Figure 1. Fundamental elements of any AI system
During the 1980s, working in 'industry', I developed AI (artificial intelligence) systems for the chemical industry, the agriculture sector and the surveying profession, and thereby learned much that is still relevant today, in practice. Since then, with Dooyeweerd's philosophy, I have come to understand why what worked worked and what did not work did not work, and this is what I discuss in this chapter.
During the 1980s, AI was different from today, though the basic principles are the same. All AI systems work according to a knowledge base, a codified version of knowledge that is relevant to the domain of life in which it is to be applied - whether playing Chess, analysing Xrays, analysing ecology of landscapes, driving automated cars, writing reports, advising farmers on the use of fungicides, helping chemical engineers cope with unpredictable and dangerous stress corrosion cracking, and so on. What runs the knowledge base is a software 'engine', which uses the knowledge base to take input (e.g. the user asks "Please write me an essay about ...") to generate output (e.g. an appropriate essay). The difference between 1980s AI and today's AI lies in how the knowledge base is compiled. These are all shown in Figure 1.
Figure 1. Fundamental elements of any AI system
Notice the humans around the AI system, signifying that AI depends on and cannot work without humans. Each human has a different responsibility. The most obvious one is the user - which in automated systems might be at one remove, such as the driver of automated cars, who is responsible for acting well on what the AI system does or tells them. The user can also include other stakeholders, such as passengers in the car, or even people on the street, and so on. The deployer refers to the people who decide to deploy AI, whether it be managers in an organisation or the student who uses ChatGPT to find material for an essay (in which case the deployer is also the main user). Most discussion of "ethics in AI" relate to those two humans, but the other two humans, explained below, also have important responsibility and discussion of "ethics in AI" should include them too.
One challenge in KEAI was tacit knowledge (Polanyi 1967), knowledge that the expert had but could not, or did not usually, tell us. The explicit knowledge, which the experts would tell us, is the kind of knowledge that is taught on courses or written about, but is usually only a part of the knowledge used in practice. The rest is tacit knowledge, of which there are several kinds: muscular tacit knowledge such of how to ride a bicycle, social tacit knowledge of how to operate well in social settings, skills that were learned 30 years ago that have become embedded in the expert's thinking, and so on. For an AI system to work well, including in those unexpected, exceptional conditions that keep on occurring, its knowledge base must be comprehensive and accurate in tacit as well as explicit knowledge. In KEAI, there are ways to make some kinds of tacit knowledge explicit, especially by asking "Why?" and "When not?" etc., which jog the expert's memory on skills they learned decades ago, but with muscular and sensory knowledge this is not possible.
What this means in practice is that KEAI is best when designed to work with the human user rather to replace them. Then some tacit knowledge is supplied by the user, and the explicated knowledge encapsulated in the KEAI knowledge base becomes very useful to ensure that the user does not overlook important questions. This is what made the expert systems in which I was involved successful: we designed them to form human-machine wholes in a way that made interaction very easy and natural so that the good quality explicated knowledge in AI system works smoothly with the tacit and contextual knowledge of the users [Basden ===; Brandon, Basden & Watson ===].
In MLAI, however, tacit knowledge can be learned by detecting patterns from such data insofar as the data has arisen from real life experience. Even the muscular or sensory tacit knowledge may be detected because its effect is detectable in spatial or physical or behavioural patterns. That is one it MLAI's advantages, especially for things like Xray analysis or self-driving cars. Where it fails, compared with good KEAI, is in two ways. One is that if it is not trained to look out for all relevant patterns, it will fail when something occurs for which it was not trained - such as when a driverless car killed a cyclist who was pushing rather than riding their bicycle [Note: Car kills cyclist]. The other is that the knowledge encapsulated is not transparent: in KEAI one can, in principle, ask the AI system "How and why did you come to that answer?" because all the knowledge has been explcitly coded, whereas in MLAI all knowledge is mixed together to become mere numbers, and cannot be unmixed (like the salt crystal dissolved in the sea). It becomes a "black box". There is much discussion about both these problems with MLAI, and some of what was learned during KEAI era is being unearthed.
With that background, we can now address two questions that are relevant to representing knowledge in a computer. (a) In which kinds of application can we expect AI to be successful and in which to fail? (b) What should we do when training MLAI to improve eventual performance? We will employ Dooyeweerd's philosophy in order to do so, and our approach will, as far as possible, apply to both KEAI and MLAI. Other questions, such as the massive power consumption and climate-change emissions of training MLAI, and the ethics of MLAI in use, are not discussed here; some discussion of these may be found in Basden [2008; 2018].
For example, for Chess-playing AI, the knowledge base encapsulates laws of the spatial aspect (for the Chess board), of the kinematic (moves), and of the formative aspect (laws of planning and achieving, and of forming strategies). It might also encapsulate laws of the aesthetic aspect (playfulness and harmony of the whole), and maybe a few others. There is little need to encapsulate laws of the physical or biotic aspects.
If the laws that are encapsulated are not accurate nor comprehensive enough, then the AI system will fail to perform well in some circumstances. It is not always necessary to encapsulate all the laws of an aspect; in Chess, the laws of the kinematic aspect that are needed are very limited: only small and very restricted movements forwards, backwards and sideways, without laws about speed etc.
In automated driving, the laws of the following aspects at least are needed:
MLAI must be trained in all the relevant laws and meaningfulness of all relevant aspects. However, it is wise to go beyond that which seems relevant, in two ways. One is that, in use, the AI system is likely to meet exceptional circumstances, which is it not always easy to anticipate. A self-driving car killed a cyclist because, though it had been trained to recognise cyclists riding bicycles, it had not been trained to recognise cyclists pushing their cycles [Note: Cyclist]. The other is extensions of use beyond the original intention should be anticipated, such as driving off-road or along farm tracks where there are animals rather than people. Or, if a variant of Chess were devised where speed matters, the Ai Chess player would need to encapsulate the speed laws of the kinematic aspect. As we will see below, human knowledge elicitation (to anticipate such possibilities) cannot be entirely bypassed.
Large Language Models (LLMs) like ChatGPT are much more compplex than self-driving cars, because their main aspect is the lingual. Indeed, because the 'mandate' of the lingual aspect is to signify 'pieces of meaning' from any and every aspect, LLMs need to encapsulate some knowledge of every aspect, and in three different ways. One is semantics, the broad sweep of the kinds of meaningfulness that need to be signified. Words can be about amounts, spaces, movements, ... friendships, money, ... goodness and faith. So, to respond meaningfully to user questions and instructions, LLMs must encapsulate knowledge about every aspect. The second is the structures of language, its syntax: LLMs must encapsulate the laws of syntax of every language group they purport to work in, and this involves functioning in pre-lingual aspects, and some post-lingual aapects. The third is the pragmatics of language, what speakers intend to as they utter. This refers mainly to functioning in post-lingual aspects, such as connotation (social functioning), truthfulness (juridical functioning), sincerity or not (pistic functioning), presuppositions (another pistic functioning), humour (aesthetic functioning) and so on. Since the actual human text used to train LLMs amalgamates and expresses all three together, the LLM homogenizes all three together in its knowledge base. ChatGPT stores all the different word-concepts it learns in a massive matrix in which each row is a word-concept with 12288 columns, each being a parameter with a numeric measure of how much each parameter is true of that word. For example, cat would score well on parameters like pet, mammal, noun and independence (if there were such parameters). ChatGPT's engine makes use of this matrix by well-known matrix-manipulation arithmetic. It turns out that LLMs only poorly and partially encapsulates laws of pragmatics, even though it processes such tacit kg in its training data [Hu et al. 2023; Sravanthi et al. 2024] - which is one important reason for LLM failure.
Understanding the knowledge base as a law-side for a virtual world, which incorporates laws of multiple aspects, can help us understand why AI is more successful in some kinds of application than others. By and large, the laws of earlier aspects are much simpler and more limited than those of later aspects. This is for two reasons. One is that the laws of any aspect depend foundationally on (and thus incorporate) the laws of all earlier aspects, in principle. For example, the laws of the physical aspect depend on those of the kinematic, spatial and quantitative, the laws of the social depend on all these plus those of the biotic to lingual aspects. So, just from this, the complexity of laws of an aspect (as expressed quantitatively) increases faster than a Fibonacci series. However, second, the laws of later aspects are themselves more complex in their own right, partly because they are less and less determinative [De Raadt 1991] with more possible choices.
What this entails is that much more ML training is required for AI applications in which later aspects are important, with ever increasing amounts of training data. ChatGPT, for which the lingual is the main aspect, which used the Internet as a source of data 'scraped' 45 Tb of text off Reddit. Given the limited time and resources available for training, this means that AI is likely to be more successful in applications meaningful in earlier aspects.
If the AI system is intended to replace humans (that is, to give reliably correct answers without much human intervention being necessary, such as in Chess or self-driving cars) the knowledge base must incorporate faithfully all the relevant laws of every relevant aspect, so only early-aspect applications are likely to be successful. However, if the AI system is to work only in close collaboration with human beings, who will detect and correct its mistakes, then it can find some success in later aspects. Thus, students who assume ChatGPT will write essays for them are likely to gain poor marks, while students who use ChatGPT to supply ideas for essays, and then carefully check and expand those ideas, then write the essays, might get better marks. The challenge is to design the user interface of the AI system so that it discourages the one and encourages the other. During the 1980s, one of my colleagues was very clear that he wanted the AI system to discourage people believing its advice and to explore the roots of it. Increasingly many recognise the need for AI to work with, rather than replace, humans, but MLAI is hampered by a fundamental inability to explain its decisions, which is essential for proper working with humans. KEAI approaches might be coming back into fashion.
The algorithm designer has the prime responsibility for making sure that the way knowledge can be encoded, the way it is stored in the knowledge base and the way it is activated by engine are fully appropriate to all the aspects that are relevant in the application. For the quantitative aspect, this is usually easy because computers already do arithmetic, though more is needed for matrix manipulations or dealing with prime numbers.
With the spatial aspect it is little more complex. Conventionality, two-dimensional spatial figures like routes or shape boundaries, such as the U-shaped woodland in Figure 2(a), would often be stored in the knowledge base as a list of (x,y) coordinate pairs.
Figure 2. Complex shape: (a) original, (b) expansion, (c) further expansion to become a shape with hole.
Suppose that the AI system needs to deal with expansions of the boundary as in Fig 2(b). (For example, in the area around the woodland birds nest on the ground and predators emerge from the wood each night up to 100 m to prey on nests and the AI system is to predict which nests are under threat.) That is still supportable using a list of (x,y) pairs. But suppose the expansion is larger, as in Fig 2(c), then we get a shape with a hole, which is no longer supportable with one list; it needs two, or more depending on the complexity of the shape. The conventional, tacit assumption no longer holds. Storing shapes as pixel arrays avoids that problem but introduces other problems. What this entails is that the algorithm designer is not merely a lowly backroom-worker but must be expert enough in every relevant aspect to know the unusual exceptions that can occur, and find ways of coping with them appropriately in the software of the knowledge base and engine.
This can explain why ChatGPT makes so many errors. First, the numbers are merely probabilistic, so can be wrong in specific cases. Second, it reduces the lingual aspect of signfication to the quantitative. Third, it is doubtful that the 12288 parameters cover all the laws of all aspects - the laws of the lingual aspect, and all the later aspects whose meaning may be signified in words. This type of error will occur even without the first two error sources. This third source of error can only be understood properly using a philosophical paradigm that allows mutually irreducible "modalities of meaning" [Dooyeweerd 1955, 24], i.e. aspects. There is a fourth type of error that afflicts large language models: cultural and contextual bias, arising because the training data (Internet text) was mostly written by affluent, educated people of the Global North. Again, this may be understood aspectually, such as that the pistic/faith aspect, which gives meaningfulness to religious beliefs, is downplayed in such writing because of the secular presuppositions that prevail in that culture.
Therefore - and this is the answer to the first question we are posing - MLAI is likely to be reasonably reliably successful in applications in which the later aspects are largely irrelevant, and is likely to be unreliable, if not entirely fail, in applications in which later aspects are of importance.
This is largely, though usually tacitly, recognised across the MLAI research community and even the industry but this fact is not properly understood and is seldom discussed. Various avenues of overcoming errors are being explored, but the presupposition of AI as a whole replacing any and all human activities is still too strong. Dooyeweerd's aspects makes it possible to recognise these issues, frame them and explore them. Understanding this via aspects can help, along with perhaps a return to some measure of KEAI. We will see this as we address the second question.
In KEAI, this requires highly sensitive and skilful interviewing of experts, understanding what they tell, and understanding of other sources of knowledge, in every aspect relevant to the domain. In my experience, the good knowledge engineer would proactively challenge the experts about exceptions etc. and would benefit from apparent disagreements among experts by exploring the root of disagreements, rather than merely accepting some kind of 'democtratic' average.
For example, in my work with two corrosion engineers, when I asked why their knowledge differed, it was revealed that one worked with chemical reactions above 300 degrees Celcius, and the other below, and the chemistry changes there. This oopened up a lot of avenues for new knowledge. I had to have at least some understanding of the physical aspect to do this. In another application, in agriculture, proactively asked what advice would be given to farmers who wanted to reduce their chemical input, but had to persevere through a barrier presupposition of ever-increasing chemical use - and was motivated to do so by my Christian beliefs about our responsibility to ecology. Again, this opened up new avenues for knowledge elicitation, and also made the final AI system more flexible and trusted when it went into use.
It is not usually easy for MLAI to take such actions.
In MLAI, completeness and accuracy are obtained by good selection of the massive amount of training data and of the parameters on which it is trained. By and large, the training parameters should be those meaningful in all the aspects relevant to the application.
There are normally seven stages in training: identifying the parameters on which to train the knowledge base, collecting data, pre-processing, selecting an appropriate model, actual training, evaluation and refining. Contrary to popular assumptions that machine learning requires minimal human input, that is true only of running the training software; human input is vital in all other stages for a good AI system that will actually work well. For example refining includes finding ways to label content as pornographic or violent, or correcting cultural bias. In every one, what is meaningful is the fundamental guiding factor, and Dooyeweerd's aspects, as modalities of meaning, can serve a very useful purpose in each by offering us a taxonomy of meaningfulness that can be used almost as a type of checklist and cue to stimulate better thinking. Here I will illustrate this with just the first two stages.
Stage 1, parameter selection. (Actually, some parameter selection can happen as part of later stages, but it it useful to have a clear idea of what kinds of parameter are meaningful and important as a guide to all the other stages.) ChatGPT has 12288 parameters. I have not yet been able to discover how they were chosen, and assume that many of them were themselves selected by some AI system built for that purpose. However, that process would have begun by humans choosing a number of parameters as a starting point. On which parameters the knowledge base is trained is crucially important, because, if any are missing, such as about road signs in self-driving cars, then the AI system will not work well. Many of them are not obvious, and must be discovered carefully by processes similar to the old knowledge elicitation. Any AI generation of parameters from the starting set would magnify any errors in the original set, especially errors of cultural bias and the ignoring of aspects. So, the AI developer has a responsibility to check and double-check the set of parameters finally employed to train the knowledge base. So MLAI parameter selection can learn from KEAI.
Since the meaningfulness of parameters is always by reference to aspects (for example, the age of a person is meaningful in the biotic, psychical and social aspects at least), separating out their aspects can be a great help and, having done this, the AI developer can then ask "What other parameters might be meaningful in this aspect?". Towards the end of the KEAI era, Mike Winfield [1995; 2006] devised a method of obtaining a good set of parameters using Dooyeweerd's aspects, including those often overlooked: MAKE, Multi-aspectual Knowledge Elicitation. An expert in the application area is interviewed, but first has Dooyeweerd's aspects explained briefly, to use as a checklist for discussion. The analyst begins by asking "Can you identify a couple of aspects that are important to you?" and then "Please tell me some parameters that are meaningful in these aspects." Ideally, aspects and parameters are plotted on a large piece of paper, as shown in Figure 3.
Figure 3. Simple MAKE diagram generated for veterinary practce.
As more parameters are identified other aspects become relevant and are added to the diagram, along with more parameters. Lines indicate relationships among parameters and even aspects. After the flow of parameters slows, it is appropriate to ask, "Are there any other aspects that we have not covered, which are important?" and, if so, these are added in similar manner. In his research, Winfield that interviewees identified almost every aspect as relevant. He found that frequently tacit knowledge would be made explicit and that interviewees greatly valued the process.
Imagine a much more comprehensive diagram of this kind. It would most likely provide a good starting set of parameters from which to begin parameter selection.
Stage 2, collection of training data. The selection of training data is likewise an expert task, the quality of which affects that of the trained knowledge base. Any bias therein, including cultural, will affect the working of the AI system, as will any inaccuracies. ChatGPT obtained 45 Tb of compressed text from Reddit [Brown et al. 2020], so we may ask how accurate, complete and unbiased is Reddit's contents. Moreover, most data available has arisen from past activity, which means that the knowledge base tends to be backward-looking and not very open to future possibilities.
Reference to Dooyeweerd's aspects (maybe via a MAKE exercise) can help in three ways, in several stages of the training process.
A useful measure of bias would be to count the number of pieces of knowledge meaningful in each aspect. I cannot find any evidence that creators of LLMs have done anything like this.
Even without such a measure, it is still the responsibility of the AI developer to strenuously aim for unbiasedness, accuracy and completeness. An exercise like MAKE can guide them to obtain training data valid for each aspect, especially those often overlooked, in that it can remind the designer in which areas the AI system needs to be trained.
2. Today's machine-learning AI can learn from earlier knowledge-elicitation AI because in both cases the accuracy, completeness and unbiasedness of the knowledge base crucially affects how the AI system operates.
3. AI will tend to be more successful and reliable when the main aspects in which its application is meaningful are the earlier ones. It will tend to be less reliable for later aspects. Yet it can still be useful if used to assist rather than replace humans.
4. The process of designing an AI system and training its knowledge base can be greatly assisted by reference to Dooyeweerd's aspects, as fundamentally distinct ways in which things can be meaningful.
5. We have addressed only two questions about AI, about success and about training; discussion of other questions, such as benefits and harm in use, and impact on society, may be found in Basden [2008, 2018].
Basden A. 2018. Foundations of Information Systems: Research and Practice. Routledge, London, UK. See its web page.
De Raadt JDR. 1991. Information and Managerial Wisdom. Paradigm Publications, Idaho, USA.
Hu J, Floyd S, Jouravlev O, Fedorenko E, Gibson E. 2023. Proc. 61st Annual Meeting of the Association for Computational Linguistics, Vol I: long papers, 4194-4213.
Sravanthi SL, Doshi M, Kalyan TP, Murthy R, Dabre R, Bhattacharyya P. 2024. PUB: A pragmatics understanding benchmark for assessing LLMs pragmatics capabilities. Findings of the Association for Computational Linguistics, ACL 2024, 12075-12097.
This page, "http://dooy.info/using/keai.html", is part of a collection that discusses application of Herman Dooyeweerd's ideas, within The Dooyeweerd Pages, which explain, explore and discuss Dooyeweerd's interesting philosophy. Email questions or comments are welcome.
Written on the Amiga and Protext in the style of classic HTML.
You may use this material subject to conditions. Compiled by Andrew Basden.
Created: 15 August 2025. Last updated: