Artificial Intelligence: An Holistic Perspective
We face five major challenges with Artificial Intelligence (AI). Two occupy most of our attention today, whether computers can ever be like humans, and how to make AI algorithms more accurate, and I will leave those till later while I discuss the others. The five challenges are:
- Using AI so that we gain benefit and not harm;
- Development of good AI systems as applications, which benefit most widely and cause least harm;
- The impact of AI on society and of society's presuppositions on AI;
And the two we already discuss most:
- The algorithms of AI;
- The "AI question" of "computer = Human ?"
This page discusses these issues employing the rich philosophy of Herman Dooyeweerd. Dooyeweerdian philosophy offers a radically different way to address many topics in the areas above, because he enables us to understand human and non-human, good and harm, diversity, coherence, meaning, human functioning, history, and the activity of scientific research. I will give links to these things where necessary. Though addressed from a philosophical perspective, I try to make it readable by those of different perspectives, because I have found Dooyeweerd's philosophy can be intuitively grasped by many.
Why should I write about AI? I began practical and academic activity in AI in the early 1980s, and developed expert systems in the 'real world' if materials technology, agriculture, quantify surveying and contract law, and advised on others. I developed a methodology for building expert systems, and innovative software tools too, then broadened my interests to human factors and digital informations systems in general, from which the above main areas emerge. In this page I apply them to AI.
Using AI so that we Gain Benefit and Not Harm
This requires that we understand benefit and harm of AI when in use. We would us Dooyeweerd's aspects to separate out kinds of benefit and harm. Each aspect, from the biotic onwards, makes meaningful a different kind of good and harm.
For how this works, see Chapter 6 of Foundations of Information Systems: Research and Practice.
Methods for systematic analysis of good and harm in each aspect have been developed. See Chapter 11 of Foundations and Practice of Research : Adventures with Dooyeweerd's Philosophy
Development of Good AI Systems
There are two kinds of AI:
- Knowledge representation (KR), which was really popular in the 1980s, and
- Machine learning (ML), which is popular today.
They are developed in different ways, but both aim to embed or embody requisite, appropriate knowledge into the AI system, so that, when run, it operates as it should, with "intelligence". KR AI is constructed by eliciting human knowledge, especially tacit knowledge. ML AI is constructed by training a general-purpose learning algorithm, usually a neural net. But the result should be the same.
The main difference in use is that in ML AI, the knowledge embodied is completely opaque, unable to be investigated, while in KR AI (expert systems), the knowledge is to some extent transparent, able to be investigated, so that answers may be given to the user's question, "Why did you recommend that?" Whether such transparency is important depends on the application; for example, picture analysis might not needed while AI used to make legal judgements certainly does need it.
Constructing application AI systems either way requires taking responsibility in four major activities that constitute the development of digital systems:
- Understanding the usage context and deciding what AI system to develop;
- Obtaining the relevant knowledge to enter into the AI system for the application (in Knowledge Representation AI, this means eliciting human knowledge; on Machine Learning AI, this means selection of the variables on which the AI system is to be trained and then obtaining and cleaning of the huge corpus of training data);
- Developing appropriate basic AI algorithms, suited to the kind of domain, or else ensuring that ones taken 'off the shelf' are adequate; also developing appropriate user interfaces and interfaces to other machines;
- Harmonizing the entire project (which is often called Project Management but in fact is the responsibility of all involved).
For how this works, see Chapter 9 of Foundations of Information Systems: Research and Practice.
Methods for systematic analysis to identify stakeholders and potential usefulness have been developed. See Chapter 11 of Foundations and Practice of Research : Adventures with Dooyeweerd's Philosophy
AI and Society (in Both Directions)
This involves addressing two major issues:
- The impact of the AI system if it, or similar systems, come into widespread use. Dooyeweerd's aspects can help us think about potential issues we have overlooked, such as indirect impacts like climate change.
- How AI might change the very structures of society and culture, and how these might impact the direction in which AI develops. These include both visible structures like laws, policies and prodedures, but also two kinds of invisible structure, the attitude that pervades society (whether self-giving or self-centred), and the beliefs, aspirations, expectations, assumptions that prevail in society, These are meaningful in the juridical, ethical and pistic aspect.
See Chapter 8 of Foundations of Information Systems: Research and Practice for discussion, and detailed development of those.
This is the development of basic AI algorithms and data structures. For example, AI system that analyse photographs require different kinds of basic algorithm from those that carry out a conversation. Each kind of AI algorithm and data structure must incorporate (embody) the fundamental laws of the relevant aspects. So, for example:
See Chapter 7 of Foundations of Information Systems: Research and Practice for discussion, and detailed development of those ideas.
The "AI question" of "computer = Human ?"
This involves philosophical understanding of the nature of computers and human beings. First, we see an AI system, or a computer program more generally, as a virtual law-side, an embodiment of laws of various aspects by which the information-based entities within its world will operate; these laws can try to match those of the real world as far as possible, or they can be modified (such as in computer games, death is not final, but players usually start again).
Given that, then there are two answers to the AI question, "Computer = Human?"
- "No!" From the perspective of pure subject functioning, only human can truly function in all aspects as subject, while the computer functions as subject only up to the physical aspect. This is why the operation of computer / digital systems is largely predictable and deterministic: the four earliest aspects have largely deterministic laws.
- "Yes!" However, from the perspective of meaningful functioning, which includes subject and object functioning - and this may often be a more realistic and relevant way to think, the AI system can be said to function in any aspect for which is it programmed or trained. Thus AI systems might discover new minerals (analytical aspect), might play games (formative aspect), might hold conversations (lingual aspect), might make financial decisions (economic aspect), and so on.
See Chapter 5 of Foundations of Information Systems: Research and Practice for discussion, and detailed development of those ideas.
This page, "http://dooy.info/using/ai.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: 7 June 2021
Last updated: 12 October 2022 completed first brief version.