Can Knowledge Be Translated (by a Machine)?

Abstract

This paper takes up the important questions whether (i) machines—by means of software—can translate texts, thereby imitating translations made by humans, and (ii) publications in the fields of scientific domains and engineering can be translated by machines to an almost perfect level. This paper examines some issues in this context, among them the question if software can be compared with the minds processed by human brains, if machines handle information or something else, if knowledge is found in software (including some deliberations on what knowledge is), and if the latest software developments are what they declare. This paper concludes that machines may assist translators but that they will not, by principle, be able to reach an almost perfect level. Accordingly the specific aim (I shall not label it objective) is to point to theoretical and analytic (parsing) obstacles to the idea that machines can produce acceptable translations in all contexts. I demonstrate that certain combinations of words will present systems with insurmountable challenges; challenges which also present human translators with questions that have almost unattainable answers. Therefore I object to the huge amount of money spent on software development for systems with that objective. I also point to the option of developing more relevant alternatives, including simpler solutions. If you ask for new insights—like you would do when perusing a research paper in order to update your professional frame of reference—there are, basically, no new insights in this theoretical scientific article. I just make it transparent what all linguists know, i.e., that in the field of machines handling language, less focus should be on what electronic systems can, or cannot, do and more focus on what details in language, and languages, will make certain kinds of handling particularly arduous; to the verge of being insurmountable.

1. Introduction

Some time ago, I received an invitation to take part in the assessment of applications for projects dealing with development of software that would be able to translate—in principle—any language into any other language and live up to the highest criteria of quality. As is well known, there is a supply of, more or less free, translation applications on the internet and a few of them are the ones mostly used by ‘ordinary people’. I know of no overview of their use or preferences in academic contexts, but my impression is (a) that also “Google Translate” is employed by academic colleagues, and (b) researchers with higher demands may use, for instance, DeepL, which has a free version. I shall have no opinion about the Google product, whereas I have seen academic manuscripts translated from German into English (whether UK or US I am not sure about) by means of DeepL. In the case last mentioned, I came across what I would prefer to call mistakes. They were not errors, in the sense that there were correct or incorrect wordings that would make them either different lexemes or polysemous words in an inappropriate context. Mistakes are sometimes a matter of ambiguity, or even style; a word is chosen but has a number of interpretations, of which one cannot be determined on the basis of the context.

2. The Issue

An example of polysemy in translation is Wittgenstein’s use of the German word Spiel—in, e.g., Sprachspiel—which can mean either ‘game‘or ‘play‘ in English. I asked Microsoft’s Bing [1] (on the application Edge) the following question (accessed on 14 August 2025):

… did Wittgenstein mean play or game by the German word Spiel?

• and got the response (boldface in the original)

A short clarification: Wittgenstein wrote in German, and the German word he uses for “game” is “Spiel.” The German word has a broader sense than the English word “game.”

There was no doubt in Bing’s mind that this was the correct answer, and why was that so? I assume that this is due to what—more or less—professionals have written in English about Wittgenstein’s language use, in due course accessible on the internet and picked up by Bing. Accordingly, it was the outcome of the academic process of filtering out irrelevant connotations, as mentioned indirectly by Bing’s note on the ‘sense’ in German. But that does not answer the question whether Wittgenstein had other things on his mind. In English, the semantics of game is mainly a reference to what people do “for amusement or fun” (my iMac’s dictionary) but also for competition, like in sports, carried out according to rules. Well, then. Not all language use is “for amusement or fun”, so is the meaning of Sprachspiel more a matter of rules for language use? And are the rules Constitutive or Regulative, a distinction suggested by the philosopher John R. Searle? If language use rules are regulative rules, then we just need to accept the rules we find in grammars and dictionaries, and that can hardly be what Wittgenstein had on his mind in Philosophische Untersuchungen [2]. It was rather constitutive rules, which are close to what children establish ad hoc and follow when they play in the playground. So, some of the implications of the word Spiel in German may be lost by choosing the word game in English. Does it matter? In the broader context of Anglo-Saxon Analytic Philosophy, it would be a little strange to expect the selection of key terms to be arbitrary, so I presume professional lectures on Wittgenstein will, with due diligence, touch upon the issue—about which word may be the better solution—taken up here.

But will a machine—i.e., a software system—be able to pay due attention to this conundrum? On the face of it, translation seems to be just substituting words in a text with other words from another language, thereby creating another, new, text, i.e., the translation. But investigating the process of translation—initially without the aid of electronic systems—appears to be a quite complicated task. My first attempts in this field of research (see Götzsche [3,4]) dealt with translation of syntax, i.e., if there could be identified a separate process in the transforming of the syntax in one language into the syntax of another language. It appeared to be fairly intricate in that one had to involve typological as well as genealogical linguistics in order to describe and explain the processes. I am aware of the fact that the electronic systems of today just disregard such difficult questions and invoke stochastic procedures in order to find so-called “patterns” (see Appendix A 1*) in texts and other media. Evidently these approaches work, which can be seen from the high amounts of machinery that use them, but there still seem to be some residuals. Can these approaches be improved until a universal level of acceptable quality is reached? And, can more or less perfect translations be seen as translations of ‘knowledge’? An answer will involve the issue of how to approach the field of linguistic semantics. As is well known, lexicographers will always experience limits to defining word meanings by other words without ending up in vicious circularity. And another issue will be: what constitutes ‘knowledge’, represented by the information offered in a text. I shall come back to this below.

3. Translation of (the Words of) a Language

The applications for the project mentioned above had to live up to the requirement that the resulting software should be able to translate scientific and engineering publications from selected languages into some other selected languages. Accordingly the criteria set a rather high bar for the accuracy of the translations; the words in the translated language should be correct and precise. In the health sciences, for instance, words may end up as being a matter of life or death.

But before we proceed, we will have to narrow down what we are talking about. The projects mentioned above that were submitted as applications for funding had, most of them, the ambition of being able to transfer (see Appendix A 2*) any kind of expression produced by humans to another kind of expression. I shall not go into the prospects of such endeavours; only confine myself to take up the ideas about translations between texts. In this context a text is defined by me as chains of words with blanks between them; and a word (see Appendix A 3*) is a string of letters. In this way I only focus on languages of which the scripts are alphabetical.

Accordingly the result of a process of translation is a text in an alphabetical language other than the language from which the words have been translated. As an example, we can pick up the Spiel instance above, of which the English counterparts may be either game or play. This seems to be straightforward: in this case Spiel is substituted by either game or play. Since Wittgenstein is, or was, a famous philosopher we must assume that both the activity of translating and the activity of reading the original and the translation is a mental activity in the minds of translators and readers. Now, imagine a language about which not many language users know the translation of Spiel: Finnish. The word is peli. Then imagine that the whole of Philosophische Untersuchungen is translated into Finnish with the title Filosofisia tutkimuksia. Most of us will recognise the first word because it is a loan word (of Greek origin) in Finnish, but not the second word. And for most of us most of the words in the Finnish translation appear only as text, in which we recognise the letters but we do not understand their combinations into words with meanings, and we may not be sure how to pronounce the expressions. For most readers in Western Europe, it is an advantage that Finland, which was for about a century under Russian rule, had adopted Latin/Roman script and not Cyrillic script, but do we acknowledge that what we have is a Finnish translation of Untersuchungen? Most of us would rely on information offered by experts, in either spoken or written statements. So, when ‘reading’ the Finnish text, the content of the text is not accessible to us; which means that specifically this text does not offer much—or any—information. Only when reading a version of Untersuchungen in a language known to us will we be able to extract the information transferred by the words. The obvious conclusion is that when presented with a piece of text that one understands, then the information is not extracted from the text. The reader’s mind carries out an interpretation of the graphics of the visual field, recognises the graphics as letters in words somehow representing speech sounds with a meaning; i.e. the text as meaning is only found in the mind of the reader. It follows that a text that is not being (and has not been) read by anyone—and of which the author is dead—virtually does not exist. You may say that such a text exists potentially, but that will be an ontological question.

One may ask how the applicants for funding of development of software systems tackle these questions. The short answer is: they don’t!

Some of the applications (see Appendix A 4*) introduced integrated systems taking care of, e.g., “the creation of precise anatomical models” combined with “complex fluid dynamics”, but since the overall aim was inter- or multilingual, a module was “a unified and advanced multilingual translation tool to convert complex scientific content”. The translation tool would utilise, e.g., “Artificial Neural Networks”, “AI-augmentation models” and “Knowledge Graph Construction”, whereas the question of the meanings of the texts handled by the systems would use, e.g., the “semantic web infrastructure” in order to create semantic tools; all of it without explicating what linguistic semantics is assumed to be, and ending up with, among other things, “context-aware machine translation”. The project would be able to open “new avenues for integrative computational modelling”, using “deep learning”, but without any technical accounts of how that should be done. The eulogy describing the infinite potential of such a project could inspire a stylistic study.

Another project would “build an assistant” that could let a researcher read “complex scientific material in the most comfortable language for individual users”. The project has the ultimate objective to make “a conversational interface powered by large language model (LLM) agents”, operating on the level of “paragraphs”, because most translation models operate on “sentence” level. It is not evident what a ‘sentence’ is: text between two full stops, or defined by syntactic criteria? One more application declares that there is “a growing need to ensure semantic fidelity” and suggests a project “focused on the design and experimental validation of a framework based on Trustworthy and eXplicable Artificial Intelligence (XAI) for the machine translation of scientific texts, semantic auditing and the detection of alterations and plagiarism in multilingual contexts, develop new hybrid models based on Large Language Models (LLMs), Neuro-Symbolic reasoning, and XAI techniques”. The ambitions give the impression of being fairly high for a work perspective of a few years’ time. But the quote also emphasises what the two most salient approaches of projects in this special business seem to be: Artificial Intelligence and Large Language Models (my underlining above). I shall come back to that below and just pick up some keywords from a number of other applications (my underlining, HG):

• “through the integration of cutting-edge AI technologies—MT, LLMs, Agentic AI, and Retrieval-Augmented Generation (RAG)”
• “a broad range of knowledge can be effectively represented using such knowledge graphs without relying on any one language”
• “Enable multilingual peer review and publication workflows”
• “Enable seamless integration of MT into peer review and publication workflows”
• “explore new techniques toward LLM fine-tuning”
• “introduces a novel, AI-driven approach to knowledge access and modelling using graph-based structures and feedback-informed translation workflows”

In none of the cases quoted here is there any (not even preliminary) technical account of how these goals are expected to be reached and, consequently, what is presented is, to put it tersely, mainly ‘old stuff’. Apart from including languages that are at best characterised as minor language communities like Estonian and other Baltic languages—an application strategy that is justified by university collaboration in order to receive funding—one application mentions Irish, a fact that seems a little more baffling. Irish is the language spoken in Ireland as the residuals of Gaelic on that island, but the amount of speakers is estimated to be between 20,000 and 40,000 (Wikipedia; different sources have different numbers) and one may question whether there is a substantial number of those speakers who use only Irish for scientific publication. And, furthermore, Irish is by Unesco classified as a “definitely endangered language”. It is hard to see why machine translation of scientific papers into and from Irish should be included, unless you see it as a technical and linguistic challenge in order to become more diligent.

Since it looks like so-called Large Language Models (LLMs) and so-called Artificial Intelligence (AI) make the point of departure for much of the prospective software development in the domain of translation of scientific and engineering publications between specific languages, I shall briefly characterise the two types of software as I see them. I am not a software engineer, but I think I understand the underlying principles of what is happening. To find out, let’s ask an AI provider:

Bing//Copilot_09/09-2025

query: what is AI? answer:

Artificial Intelligence (AI) refers to the development of computer systems capable of performing tasks that typically require human intelligence, such as reasoning, learning, decision-making, and perception. AI encompasses a wide range of technologies, including machine learning, deep learning, and natural language processing (NLP).

Taken literally, this is a little confusing. By “computer systems” is plausibly meant “low voltage” electricity currents, also called software, run by a machine called a computer, but the assumption that “reasoning”, “learning” and “decision-making” are a matter of human intelligence is quite puzzling. My wife’s dog—mentioned above, answering to the name Shirlock—is able to learn, make decisions (e.g., if to eat the food or not), so, as a trope, we say that she is very intelligent; but she is not human (see Appendix A 5*). And another conundrum is what meaning the word intelligence has in this context. As is known by some academics, the theoretical term “intelligence” “has a bizarre origin” (citation from Götzsche [5] p. 583 referring to Richardson [6]) and might as well be understood as cognitive capacities leading to the “mondo cane” capabilities just mentioned. But the good question is whether a software system does the same thing as the human mind with its mental processes by its cognitive systems run by the brain. This is what is claimed by another question to an AI provider:

Bing//Copilot_09/09-2025

query: how does AI work? answer:

… AI emulates human cognitive abilities, …

Professionally, I totally agree that AI actually emulates something; but it is not human cognition. After having read Caltech’s Beginner’s Guide to AI [7], I realise that AI applies anthropocentric projections the same way we as humans do with other things. When my wife and I talk about her dog, we perceive her as a kind of (small, she is a Cairn Terrier) human being, and AI protagonists often talk about machines ‘understanding language’ and ‘think and act like humans’. Thus, a key expression in this context is that “ML (Machine Learning) uses networks like the human’s brain” as a requirement for doing this. This is understandable as a kind of simplification for the uninitiated; but basically, the reasoning is flawed (and there is some scepticism “out there” as to the use of such language; see Appendix A 6*). We have, as of today, no justified theory about how the neural networks of the brain are creating what some researchers call a “model of the world” in the minds of humans, so how could artificial networks be expected to do the job by that analogy?

In my view, text-handling AI is a label denoting software that is able to mimic text in the form of strings of letters from an alphabet to the extent that humans accept it as a quasi-human product. I suppose that the procedures are well known: algorithms that search the internet for examples and patterns of letter strings that somehow fit the queries and then will be busy calculating possible and plausible combinations of such strings. But it should be emphasised that nowhere in the software process is any information involved. As mentioned above, information is not construed until a human has interpreted strings of letters as such information. If the piece of information is relevant for each query, it is, in the end, a matter of the existence of a pragmatic context. And I shall not go into the issue of whether the energy consumption—which is, to my knowledge, enormous—is worth the trouble. In that context it is a peculiar fact that the “electronics of the brain” runs with the speed of max 120 m/s (432 km/h), whereas the electronics of computers runs at the speed of light and is, anyhow, not able to make perfect imitations of human brains that consume a fraction of that energy. Maybe a software system is able to win over humans in all quantitative games but is not able to “understand” anything.

Then, just a short comment on so-called “Large Language Models”. One should observe that these “Models”—including their technicalities like “vectors” as “numerical entities in a multi-dimensional space”—are not language models, in the sense that a ‘model’ is a theoretical description and explanation of some empirical facts. We know that languages exist, and we know that there are speech sounds and, sometimes, writing systems like alphabetical script. And we have fairly good insights into the descriptive grammars of many languages but there is no general agreement about which kind of theory of syntax and morphology is the most reasonable one; there are actually hundreds of theories on the market. Then there is the question of semantics. We know, for sure, that language meaning works but we know very little about how the mind, as processed by the brain, does the trick. When my wife tells me that ‘Shirlock (her dog, see above) is sleeping on the sofa’, we both know what we mean but we do not know how our brains know it. If you do not believe me then read Cobb [8].

And when we talk about meaning we may include the domain of ‘reasoning’. One of the issues with answers to AI queries is the potential lack of consistency. In a paper by Fengxiang Cheng et al. [9]: ‘Empowering LLMs with Logical Reasoning: A Comprehensive Survey’, (posted on Academia.edu 2025), the authors say that ‘there are still significant challenges to the logical reasoning abilities of LLMs’. They propose a number of methods amending the situation, among them translating natural language into symbolic language and back again. The idea illustrates the problem very well. The authors are themselves aware of some of the obstacles but the underlying assumption is that there are a few symbolic languages, such as ‘logic programming (LP), first-order logic (FOL), constraint satisfaction (CSP), or boolean satisfiability (SAT) formulation’ (Section 2.1), while there is nowadays—after classical logic was symbolised in the 19th c.—actually a plethora of, what I would prefer to call, coding languages including mathematical tools that may be candidates for such a kind of back-and-forth exercise. And which solutions in “translating” (which may not be the best word) natural expressions into some set of symbolic expressions will be the best, or just be able to do the job?

4. Translation of Knowledge in the Words of a Language

The question that was the point of departure for this paper was the, sort of, competition between applicants for funding of software that would be able to translate scientific and engineering publications between a number of languages. This raises the issue to what extent one can say that knowledge can be translated; and first, it is a matter of what knowledge is. This question has had philosophical minds occupied for millennia so, as a preliminary and rudimentary approach, I shall propose the following.

When we say that we know something, we would like to be sure that this knowledge is certain. In the Old Greek context, this was expressed in propositions like (in the format of today) ‘Shirlock is sleeping on the sofa‘, and propositions were said to be either true or false. Over the centuries, it ended up as the standard definition in textbooks on epistemology: ‘knowledge is justified true belief’. Apart from the tricky question about what “beliefs” are, this seems fair enough, and formal (later symbolic) truth-functional logic was born. I would like to suggest some modifications to this approach. First, I would say that if I know something for certain then it must be a fact, indeed, taken to the extreme it must be an empirical fact; at least if we are talking about science and engineering. Stanford Encyclopedia of Philosophy (SEP)_ [10] has a number of entries each of which includes the word knowledge, or words affiliated with the ‘notion’ (see below), and the main SEP article is ‘Knowledge how’, and it is a traditional exercise in Analytic Philosophy. This has the effect that ordinary vernacular words are made subjects of meticulous scrutiny (over 49 pages excluding references), which, to me, is not always illuminating. Instead, I shall propose a distinction between universal, particular and singular empirical facts. This seems to be in accordance with Kant’s types of Categorial judgements on Quantity (see Journal of Logics and their Applications, Volume 4, Issue 4 [11] p. 845)

A universal fact is what there among scientists is a consensus on. We are pretty sure that gravity works—even though we do not know what gravity is—and likewise we are sure that the Moon is moving in a trajectory circling around the Earth; and if it will in the end “fall down on Earth” or fly out into deep space belongs to a somewhat distant future. A singular fact is what is also called a contingent fact. If I say to my wife, ‘It’s raining now!’, that implies that we will not take a walk right now, then that may change, or I may be wrong, or just joking. A particular fact is something in between. ‘Most people have a mother tongue’ but some people suffer from aphasia. But ‘Do children whose mothers took paracetamol during pregnancy risk autism,—currently (September 2025) disseminated on the internet—is not a statement supported by experts, so it is not a fact; it is put forward as a hypothesis, and in this case one may question if it is really a scientific hypothesis. Meanwhile, ‘we were in England for our holiday’ is a fact depending on the meaning of we and if we really were there. So, one could say, informally, that a fact is ‘a state of affairs that no sane human will deny’, maybe supplemented with the phrase ‘based on available information’, and also that a fact will be hard to deny without contradiction: ‘gravity is working but I think I can fly on my own’. Whatever you would consider a fact, it is for humans either what one has experienced, or one has heard or read. It is generally agreed that scientific endeavour aims at confirming facts and suggesting theories about facts, often in the form of hypotheses. Theories are always preliminary, sort of ‘to the best of our knowledge’, see below, and should there be consensus on a specific theory it is most likely wrong. This reality with its degrees of uncertainty is part of the knowledge-labour done by scientists, and also part of their publication activity

4.1. Subsection: On Knowledge and Science

This year, exactly 30 years ago, I was assigned to a workplace in the room of a, by me highly estimated, professor at the Danish university I retired from in 2019, because, for logistic reasons we had to share an office. Often, we had some interesting debates on professional matters but one day my kind of soulmate looked at me and said, with a touch of horror betrayed by his voice, ‘But Hans, are you an empiricist?’ I admitted that in Theoretical Linguistics I was in favour of empirical evidence and mostly adhered to the English language empiricists Lock, Berkeley and Hume. But this is now history, and it is a story, an anecdote, that I have told a number of times with the consent of my colleague in order to illustrate the fact that one can have different world views over one’s lifetime but keep one’s empirical ambitions. As of today, I mostly identify as a philosopher of science in favour of Kantian Transcendental Idealism (TI). What the details of this are have been elaborated on by a (not small) number of intellectuals since Kant presented his thoughts in the 18th c., and the following is my own interpretation of some of his main thoughts. And I shall not, in this context, have recourse to the German nomenclature; I will save that for another context.

According to TI, humans know something either by means of reason, so-called a priori knowledge expressed in analytic statements, or by means of experience, so-called a posteriori knowledge expressed in synthetic statements. What we perceive in the outer world is only phenomena and we have, per se, no knowledge of the essence of objects. In order to have such knowledge, we have to transcend the phenomena and mentally enter the “world” of ideas, hence idealism. This has nothing to do with Platonic Idealism, according to which our perceived world is not the real world; which is instead the world of “real ideas” of which what we see are only instances of these ideas. In a modern wording, one might say of Kantian Idealism that what is sort of ‘behind‘ the phenomena, the essence, has to be ‘construed by theories‘. Theories are expected to describe and explain phenomena, and scientific theories are—pace Karl Popper in case they have not (yet) been proven to be wrong—to the best of our knowledge regarded as (almost) certain knowledge. This is what modern civil society pays scientists to take care of, and also corporate research in, e.g., laboratories has to live up to these criteria. And, of course, not only the proviso by Popper applies; things work, as just outlined, not in an ideal world, and we all know of, shall we call them, complications in research. I shall therefore recommend that research is not confused with science. In my view, not all research is science and not all science is based on so-called research. In the modern landscape of higher education, quite a few more or less academic subjects are labelled “science”; e.g., in German Litteraturwissenschaft is by definition a science, and so is Filmwissenschaft, in German defined as a Kunst- und Kulturwissenschaft. I am not the one to set up conceptual borderlines in academic life but if these German academic subjects are sciences then Richard Feynman’s well-known and often-quoted wording “the scientific method” (see the Caltech Lectures [12]) does not apply.

When selecting the Kantian approach to knowledge and science, I deliberately choose the basis offered by Robert Hanna [13] in his published volumes on Kantianism, primarily his Kant, Science, and Human Nature (2006). The words knowledge and science are found on more than 50 pages and each specific context contributes to the clarification of this stance in the philosophy of science. But I shall not further develop its relevance to the theme of this paper.

Assuming that a piece of research lives up to the criteria of representing science in an acknowledged scientific field—say Theoretical Linguistics, Mathematics or Theoretical Physics—and the paper in which the results are presented has been approved by colleagues, then a question of the translation of the paper involves three obstacles and phases: (i) what is the language of the original paper? (ii) what is the mechanism of translation: human or electronic? and, (iii) what language is the target language? In order to illustrate the resulting dilemmas, I shall, below, select text from some scientific papers, mainly paragraphs, and have DeepL translate the texts between some languages.

But before that I will take up some details in a specific language. On 13 September 2025 the Danish newspaper Berlingske mentioned the following (headline)

I Mette Frederiksens bagland drømmer man om noget helt andet end formanden [14]

‘In Mette Frederiksen’s camp, they dream of something completely different than the chairperson.’ DeepL, free version

The syntax in this piece of text displays the “unfinished sentence dilemma issue” (in other grammatical approaches (take Chomskyan Grammar, you choose) it may be called other things). The phrase noget [helt] andet end ‘something […] different from‘ (I would say that from is the more appropriate here) in this context means that two things are being compared and of them the first one has, to somebody, the highest priority. The word chairperson may be seen as a cultural gender reflection, or it may be seen as a translator’s ignorance of first name genders in Danish, and the word camp looks like an almost desperate try. In Danish, bagland is an idiomatic expression referring to ‘respected members of the Danish prime minister’s political party’. But the main thing here is the comparison in which the bagland either is (a) dreaming about other things than about their Prime Minister or (b) is dreaming about other things than their Prime Minister is dreaming about (end formanden [drømmer om]). The interpretation last mentioned is presumably what is spontaneously chosen by Danes because of their knowledge of the political pragmatic context in Denmark. So, how could DeepL know? Anyhow, in a number of relevant languages we get the following:

I Mette Frederiksens läger drömmer man om något helt annat än ordförandeposten. (Swedish)

• Here we have bagland as läger ‘camp’, which is like in English, but in Swedish they dream about (taken for granted: ‘who should be elected to become … ’) ‘party chair’.

In Mette Frederiksen’s achterban droomt men van iets heel anders dan de voorzitter. (Dutch)

• In Dutch it looks as if one gets the same meaning as in the Danish (b) version, viz. social democrats dream about something else than what the party chair dreams about.

In Mette Frederiksens Hinterland träumt man von etwas ganz anderem als der Vorsitzende. (German)

• In German the combination of etwas and als der Vorsitzende actually yields an (a) translation, based on the fact that the implicit als [von] takes die Vorsitzende in the dative case: der Vorsitzende.

Nel campo di Mette Frederiksen si sogna qualcosa di completamente diverso rispetto al presidente. (Italian)

• To the best of my knowledge, the Italian translation has the same dual meaning problem as Danish, whereas French,

Dans l’entourage de Mette Frederiksen, on rêve d’autre chose que la présidente.

• actually gets it right: (b). And you make take my word for it but in Finnish it comes out like in Swedish:

Mette Frederiksenin tukijoukot haaveilevat jotain aivan muuta kuin puheenjohtaja.

• I could have kept on with high numbers of reversed translations and have found exciting details but the above seems to me to cast some light on what hurdles human and electronic translators encounter. And we can try another, sort of classical, dilemma (DeepL):

Det var en meget bevæget hustru til den dræbte højrefløjsaktivist og Trump-støtte Charlie Kirk, der i nat dansk tid for første gang siden drabet på sin mand udtalte sig til offentligheden.

‘It was a very emotional wife of the murdered right-wing activist and Trump supporter Charlie Kirk who, last night Danish time, spoke to the public for the first time since her husband’s murder.’

DR news app 13/09-2025@08:55

The syntax/semantics interplay in this piece of text displays what a Danish grammarian might label ‘the Danish anaphoric pronoun dilemma’, in that the phrase in Danish siden drabet på sin mand is expected to correspond to the English phrase since her husband’s murder. The English version has chosen an ‘s-genitive’, a solution which, to the best of my knowledge, is a standard solution in both English and American. It is, by itself, a little peculiar because it presumes a kind of possessive relation between a human being and the murder of that human being. But it can, of course, be seen as an idiomatic expression. I have tested the phrase in a minor number of languages, and they all illustrate a special grammatical and semantic phenomenon in Danish, viz. the above mentioned ‘Danish anaphoric (actually more correctly ‘anaphora’) pronoun dilemma’. We will take a look at the syntax of the last sentence:

$$

siden drabet på sin mand udtalte sig til offentligheden

adverb phrase--------------- verb-------- prep phrase

‘spoke to the public for the first time since her husband’s murder’

 verb prep phrase prep phrase------ prep phrase--------------------

$$

• The interesting thing here is the adverb phrase in Danish:

$$

siden drabet på sin mand

adverb phrase-------------

$$

• in that it contains a prepositional phrase:

$$

på sin mand

prep + pronoun + noun

$$

The pronoun sin goes back to Old Norse—and is also found in Old English—and it refers back to one or more words in front of sin. Danish grammar tells us that it refers back to, mainly, a subject in the sentence syntax, and the only subject found in the previous sentence (actually by way of der ‘who’) is hustru ‘wife’, and one may therefore conclude that the Danish sentences in their combination of words mean that hustruen havde dræbt sin mand ‘the wife had killed her own husband’. Not many Danes would (because of the pragmatic context) notice the implied semantics but there are people who spend their time finding “errors” made by other writers. Twenty years ago, I had an email exchange with a Danish politician who, on TV, claimed that sin/sit ‘his/her’ is the only correct use (meaning that there is no Danish anaphoric pronoun dilemma) of that anaphoric pronoun in Danish in, e.g.,

$$

han tog sin cykel og kørte hjem

‘he took his (own) bike and went home’

• while

han tog hans cykel og kørte hjem

‘he took his bike and went home’

$$

is incorrect because hans may refer to somebody else than the possessor of the bike. I pointed to the fact that hans is in general use in Danish dialects (analogous to English), but he only explained to me that I was mistaken. The dilemma mentioned above is that if you use sin, you apply the anaphoric semantics, whereas if you use hans, you apply the gender difference hans ‘his’ vs. hendes ‘her’; but you cannot have both in one word in a Danish phrase. My point is here that electronic translations will have, at least, some trouble handling this phenomenon (all DeepL):

… since the murder of her husband, she has spoken publicly (British English)

… seit dem Mord an ihrem Mann sich öffentlich äußerte (German)

… sedan mordet på hennes man har hon uttalat sig offentligt (Swedish)

… depuis le meurtre de son mari, s’est exprimée publiquement (French)

… dal momento dell’omicidio del marito ha rilasciato dichiarazioni pubbliche

        (Italian)

… murhan jälkeen puhui julkisesti (Finnish)

        ‘murder after spoke publicly’

The suggestions by DeepL have to be given the credit that there are alternative suggestions to choose between; but one has to have some knowledge of each language for the alternatives to offer some meaning. By all means then, the options above show no dilemma like the Danish grammar/semantics interface but a meticulous examination of the languages applied here will—I feel confident about it—uncover other dilemmas in these languages. And such dilemmas are not easily solved, not even by invoking ‘translations’ into “formal languages” and back again the way it is suggested by Fengxiang Cheng et al. above. The good question is, of course, “does it matter?” Since Danish readers may not notice that siden drabet på sin mand is a peculiar expression, we can all live happily with it; except for in far-from-perfect translations. My concern will point to the effect on the national language of Danish. Any language does change all the time through language use, but I have in general warned against inconvenient lexical and idiomatic changes like, for instance, the ubiquitous spread of the word dele ‘share’. It has, to my knowledge, been taken up from American share about information in some software format, especially on so-called social media. In Danish, we have a number of nice expressions meaning ‘tell’, fortælle, berette, etc., but dele is now often used also in spoken exchanges and it may end up limiting the vocabulary of Danish. And if electronic translations are taken at face value so that Danes think their suggestions are “correct” Danish, then it looks like we have a problem.

End of subsection

Back to ‘electronic translation of scientific and engineering documents’, or, ‘can knowledge be translated (by a machine)?’ You could say that the textual cases just analysed are trivial and that in science documents, the language is much more precise and does not present you with tricky expressions. That seems to be just the point. From the account presented above the reader may have got the impression that I am against or an opponent to, what I prefer to call, electronic translation. That is not the case. What I am sceptical about is (i) the balance between the use of machines and electrical power to achieve something relatively simple for professional translators and (ii) the question, why all the fancy labels—also in professional contexts—attached to the handling and the output of software processes? As for (i), it is a matter for corporate and political projects; and one of my experiences with research and science is that politicians should not be involved in decision-making concerning technical or technological issues. It has gone, and will always go, wrong. Regarding (ii), one may rephrase it and ask ‘why the mystification’ hidden by the fancy terminology? Machines and software output may mimic human acts and products, e.g., presenting alphabetic tokens on my computer monitor as effected by my pressing the relevant keys on my keyboard, or coming from other sources. But, by all means, the processes in my software systems (apps) do not imitate or emulate the mental or cognitive processes in the neural systems of me or other humans. As argued above, there is no information in the machines or in the software. It only becomes information when it has been read, or otherwise perceived by a human. All other functions are just mechanical and can make machines, like cars, adjust their functions, which points to the limitations of machines and software: it is maybe artificial but not intelligent in any of the meanings of the word. And so, with knowledge. Knowledge is not being translated.

After Gutenberg and the printing press, technology has developed in a way previously never seen before in human history; but it is a mistake to call it “Information Technology”. What has been developed is the medium, or mediums, for replicating written materials, or other more or less symbolic graphs, that can be perceived and processed by humans as information and knowledge: in runic inscriptions, it was stones or wood or bones; later, paper was used; and now we watch pixellated graphics on monitors. In this context, I would like to point to two issues here. One is the way knowledge is gained by individuals, i.e., by experience or by having been told or by having read about something, see above on Kant and types of knowledge. In order to understand explanations of objects or events you need theories that are able to, let us call it, explicate those objects or events. Take reading, for instance, as an ironic digression in this context. It is one of those research fields that needs explication. On ResearchGate on the 4th Sept. 2025, a question was asked: ‘What is the terminology associated with the reading wars?’ [15] In spite of my insights into reading research, I did not answer but I read one of the answers. It mentioned seven disparate approaches:

Whole language

Phonics

Balanced literacy

Science of reading

Structured literacy

Phonemic awareness

Psycholinguistic guessing game

For the uninitiated, it can be mentioned that Phonics and Phonemic awareness are methodologies stressing the connection between speech sounds and script; while what lies behind the other labels, seems to be more opaque. The different perspectives are so heterogeneous that the only conclusion about the research field of reading is that ‘we do not know how it (reading) works’; or bluntly: ‘we haven’t the slightest idea’. So, when knowledge is presented in text we have learnt to interpret the expressions as information, which in due course can build knowledge in our cognitive systems. And in spite of the fact that we do not know how it works we can be sure that it does work. What has made the output in some form of text, in English, Finnish, Russian or Japanese, is not important, as long as we as readers accept what we perceive as the knowledgeable content and are able to integrate it into our general understanding of the world as an effect of our individual accumulated experience; and that, therefore, it can be applied in our pragmatic context (see above). So, ‘can knowledge be translated (by a machine)?’ So far, yes, with the caveat that follows from the narrow definition of “translate” presumed in the account above. Can that be demonstrated?

4.2. Subsection: Scientific Papers for Translation

To answer that question, I have tested a few “translations” into a few languages of my choosing using DeepL (see Appendix A n*):

David L. Share: ‘Phonological recoding and self-teaching: sine qua non of reading acquisition’, in Cognition 55 (1995) pp. 151–218. [16]

(English into German)

C. S. Unnikrishnan (2025): ‘Information versus physicality: on the nature of the wavefunctions of quantum mechanics’, in Academia Quantum, Research Article, Published: 7 May 2025. [17]

https://doi.org/10.20935/AcadQuant7696

(English into German and French)

Akshansh Mishra (2025): ‘Machine learning-driven optimization of TPMS architected materials using simulated annealing’, in Machine Learning for Computational Science and Engineering 1:1. Springer. (pp. 1–20) [18]

https://doi.org/10.1007/s44379-024-00001-z

(English into French and Swedish)

The outcome is, surprisingly if you are a sceptic, that the target language texts are, in general, well formed and accurate. The general reason seems to be that the English wordings of the scientific papers reflect accurately the line of reasoning behind the piece of research carried out, an assumption that underpins the perception of the text as an unambiguous description of that contribution to science. Actually, it undermines the idea presented by Cheng et al. [9] (see above) that translations between natural language and some kind of formal language (code) might improve “the logical reasoning abilities of LLMs”. If scientific articles are actually written in accordance with the tradition(s) in a particular scientific domain then such a “back-and-forth” logic approach will be redundant.

The other issue is how machines and humans are able to do the job at all. And frequencies are at stake here. Machines apply statistical methodologies to strings of letters, or other kinds of expressions, possibly supplemented with some rule-based procedures. Based on specific research papers, I shall suggest that the human brain also works on frequencies, even though the “statistics” is not conscious. As mentioned in Götzsche [5], the brain likely works as a prediction machine but not quite as a calculator of probabilities like a machine. The brain has a “model of the world”, what I would prefer to call a Mental Universe (MU), but it is not updated all the time. It would be too energy consuming. It only updates changes in relation to what is predicted. This works in all our kinds of behaviour, including reading. It follows that knowledge working as part of this “model of the world” (HG: MU) is only updated to the extent that the content is not predicted by the ‘MU-processes’. Compared with machine software, the brain works the other way round. Instead of continuously searching the internet for appropriate data, like software does, the brain is basically introvert. It is, mechanically, seeking confirmation of a, so to speak, ‘steady state universe’ until the point where it has to adjust to adapt to new circumstances, whereas, to my knowledge, machines working with AI will all the time be trawling the internet in search of updates. In my view, this leads to the opinion that machines will not—at least not this time—lead to job redundancies in the form of takeovers by machines. And new ways of doing things will inevitably lead to the invention of new jobs—among them the exchange of knowledge—and we have no need for luddites.

We are now back at the initial question, which is the title of this paper: Can Knowledge be Translated (by a Machine)? The question—in its complementary meaning (as its negation as a question and instead stated as a proposition)—has to be analysed in its parts. I have entertained the theoretical concept, the notion, of ‘knowledge’ above and I think that we can ascertain two hallmarks of the idea of ‘knowledge’: (i) ‘knowledge’ as a concept is very hard to define in an acceptable way, even though one has the best of intentions, and (ii) different contexts may assume different meanings of the word. A search on Stanford Encyclopedia of Philosophy [10] offers 1921 entries with the word knowledge, or affiliated words (like wisdom and epistemology), as part of the title of the entry (accessed on the 22 September 2025).

In accordance with my deliberations above, it is also problematic to claim—as is the general understanding—that knowledge is a property of a text, whatever medium is used, as a kind of hidden substance that has to be uncovered. I maintain that a text, in the understanding of this scientific paper, is a number of strings of letters which has to be read for it to activate information and knowledge in the mind, i.e., the cognitive systems of the brain of the reader. Beyond that, some of the information and knowledge may be new or alternative to the reader and may be merged with the reader’s mental universe. Very few readers are able to remember all the words of a text, and not many are able to remember most of the words of a scientific paper. But specialists in a scientific field (I am not happy with the label “expert” and prefer the category ‘specialist’) will in general extract—maybe by abstraction—the main terms and lines of reasoning in a paper and let them be blended with their general frame of reference in the relevant field. A presupposition is, then, that a paper has been written by some specialist and that the words as expressions, phrases and syntax have a semantics that is the expression of the knowledge of the author. To what extent the knowledge in the minds of the author and the reader respectively correspond to each other is a difficult question and it shall not be taken up here. Uncertainties and misunderstandings happen all the time but one must assume that most scientists read papers in order to, can we call it (?), ‘learn’ something, be it writing an assessment as a peer reviewer or for extending one’s theoretical spectrum. If we adopt the scenario in which the reader would like to read a paper in his own language we will have to introduce an extra reader, the translator, who may well be an author with proficiency in more than one language.

As long as humans have been able to use a (kind of) language to exchange thoughts, and accordingly knowledge, between each other there seems to have been different languages, and the art of translation goes way back into the past, documented by the number of translated texts, specifically texts that are by evidence known to have been translated. A case in point may be biblical translations in the Christian tradition. The etymology of the word translation is well known and includes the information that its origin is a past participle of Latin transferre (cf. the word transfer in English), meaning that something is ‘carried over’ or ‘conveyed’, and specifically that something could well be said to be knowledge. For the process of conveyance to take place, the extra reader must have proficiency in both the language of the original text and in the language of the text that is expected to be the outcome of the translational process. The translator therefore becomes a reader and a writer at the same time, but we must emphasise that, in case we assume that knowledge is transferred, it must be ‘situated’ in the mind of the translator, also acknowledging the fact that there may be gaps or addenda found in the translation if the texts are examined in detail. But the potentially slightly different sets of knowledge must be found in the mental universe of the translator, appropriately in the form of what I call (see above) a pragmatic context. It follows that the translator understands the content of both the original and the translation respectively.

If we would like to give consideration to the idea that a machine in the form of a software system should do the same thing, viz. carry out the process of translation of paper TXT^orign into another language so that it becomes paper TXT^trl, we must look into what we know of the workings of such a system. Programmers and linguists may remember Systran, which was one of the systems in the second half of the 20th c. It was a rule-based system, and I remember colleagues telling stories about it, viz. that when it was to be demonstrated at a special event the Systran people had to ask IBM for help because it did not work. Later it implemented a rule-based and statistical machine translation system, and this later approach is also the solution in translation software of today, in that it works by doing the following (see [19] p. 126):

… a deep neural net is constructed by stacking layers of nodes connected via regression models. The model as a whole takes the explanatory variables X as input, processes them successively through the middle layers, and finally yields outputs in the response variable Y. This means that a deep neural network is itself a gigantic regression model. Indeed, the whole model can be written down in a compact form as

y = g(x;w)

Taking our text TXT^orign as input, the output will be the text TXT^trl, and none of the processing will be done by means of anything beyond software, i.e. ‘low voltage’ electric currents. As is well known, the behaviour of the software is run by a machine (programming) language that is numerical. As mentioned above it is not likely that the human brain works in the same way; but of course, we do not know for sure. So, I think that to claim that the machine with its software is handling information, or even knowledge, would be a little controversial

End of subsection

So, as for machine translations, one can say that we have the same problem in the form of a gap between an original text—presumably written by a human—and a Machine Translation untouched by humans. One can, furthermore, say that if the knowledge of the writer of the original includes a kind of “manual” and a “toolbox” to “encode” (see Appendix A 7*) the text, then one must also assume that any reader, and also a translator, will have both a “manual” and a “toolbox” for “decoding” the text for it to become knowledge in their minds. Assuming that both the original and the “translation” are found on a computer screen—or maybe have been printed—the knowledge associated with the graphics is, accordingly, only found in the minds of the writer and the readers respectively, and if an output “translation” is sufficiently appropriate based on software that trawls the internet and then calculates possibilities and plausibilities, it can be seen as a combination of craftsmanship and luck. Craftsmanship is the form of applying regularities (what some people call “patterns”), but about luck we cannot really be sure. Since knowledge—or information for that matter—is not an intrinsic attribute of electronics, I shall propose to adapt an alternative label to so-called “machine translation”: it is actually Artificial Translation (AT), i.e., “artificial” in the sense of ‘not genuine’ (etymologically from the 1640s) and implying ‘not to be trusted’. To wrap up, Can Knowledge be Translated (by a Machine)? No, machines cannot translate anything, neither texts nor knowledge. They can, of course, mimic, imitate, emulate, simulate, resemble, and (in other ways) look like translations, you choose; and available tools of that kind may sometimes be very useful and supportive because human minds are not very good at all details and exact words or wordings. But you should not confuse that with intelligence or other peculiar or mystifying things attributed to humans, or with pure magic.

By coincidence, when I was working on this paper in September 2025, the MIT Press mailing list sent me a poster on a new book, What Is Intelligence? The blurb indicates that the brain is seen as a computer. Since the first analogies between man and machine—i.e., organisms seen as machines—appeared in the 16th c., this suggestion seems to be the latest offspring of that idea. I downloaded an excerpt from the book, in the format of The MIT Press Reader, and the resemblance between brain and computer is here taken a step further. The subtitle says the following:

Alan Turing and John von Neumann saw it early: the logic of life and the logic of code may be one and the same [20].

• It is not quite transparent what “logic” means but there may be a broad conception of both “life” and “code”. Next the author refers to a statement by John von Neumann, who predicted,

… that life, at its core, might be computational …

• and refers to an event in 1994 when a “pixellated machine” has “read a string of instructions, copied them and, built a clone of itself”. An analogy is then proposed:

self-replicating systems read and execute instructions much like DNA does.

The key question is here what fills the theoretical function of being a stepstone, a link or a liaison, between instructions and executions producing, in the end, e.g., a clone? In the context it is not quite obvious. It looks as if it is assumed that DNA instructions are of the same kind as instructions in a software program, but the analogy cannot easily be justified simply because we do not know in detail how DNA instructions build the organ of a functional brain, and in the end human cognitive functions. The ideas put forward here are at best committing the fallacy of analogy and seem close to being a version of what may be called a ‘blueprint theory’ (see below).

There is, of course, the theoretical option that you may say that the one-dimensional processes of the electronic calculations (also?) represent information about the semantics (and pragmatics?) of the texts involved in a translation procedure. I would like to call this the ‘blueprint theory’ of translation, and the term and the approach, is inspired by the biologist Richard Dawkins in his The Blind Watchmaker [21]. (On p. 305 he mentions the Lamarckian ideas about biological evolution and attaches the label ‘blueprint theory’ to traditional and modern Lamarckism. I shall propose to sum up the approach as the idea that when a two-dimensional blueprint is a scaled down (for instance) graphic version of the real thing—a house for instance—the miniature version is a description of the real house. But if you assume that the DNA string is a miniature description of the body of a living organism, you will commit the fallacy of division, viz. believing that what is true for the whole is also true for the parts. The DNA string is not a minor organism but instructions for ‘building’ an organism ([21] p. 306), and only in the form of such instructions being replicated in the cells of that organism. The mechanism of encoding a blueprint into a one-dimensional signal is relatively easy, like doing it with the two-dimensional entities of an alphabetical text. But the instructions to sort of encode meanings into a text, which is a string of words that are situated in the cognitive systems of the human brain, is not a straightforward piece of work. So, it looks as if these ideas do not grant the solution to the vexed question about what machines do when an input is transferred to an output, called a translation.

One of my philosophy teachers in the 1970s said on one occasion that there are some questions that cannot be answered before ‘the light of consciousness would be lit in a computer’. Maybe he implied that the event of this happening was not just around the corner; and so it seems today.

5. Conclusions

What I am sceptical about concerning the project applications, referred to as a point of departure for this paper, is then, to what extent are such mega projects relevant, or worth the money? If DeepL is so good at transferring the scientific knowledge as has been pointed to above, why should public, and corporate, money be spent on extending existing software building mega-systems, ever bigger and more power consuming? What I would recommend is an incentive to make available systems simpler, cheaper and less power hungry. But apparently this is not where the money is?

In the account above, I have presented an issue that has to do with the current interest in expanding the software business to still new domains of human life, and I shall sum up my observations in a few statements:

• I am not opposed to the development and use of so-called AI, nor to the development of so-called LLMs.
• So, I shall not try to contribute to undermining the use of AI by ordinary people (not that I am sufficiently megalomaniacal to think that I will be able to influence anybody).
• Software in the form of electronic machines is a very useful and practical tool (even though you may consider the energy consumption/output demand balance mentioned above).
• But I am sceptical of the idea that machines can be compared to human brains the way they are in the media and also in research papers, when they talk about “thinking machines”. The similarities are not there, even though the output may resemble human output and despite the fact that software engineers do not know what happens in certain (so-called “deep”) processes.
• And I am worried about the disruption (Americans may call it havoc) that will be the consequence if influential groups begin to believe everything that an AI system offers as an answer to controversial questions.
• In the more limited range, I might be worried if answers to certain questions are being perceived as facts and if theories and explanations are spread without there being any basis for them.

When humans first began to become self-conscious, most likely by means of language in its incubate forms, they did, I think, create myths in order to answer the question still around, ‘who are we and why are we here?’ I asked Bing on the 23 September @16:52, ‘what are humans and why are we here?’, and the answer was (boldface in the original)

human being (Homo sapiens) Human beings (Homo sapiens) are anatomically similar and related to the great apes but are distinguished by a more highly developed brain and a resultant capacity for articulate speech and abstract reasoning

supplemented with the usual list of links and commercials. ‘Why’ was not answered. Maybe Bing does not know.

And, well, I almost forgot about the project assessments above that I was a part of. Yes, there was no doubt that funding the best projects would boost research in the field and trigger the development of great new products underpinning science and disseminating scientific knowledge for the welfare of humans. But I must admit that I did not consent to all the opinions about ‘high quality’, which are, albeit just about technicalities. The projects will have more severe challenges, data for instance! There seem to be no reliable estimates over what, and how many, languages are available on the internet—for instance in the form of websites. But in relation to the between 6000 and 7000 languages in the world, apparently only a few hundred are found on websites. I tried to search the ‘Gujarati’ language (Indian) in the original Gujarati script on OpenAI (https://openai.com/) but I could not make it accept the script. The language has about 60 million speakers, which is a little more than the 20,000 to 40,000 speakers of Irish. Maybe we are just in the beginning of the era of translation automatons and automaton interpreters, but the overall impression of the whole idea of having programmers work hard on making systems be able to produce “perfect” translations of papers in science and engineering is more of the following type: this is a solution looking for a problem.

5.1. Postscript

The title of this journal is Knowledge. And the ambition of this article is to point to the predicaments that may arise if translations of what one regards as knowledge are seen as just substituting some words written in a language with words picked up from another language; i.e., the way it is done by software running on computers. It is not only a matter of linguistic context—or social context for that matter—or the semantics and pragmatics of the expressions in the languages used; it is also about what counts as human knowledge. In the philosophical and scientific traditions respectively (of which western philosophy goes back about 2500 years, whereas science is a relatively new idea), this question is dealt with in epistemology and ontology as an ongoing process, while scientific ‘paradigms’ replace each other, sometimes rather frequently. As is well known, there have been, should we call them (?), ‘problems’ with data and study methods in research articles published in acknowledged publications, and such difficulties may have to do with the overall worldview of the scientists.

In his 2024 Folio Society publication: Death by Black Hole [22], the author Neil de Grasse Tyson points to some hyperbole statements in recent history. Albert A. Michelson said in 1894 that “The more important fundamental laws of physical science have all been discovered, … “ ([22] p. xxi), and the astronomer Simon Newcomb stated in 1888 that “We are probably nearing the limit of all we can know about astronomy” (ibid.). As we all know, a few things have happened since then, among them Max Planck and ‘quantum mechanics‘ (p. xxii). This leads Tyson to note that ‘What I do know is that our species is dumber than we normally admit to ourselves. This limit of our mental faculties, and not necessarily of science itself, ensures to me that we have only just begun to figure out the universe.’ Beginning with a number of insights coming up at the end of the Middle Ages, some of the solid forms of human “knowledge” that had prevailed for millennia, presuming a given dualism between heaven and earth, were undermined by Newton and his ‘universal law of gravity’ ([22] p. 11), i.e., physical laws apply in the whole (known) universe. And ‘This universality of physical laws drives scientific discovery like nothing else’ (ibid.) As mentioned above and emphasised by Tyson, ‘In spite of all this boasting, all is not perfect in paradise’ ([22] p. 15). We do not know what ‘dark matter’ is, or whether there is ‘dark matter’, so the almost quasi-megalomanic statements of some of the scientists of the 19th century have been weakened, to say the least.

The vexed question about science in the 21st c. as waves of mistakes is reflected in the language use of researchers in most scientific domains, or should we call it the usage of special words working as terminology, i.e., scientific nomenclature used in the technical understanding of a field? Seen from the outside, they trigger the core theoretical concepts of the field in question. In, and by, so-called “communities of scientists”, this ‘package’ of words and concepts is presented as knowledge. As implicitly mentioned above, the history of science is one long narrative of misunderstandings, which are in the end being stepwise undermined. And such kinds of “knowledge” are, as is well known, by Thomas Kuhn called ‘paradigms’. The wordings often diffuse into the vernacular of ordinary people. Take, for instance, the word digital. It now means any device that is able to process electronic (i.e, in the meaning of ‘low voltage electrical current’) mechanisms, in spite of the fact that electronic mechanisms need not be digital. Or take the compound noun climate change, a ubiquitous label for many things. In debates it is highly sensitive and, for the record, here I take no stance on the substance of the debates. I shall just point to the etymology of the word climate. It originated

‘from Greek klima “region, zone”, literally “an inclination, slope” thus [the] “slope of the earth from equator to pole”’; see the Etymonline app. So, it was in ‘late 14c., [the] “horizontal zone of the earth’s surface measured by lines parallel to the equator,”…‘from Old French climat “region, part of the earth,” from Latin clima (genitive climatis) “region”’. https://www.etymonline.com/search?q=climate

This means that the original semantics included the essential feature local. From 1975 it (all) changed with the effect that you could manipulate the words and make the nominal compound global climate. The word is, accordingly, a contradiction in terms, i.e., the outcome of an informal fallacy. In principle, you are of course, as a researcher, both able and allowed to call an entity anything you like if your technical terms are well defined, even though nomenclatures are mostly framed in accordance with the Old Greek and Latin tradition. So, what was in the past “local” now has become “global”, and we have the label global climate change. From a linguistics point of view, the expression is meaningless, and a consequence is that its meaning is not just vague or polysemantic but blurred. Nobody knows what an essential meaning of these words is, and even a contextual meaning may not be transparent. It follows that it may be extremely difficult for a human listener or reader to ascertain what people talk or write about when using the words mentioned here. Hence it seems, so far, to be an insurmountable task for an electronic system to produce reliable scientific definitions of labels like this one that can count as knowledge. Personally, I have one all-embracing principle for wisdom as the ultimate sort of knowledge: to be aware of cognitive domains into which I have no, or very little, insights; i.e., in general: to know what you don’t know.

5.2. Post-Postscript

Well, okay. Not in the sense it is illustrated by the French mathematician Bernard Beauzamy in a paper [†] with the title ‘La mystification du Jumeau Numérique: le Paradis Artificiel du Pauvre’ (online 2023). This is a comment on another paper [‡] by the same author, ‘Démonstrations de Sûreté pour la Pile à Combustible. Analyse critique des documents existants’ (online 2022). The paper last mentioned deals with the question of safety when producing and handling fuel cells, including potential complications when using hydrogen in processes. Bernard Beauzamy has a quotation of a remark made by the Directeur du “Digital Experience Center Siemens”: ‘qui a cru bon d’annoncer qu’il ne connaissait rien aux mathématiques’; literally, in my translation (HG), ‘who saw fit to announce that he knew nothing about mathematics’. This statement elicited the following remark in the paper first mentioned above: ‘plus personne ne sait, à l’heure actuelle, quelles sont les limites réelles des technologies dont nous disposons’; in plain English, ‘No one knows at present what the real limits of the technologies at our disposal are.’ But maybe we will all end up in the “Paradis Artificiel du Pauvre”.

† https://www.scmsa.eu/archives/BB_Jumeau_Numerique.pdf

‡ https://www.scmsa.eu/archives/SCM_Pile_Combustible_2022_06_06.pdf