MS&T – for example, when measuring instrumentation is applied in biomedical engineering. In the latter case, the evaluation of uncertainty must be precise enough to exclude the risk of harmful measurement-based decisions concerning human health or wellbeing. Less evident may seem the risks related to unreliable measurements embedded in such technical objects as autonomous vehicles, social robots or industrial control systems.

Over the 25 centuries of development of Western philosophy, many proposals for systematically solving moral problems have appeared. Fortunately, their deeper analysis leads to the conclusion that each of them can be viewed as a combination of three monistic ethics:
– virtue ethics which considers an act to be morally good if it is performed by a virtuous person;
– deontological ethics which considers an act to be morally good if it constitutes the fulfilment of a duty or law;
– consequentialist ethics which considers an act to be morally good if its consequences are good.
The importance of the above components differs from one ethical system to another, but regardless of their proportions a universal methodology of making morally significant decisions may be formulated – the methodology inspired by the practice of making engineering decisions based on mono-criterial optimisation. The key for its formulation is the analogy linking:
– the consequences of our actions with optimisation criteria;
– the relevant moral norms with optimisation constraints;
– the virtues of decision-makers with the numerical advantages of optimisation algorithms.
Using this analogy, we may postulate that – when faced with the need to undertake a morally significant decision one should make the best choice in terms of the anticipated consequences, but only within the space of admissible solutions, defined by relevant moral norms. The practical implementation of this pattern of decision-making requires the use of causal models for prediction of consequences of an act and a catalogue of norms generating constraints to be satisfied by that act. Numerous attempts to compile such a catalogue have been made – at least since 1942 when Isaac Asimov, in a short story published in Astounding Science Fiction magazine, proposed three laws of robot ethics. Over last eighty years, numerous organisations have proposed various sets of principles for ethical AI [88]; among them, a set of 23 recommendations for developers of AI systems in a document launched during the Al conference organised in 2017 in Asilomar (California, USA) by the Future of Life Institute, and called 'Asilomar AI Principles' [89]. The Canadian philosopher of science and medicine, Paul R. Thagard, has demonstrated [90] that they may be logically derived from four principles of bioethics, introduced in 1978 by the American ethicists Thomas L. Beauchamp and James F. Childress [91], viz.: (1) the principle of beneficence, which obliges us to maximise the health benefits of patients; (2) the principle of non-maleficence, which obliges us to minimise the harm associated with experimental activities with their involvement; (3) the principle of respect for autonomy, which obliges us to respect patients' decisions regarding their own lives; (4) the principle of justice, which obliges us to give patients everything they deserve, to treat them equally, fairly and impartially.

The main advantage of the above-outlined methodology of solving morally significant problems in AI-aided engineering, in particular – in engineering related to MS&T – consists in its high potential for automatic implementation. Moreover, this methodology is helpful in integrating ethical issues with technoscientific issues in the areas of morally sensitive applications of measurement, such as social robotics or biomedical monitoring. It enables a productive collaboration of engineers with professional ethicists, based on a common language resulting from the above-proposed analogy between constrained optimisation and making morally significant decisions.

The recognition of the need to include ethical considerations in engineering education seems to be a harbinger of the imminent recognition of the need to significantly enhance technoscientific curricula with psychosociological and humanistic contents – to overcome a divide between sciences and humanities that appeared in the 19th century and deepened in the 20th century as a result of overspecialisation. Already in the late 1960s, Charles P. Snow, in his lecture The Two Cultures and the Scientific Revolution [92], formulated the thesis about the incompatibility of two cultures having common historical roots in Western philosophy, viz. sciences and humanities. He argued that the representatives of these cultures had ceased to understand each other, and warned about the dangerous consequences of the deepening intellectual gap between them. Over the past six decades, Charles P. Snow's thesis has been the subject of critical discussion, which has made many thinkers aware of the need to reintegrate sciences and humanities. However, it was not until the age of ubiquitous use of AI tools that awareness of this need became common. The progressing convergence of technoscientific disciplines critically dependent on IT tools (including AI tools) requires an adequate convergence of technoscientific and humanistic education at all its stages – the introduction of selected elements of humanistic knowledge to technoscientific curricula and selected elements of technoscientific knowledge to educational programmes in humanities. Sustainable development and implementation of modern technologies is impossible without close interdisciplinary collaboration which involves not only experts in various domains of engineering, but also representatives of social sciences (especially psychologists and sociologists) and humanities (especially ethicists and philosophers of technoscience). In order to effectively communicate, they must master a kind of common language. This is only a very brief ad hoc justification of the need to reintegrate technoscience and humanities in academic education, but there are also some deeper reasons (of philosophical nature) supporting this idea.

The essence of the unity of sciences and humanities has been accurately grasped on the portal ScienceFather: a synergy resulting from the integration of humanities and sciences implies the fusion of the analytical rigor of scientific inquiry with the nuanced insights of humanistic inquiry, and in a holistic understanding of the world encompassing both empirical knowledge and human experience [93]. Albert Einstein justified this essence by referring to the Western intellectual tradition. According to him 'All religions, arts and sciences are branches of the same tree': their aim is to spread moral and cultural understanding without referring to force [94]. Further justification may be found in the preface to the 2018 report The Integration of the Humanities and Arts with Sciences, Engineering, and Medicine in Higher Education, where its authors recommend – despite limited causal evidence on the impact of integration on students – that further effort be made to develop and disseminate a variety of approaches to integrated education and to carry out research on the impact of such forms of education on students [95].

Examples of partial integration of technoscience and humanities in academic education may be already found in many Western institutions of higher education. Here are selected examples:
– Lehigh University (Bethlehem, Pennsylvania, USA) offers an undergraduate programme leading to the Integrated Degree in Engineering, Arts and Sciences. It is dedicated to students with high creative potential – future innovators able to use different styles of thinking ('renaissance thinking for the technological era') [96].
– The graduate programmes in Human-centred Artificial Intelligence are offered by several European universities (Göteborgs Universitet [97], Vrije Universiteit Amsterdam [98], Università degli studi di Milano [99], as well as by the consortium including Technological University Dublin, Università degli Studi di Napoli Federico II, De Hogeschool Utrecht, Budapesti Műszaki és Gazdaságtudományi Egyetem [100]); the corresponding curricula provide students with a deep understanding of how Al is transforming society, as well as with a broad theoretical preparation for analysing the consequences of the proliferation of Al tools.
Such educational offers are multiplying as IT employers increasingly value interdisciplinary competences in their employees, such as the ability to clearly communicate thoughts (both in written and oral form), productivity in teamwork, the ability to make morally significant decisions and the ability to apply knowledge for solving complex interdisciplinary problems – cf. [101]. They expect that graduates of engineering studies will not only be able to take up employment in a profession corresponding to the profile of their education, but also will manage to adapt their competences to changing needs in the following years of their professional careers – cf. [102].

According to the report about the global labour market, published in 2023 under the auspices of the World Economic Forum, over 60% of surveyed employers indicate the growing importance of the following competencies: creative thinking, analytical thinking, technological literacy, curiosity and lifelong learning and the broadly understood ability to respond flexibly (resilience, flexibility and agility) – cf. [103]. It seems that in today's approach to technoscientific education, including teaching of MS&T, we are inclined to overvalue purely technical skills and undervalue skills from the other four categories – the skills relying on the art of asking questions.

Why is asking questions so important today? In times of gigantic databases and intelligent search engines, in times of ubiquitous sensor networks – collecting measurement data that may be useful to someone somewhere-sometime – the ability to extract the useful and reliable information from these resources and to assess its credibility is becoming a key competence. Asking good questions is a form of creativity, the value of which we appreciate when using AI tools such as ChatGPT. The socio-economic importance of new professionals, called prompters, is growing rapidly. As a rule, holders of diplomas in humanities are better candidates for this profession than engineers, but this observation should not imply a conclusion that an engineer cannot acquire prompter qualifications by attending classes on logic, rhetoric, language culture or propaedeutics of philosophy.

Being aware that modern economies are based on knowledge, we must recognise that: (1) verified (or warranted) information and the methodology for obtaining it are crucial for the lifelong professional career of an engineer; (2) mastering this methodology is more important than learning detailed specialist contents that will, in practice, become outdated within three to five years after their appearance. The ability to learn independently and to define learning goals independently – these are the main competences of the future, with which we should equip our graduates. The implementation of this postulate requires very careful selection of the contents to be memorised a selection that gives priority to structural and methodological knowledge because this type of knowledge plays a fundamental role in finding and verifying information on any topic, and in particular – in asking questions skilfully.

In the 1990s, efficient use of computers and measurement instrumentation was impossible without certain hardware- and software-related skills. Today, such competences are not needed, user manuals have become obsolete; we expect that, when carrying out a specific task, we will communicate with a measuring system via an intelligent multimedia-based interface enabling us not only to operate this system, but also to compensate for our lack of precision in operation, to suggest a spectrum of alternative ways of operation, etc. Designing such an interface requires creative cooperation between an instrumentation engineer and a person whose competences include verbal and iconic communication, psychology of human needs and reflexes, as well as a methodology of evaluating new technologies from the point of view of ethics and ergonomics.

The conclusions presented in this section are intended to integrate key elements of the emerging paradigm of MS&T teaching announced in the title of the article – the elements that have been introduced and analysed in the previous sections. Many of them are not specific to MS&T, and can be applied to teaching students of various subfields of IT – and even of those subfields of engineering which are not focused on IT. This is one more illustration of the phenomenon called convergence of technologies.

Taking into account the omnipresence and key role of measurement in industry, medicine, economy and other sectors of social practice, we teachers of MS&T – should recognise the pragmatic and ethical implications of the evolving technological and socio-economic environment of educational practice. The corresponding evolution of the engineering curricula (rather than of isolated courses devoted to MS&T) should consist in inclusion of the following contents:
– general methodology of mathematical modelling of material entities;
– procedures of effective use of mathematical models in conceptual and functional design of measuring systems;
– strategies of responsible use of AI tools for this purpose;
– elements of research-and-engineering ethics, enabling the dialogue between engineers and professional ethicists over AI-induced moral issues.
The scope and complexity of the above-listed contents should be fit to the level of instruction: they should be more basic and illustrative at the B.Sc. level, and significantly more advanced and abstract at the Ph.D. level. The range and strategy of introducing postulated changes should also be fit to the local educational environment. A considerable inertia of curricula evolution is characteristic of most academic institution, but this cannot be used as a justification for procrastination. The appropriate change of selected syllabi, which as a rule is easier from institutional perspective, may be a good anticipatory step preparing the ground for reforming an entire curriculum.

If the use of mathematical models is concerned, the meta-model of measurement, referred to in Section 3, could become a convenient platform for incorporation of the methodology of mathematical modelling into MS&T courses. If the ethical contents are concerned, Chapters 11-20 of the textbook [104] could be a good starting point even at the B.Sc. level studies. The first-semester students of cybersecurity and the students of internet of things, exposed to such contents at the author's institution, are able to grasp them; so, it is reasonable to expect that MS&T students at the B.Sc. level will be able to follow their suite.

Over the past decades, we have come to realise that a very limited share of MS&T graduates are going to effectively use knowledge and skills concerning the physical phenomena underlying functioning of sensors and other hardware components of measurement instrumentation, while a great many of them will need some competencies concerning methodology of mathematical modelling and data processing, in particular – of using AI-based algorithms. It would be, however, a mistake to deprive the MS&T students of any opportunity to gain understanding of those phenomena – not only for philosophical reasons, but also for the need to prepare them to effectively communicate, over their professional careers, with the specialists of empirical sciences.

We have already learned that many operations related to the design and utilisation of measuring instrumentation may be ceded to Al systems, in particular, such operations as searching databases of technical information about sensors and other hardware components of measuring systems, optimal fitting of those components to the design requirements, automatic calibration of measurement channels, automatic generation of technical documentation and its translation in various languages, etc. Consequently, the skills related to the efficient use of AI tools are getting increasingly more important than memorising catalogues of hardware and software components of measuring instrumentation; the ability to use general methodologies of designing technical objects and general principles of designing measuring systems are getting increasingly more important than technical data of measurement bridges or structures of analogue-to-digital and digital-to-analogue converters. We – teachers of measurement science and technology – should realise that 'Academia needs to adapt to a more AI-integrated system [since] Al assistance is not a temporary trend, but a tendency of technological progress, which should be treated with cautious openness' [105]; that already today advanced Al tools contribute to the most creative activities of technoscientific research – cf., for example, [106]. Several advanced AI engineering laboratories all over the world are working on developing AI tools intended to support the research process carried out according to the so-called scientific method, the essence of which is generating hypotheses and their experimental verification. A good example is AI co-scientist a multi-agent system designed as a collaborative tool for scientists [107]. The advanced, specialised AI tools do not eliminate general-purpose tools, like Microsoft Copilot, from everyday research practice; for example, Google Translator was used to optimise the wording of some paragraphs in this article.