Roger Clarke's Web-Site
© Xamax Consultancy Pty Ltd, 1995-2026
Infrastructure
& Privacy
Matilda
Roger Clarke's Web-Site© Xamax Consultancy Pty Ltd, 1995-2026 |
|
|||||
| HOME | eBusiness |
Information Infrastructure |
Dataveillance & Privacy |
Identity Matters | Other Topics | |
| What's New |
Waltzing Matilda | Advanced Site-Search | ||||
Revision of 3 October 2025
© Xamax Consultancy Pty Ltd, 2025
Available under an AEShareNet 
licence or a Creative
Commons 
licence.
This document is at http://rogerclarke.com/EC/AIBB.html
This page provides a backgrounder in so-called 'Artificial Intelligence' (AI) and a number of closely associated technologies. Unlike most such backgrounders, it does not wear rose-coloured glasses. There have been serious problems with the notion of AI from its very beginnings, and these have compounded and deepened over the decades. The infectious enthusiasm of earlier times has been replaced by irresponsible misrepresentation of technologies and their capabilities. This document brings a sceptical eye to the subject-matter, as an antidote to the widespread excesses of over-optimistic providers and promoters, and the blind faith of a large proportion of users and user-organisations.
This backgrounder commences with a short review of the state of information technologies during the first quarter of the 21st century. It then introduces the often vague notion of AI, identifies some exemplars, discusses the different kinds of languages used to develop them, and outlines the threats inherent in AI. The nature of the data on which the technologies depend is considered. Attention is then turned beyond intelligence and software to associated technologies in the areas of data analytics, content creation, and physical action in the real world.
Finally, an alternative conception of the overall field is proposed, which enables an integrated view of the whole. This lays a foundation for sober assessments of how humankind can harness the ideas that are valuable, can identify and avoid ideas that are inappropriate and dangerous, and can manage the predictable harm and the contingent risks that are inherent in these powerful information technologies.
Advanced information and communications technologies promise and threaten further substantial changes in economies, societies and polities. Those technologies also directly impact individuals' experience of, and interactions with, the physical world, and in and with virtual worlds.
The purpose of this short document is to outline these technologies and the mutually reinforcing nature of the interactions among them. Links are provided to accessible explanations of some aspects at somewhat greater levels of detail.
As a starting-point, I'm assuming that readers are familiar with information infrastructure and tools of the early decades of the 21st century. Key elements are computer-based devices designed for individual human users. Broadly speaking, three categories are usefully differentiated: general-purpose computers referred to as desktops and laptops; tightly-controlled 'appliances', particularly mobile phones and tablets; and smaller forms such as cards, watches and rings with computing chips embedded in them. Large-scale computing resources used by organisations are referred to as 'servers' and 'server farms', to distinguish them from individual users' 'client devices'. Increasingly, computing capabilities have also been embedded into other artefacts as diverse as motor vehicles, many components of buildings, and cameras.
Inter-communications capabilities among devices have long since become entrenched, including 'wired' forms using metal and optical fibre cabling, and 'unwired' forms using various parts of the electromagnetic spectrum. Some channels are designed for communications over short distances and others for tele-communications over distance. Many services are overlaid over communications infrastructure, some intended for specific purposes and others for general usage.
These linked computing and communications facilities have had dramatically increasing volumes of data available for processing and transmission. During the second half of the 20th century, 'data capture' into machine-readable form needed to be performed -- although the act is better described as 'data creation'. Each data-collection was created for a purpose. The definitions of what each data-item meant, and how carefully its quality was assured, had little regard for any other purposes.
By the beginning of the 21st century, on the other hand, data was mostly 'born digital'. Some digital data is input by organisations' employees, and much of it by people external to the organisation that acquires it. Some of the data arises from conscious acts by customers, clients and other individuals using self-service facilities. However, people's actions generate a great deal of data without their knowledge. Data of that kind is transmitted and widely used without consent by the people it relates to. Here is further discussion of the digital surveillance economy (s.4).
In addition, some digital data is generated automatically by computer-based systems. This is done by combining data that those systems already hold with data arising from new transactions. Some data is also generated by so-called 'analogue to digital converters' (ADCs), which sense some feature of the real world (such as temperature, movement, sound, light, or some other segment of the electromagnetic spectrum) and generate digital data.
The resulting, vast volumes of data are subject to a great deal of expropriation by additional organisations, and re-purposing to many and varied uses. This breaks the linkage of data's meaning and quality with the original purpose of collection. Expropriated data from multiple sources is then subjected to data consolidation and inter-relationship with other data, with little regard for compatibility among the data's varying meanings and varying quality-standards. This results in a great deal of scope for material errors in inferences drawn, decisions made, and action taken. Despite that, organisations commonly assume that data has adequate reliability and quality even though it has been expropriated, re-purposed and consolidated.
A further element that is all too often overlooked is the existence of means whereby artefacts can directly impact the real world. The generic term 'actuator' encompasses robotic arms, card-sliders in ATMs, the automated operation of sluice-gates in dams, programmed changes in traffic-lights, and the hydraulics of aircraft controls. It is also applicable to workaday functions such as the automated despatch of communications (invoices, letters, emails).
Here is a somewhat deeper summary of those technologies (s.4.1) that was published in 2018. That segment is followed by a review of aspects relevant to polities, economies, individuals, groups and societies (s.4.2).
Computing devices, communications links, data, and actuators represent inert infrastructure, with latent power. Software utilises that infrastructure in order to apply that power. Software is a generic term for individual programs and suites of programs that contain, or may give rise to, precise instructions in the particular 'machine-language' or 'instruction-set' at the heart of each computing device. It's useful to distinguish three categories of software.
Firstly, the term 'systems software' refers to software deep inside computers. It overlays the bare device and networks, and provides services to the developers of application software. However some kinds of systems software are of direct value to users. On example is tools to manage documents and other files, such as Finder under macOS and File Explorer on Microsoft Windows.
The second category is 'software development software'. This includes compilers and interpreters to enable programmers to write software in a convenient language and have it converted into the relevant device's machine-language. Other important examples are editors for capturing and editing programs, and tools for testing programs and for identifying the errors that are causing software to deliver results that are not consistent with the program designer's intentions.
The third category 'application software' is what users of devices are familiar with This applies the device's capabilities to humanly-useful functions such as managing a photo-album or a family-tree, supporting message interchange with other people, or enabling the preparation and editing of documents. During the period 1975-2000, increasing numbers of people had access to personal computing devices that ran software under their control, and maintained their data locally.
Since the beginning of the 21st century, however, IT providers have changed the architecture of computing and communications technologies. The new pattern is referred to here as 'remote software and data services'. Here are the key steps that enabled IT providers to gain ascendancy over consumers:
Through these mechanisms, technology providers have steadily captured their customers by withdrawing access to application software that runs on the person's own devices, and attracting people away from general-purpose computing devices towards provider-controlled 'appliances'. Their purpose was to trap people into reliance on remote services such as Microsoft 365, Google Docs, digital music streaming and ancestry.com; and they've succeeded
The languages in which developers write software are usefully categorised into different generations. In the 1st generation, all programming was performed in the particular device's machine-language. Shorter and more convenient expressions were devised, giving rise to 2nd-generation 'assembler languages'. These were complemented and mostly replaced by a 3rd generation of languages that supported the expression of solutions to the kinds of problems addressed in particular professions. For example, science and engineering had a focus on translating formulae -- for which Fortran was devised. For accounting applications, on the other hand, a common business-oriented language was developed, COBOL. The more technical forms of the 3rd generation express solutions as 'algorithms', whereas, in the more pragmatic language of business, the form of such solutions is referred to as 'procedures'. So the 3rd generation is generally referred to as comprising algorithmic and procedural languages.
A 4th generation, commonly thought of as 'program generators', took advantage of the fact that a great many programs share a lot of structural similarity. Much of each new program can therefore be generated from a few parameters, and a 3rd generation language can be used for the smaller amount of code needed to customise the generic solution to a specific purpose.
Up to the 4th generation, the focus is on problem-solutions. The 5th generation shifts to a higher level of abstraction, and instead model a problem-domain. Successful approaches in this generation used logic programming and rule-based expert systems.
The 6th-generation approach is yet more abstracted from the level of problem-solutions. It abandons the rational modelling of problem-domains. It instead relies on untidy heaps of empirical data being processed by general-purpose analytical tools in order to draw inferences. Typical of the 6th generation are AI/ML techniques such as neural nets. These are further discussed below. Here is a lengthier explanation of the generations of software development languages (ss.2-8).
On the basis of that scan of information technologies prevalent in 2025, the following sections consider a number of advanced, powerful and inter-linked technologies. Together, they are promising and threatening changes of a highly transformational nature, that is disruptive and even destructive.
The term 'artificial intelligence' (AI) was coined 70 years ago. The meanings ascribed to the term have been many and varied. My contention is that, even in its most coherent form, it embodies, and always has embodied, a serious and harmful misconception. A major shift in the discussion is long overdue. This section begins the discussion by outlining three interpretations of what AI means. It then identifies some key exemplars of AI, and some tools used in the development of AI software. Finally, it offers an initial broad scan of the dangers inherent in AI.
In 1955, a group of researchers applied for funding for a symposium to discuss the use of a computing device to create artificial forms of human intelligence. The event stimulated many people's imaginations, and the term quickly entered the mainstream. Within five years, however, what started as "conjecture" and "hypothesis" had been degraded into wild predictions. These were about "duplicating the problem-solving and information-handling capabilities of the brain" and "substituting machines for any and all human functions in organisations", which were confidently predicted within one-to-two decades, i.e. by 1980. More is available here, on the initial definition and subsequent claims (s.2).
The original conception can be seen as a 'grand challenge'. One reason to doubt its value is to ask: With close to 3 billion human intelligences on earth in 1955, and over 8 billion 70 years later, what benefit does humankind or the biosphere gain from devices that replicate human intelligence?
The original idea has come to be referred to as 'artificial general intelligence' or 'strong AI'. Most AI practitioners have long since divorced themselves from that objective. They commonly distinguish their more practical approaches by referring to the original notion as having "aspired to replicate" human intelligence, whereas they regard themselves as being "inspired by" human intelligence. AI practitioners tend to focus on particular ways of tackling particular categories of problem, and only occasionally venture into meta-questions about the nature of their endeavours. As a result, a sufficiently comprehensive and coherent definition of AI in contemporary practice is difficult to find. An interpretation based on multiple sources is:
Whether or not an application is AI software does not depend on which generation of software is used in its development. Among practitioners, the key factors that define AI are partly apparent world-awareness (1. "evidences perception and cognition") and partly embodiment in the software of intentionality (2. "has goals" and 3. "formulates actions" to achieve them, and possibly 4. "implements those actions").
A third interpretation of AI appears to be achieving mainstream adoption among inter-government agencies. After some simplification of bureaucratic expression, the European Commission, the Council of Europe and the OECD consider an "AI system" to be a computer-based set of interacting processes that has some level of autonomy and some sense of objectives, and that draws inference(s) from input to generate output.
In comparison with the interpretation of AI practice, above, this adopts a weak form of 2. (in that "has goals" is instead "has some sense of objectives"), a very limited form of 3. ("formulates actions" to achieve goals is instead "draws inferences"), and uses the undefined term "autonomy" which only possibly extends to 4. ("implements those actions") and only by vague implication encompasses 1. ("evidences perception and cognition"). More is available, on this interpretation of AI in a regulatory context (s.2.2).
The variations in interpretation of AI are considerable, and this ensures mutual unintelligibility of discussions about AI among communities that have diverse disciplinary and professional backgrounds, and diverse objectives. The final section of this backgrounder argues that confusion will reign, technologies will run out of control, and serious harm will increasingly arise, unless and until two steps are taken: the abandonment of the misconceived notion of 'artificial' intelligence, and the substitution of terminology and definitions that reflect the human purposes to which the technologies are applied.
As a complement to that abstract and dry discussion of what AI means, it is helpful to identify some key exemplars of AI and AI systems that are in use, or that are being piloted. Many of the standout success stories from AI research had been in areas of pattern recognition. Progress in relation to images gave rise to optical character recognition -- OCR, automated number plate recognition -- ANPR, and computer vision. There have also been successes in the area of sound, such as music, by means of acoustic fingerprinting techniques.
The field of natural language processing (NLP), and in particular natural language 'understanding' (NLU) analyses words and phrases in text, in order to extract a structured representation of the text's content. For example, it may recognise the nouns that define the subject-matter, and identify verbs and verb-forms that are related to the requestor's intent, thereby distinguishing a question (seeking information) from, say, an instruction to conduct a transaction. The end-result is a set of stored tokens that contain a model of the source-texts. That model is then available for further processing for various purposes. A rather different development has been natural language generation (NLG), whose function is to synthesise syntactically acceptable text from some structured representation of information.
During the 2020s, those two longstanding areas in the natural language space have been joined by a third. As a prelude to that third area, the field of machine learning (AI/ML) needs to be introduced. This comprises a suite of techniques that firstly identify patterns and structures within a heap of data provided to them, and then express them in a model. That model is then used to classify new instances of data. The techniques are being applied to the drawing of inferences, which are then available to decision-making processes. Software development techniques for AI/ML are considered in the immediately following section, and the topic is revisited briefly later, in the context of data analytics.
A recently-emerged category of AI applies similar techniques to AI/ML, but for a different purpose. Rather than using models to categorise instances of data, as AI/ML techniques do, the purpose of Generative AI is to produce new instances of data that are intended to have particular attributes. GenAI models can produce new instances of desired data intended to represent many different aspects of reality. For example, large volumes of audio (sound) data can be used to generate a model, which can be requested to synthesise new sounds. Those sounds might resemble (and perhaps be indistinguishable from) music, or a particular human's voice. Diagrams and animations can be produced in a similar manner. So can images, and hence video (which is merely a linked series of images, perhaps with an accompanying sound-track). GenAI models that handle text are considered later in this backgrounder.
As indicated earlier, AI software may be produced using any generation of software development tool, and in practice many have been produced using 3rd-generation algorithmic or procedural languages. However, 5th-generation logic programming and rule-based expert systems have been natural alternatives for many AI practitioners.
The 6th generation of tools is less driven by rational approaches to problem-definition, problem-solving and problem-domain modelling. Rather, they exploit large volumes of empirical data by means that generally involve far less systemic reasoning about causality, and far more reliance on correlation. Multiple approaches fall within the generic notion of machine learning (AI/ML). The scale of computing power and high-speed memory reached a flex-point in about 2010. Since then, variants of artificial neural network (ANN) techniques have been the subject of a lot of research, and some amount of active deployment. The following section considers their application to data analytics. The section after that discusses their use for Generative AI, including text-generating tools such as the ChatGPT family. More on the differences between 3rd, 5th and 6th generation techniques is here (ss.3.3., 3.4) and here (s.6).
One particularly vital distinction among the 3rd, 5th and 6th generations of software development tools needs to be appreciated. With 3rd-generation approaches, an algorithm or procedure exists, and from that a humanly-understandable explanation for an inference, decision or action can be readily constructed, e.g. 'After computing your remaining disposable income, and comparing it against the repayment schedule, it fell short of the threshold we require to provide you with that loan'. In practice, only a small proportion of inferences, decisions and actions are disputed. Hence, in the large majority of cases, no-one calls for an explanation and none is offered. What matters, however, is that disputes are capable of being investigated. It tooks years for the depravity of Robodebt to be confirmed by Australian courts. That was possible, because the software was developed using a 3rd generation tool, and hence the illogic and illegality of the scheme could be exposed.
With a 5th generation tool, it may be theoretically feasible to incorporate into the design the generation of a (perhaps somewhat differently structured) humanly-understandable explanation. This might not identify all of the relevant factors, e.g. it might focus on a single disqualifier, such as age, without including mention of the fact that disposable income was also too low to justify a loan. (This is akin to the expression commonly seen in court judgments, along the lines of 'Having already determined the matter, I have no need to give further consideration to whether factors X, Y and Z apply').
The situation is very different in the case of the purely empirical approaches that are typical of 6th-generation development tools. No humanly-understandable rationale exists. Furthermore, such explanations as can be generated are of the nature of 'The state of the weightings on the nodes of the neural network at the time your data was run against it was such that your application was declined'. A field of 'eXplainable AI' (XAI) has arisen in an attempt to address this inadequacy. However, the field remains at this stage 'aspirational', meaning that little or no progress has been made to date.
The substantial absence of an ability to express a decision-rationale creates serious problems for a large proportion of applications produced using 5th and 6th generation software development techniques, including contemporary approaches to AI/ML. Without a humanly-understandable explanation, it is infeasible for decisions to be reviewed, audited, and either justified or withdrawn. In short, accountability is undermined because affected parties, tribunals and courts cannot be provided with humanly-understandable explanations for decisions. This has substantial implications for the less powerful (typically, consumers, citizens and small business enterprises), but also for regulators, courts, policy agencies and parliaments, and for risk management and the insurance industry.
A great deal has been written about the threats that AI entails. Unfortunately, a great many different approaches have been adopted, and the opaqueness of many of the technologies invites confusion. The threats can be usefully clustered into five categories::
An attempt at a summatory sentence is:
AI gives rise to errors of inference, of decision and of action, which arise from the more or less independent operation of artefacts, for which no rational explanations are available, and which may be incapable of investigation, correction and reparation
Further discussion is available here on the threats inherent in AI (s.4) and again here (s.8), and on broad areas of negative impact (s.9).
A wide variety of techniques have been developed to enable inferences to be drawn from data, as a basis for decision-making and action. Such activities are variously referred to as 'data mining', 'data science' and 'data analytics'.
For decades, much of the data that was available for analysis was in data collections created by a particular organisation for a particular purpose. The last few decades, however, have seen far more data created. Further, since about 2000, vast volumes of data have been traded and expropriated out of the data's original context. Proponents have waxed lyrical about the promise of such 'big data' for new insights, competitive advantage, and profit. Disparate data collections have been consolidated, and data analytics techniques have been applied to consolidated data heaps, with little regard for the diversity of data purposes, data meanings and data attributes. Here are discussions of quality issues in relation to big data (s.3), and issues arising from analytic techniques applied to it (s.4). Here is guidance for responsible use of big data analytics (s.3).
AI techniques referred to as 'machine learning' (AI/ML) have been applied in the data analytics field. In one commonly-adopted approach ('supervised learning'), a (usually quite simple) model of some potentially relevant factors is provided to a piece of generic software. A 'training set' of data is fed in, resulting in weightings being associated with the various factors. The process of creating this artefact is referred to as 'machine learning', although 'computer-supported model development and adaptation' might be a better descriptor. New instances of data can be run against the resulting model, enabling the instance to be categorised, and inferences drawn. A looser approach, referred to as 'unsupervised learning', relies heavily on the analytical tool not just to compute weightings on predefined factors but also to postulate elements of the model. It represents something approaching 'automated model development and adaptation'. Here is further discussion of the transition towards AI-based data analytics (s.2).
Earlier techniques in data analytics were algorithmic or procedural in nature, and hence the rationale underlying inferences could be extracted and explained. On the other hand, many of the most common AI/ML techniques that have been developed apply 6th-generation artificial neural networks (ANNs). As mentioned earlier, techniques of these kinds are incapable of generating humanly-understandable explanations. The combination of (a) the consolidation of data from disparate sources with diverse meanings and quality attributes with (b) opaque inferencing techniques, is a recipe for organisations to deceive themselves. This creates a high likelihood of an organisation making materially bad decisions, resulting in harm to themselves and to other affected parties, without the ability to investigate the sources of the errors. Here is further discussion of often-mistaken assumptions about data and inferencing processes (ss.4.2-4.3), of the opaqueness of much AI inferencing (s.4.4), of the failure to impose obligations on developers and users of AI (s.4.5), of the risks they give rise to (s.3), and of guidance for responsible use of those techniques (s.4).
Reference was made earlier to two relatively mature exemplars of AI: natural language processing (NLP) and natural language generation (NLG). During the 2020s, a new exemplar of AI burst into prominence, which combines with NLP and NLG to deliver new capabilities that are promising, exciting, and dangerous.
Generative AI is a general term for the production of synthetic content in any form, including image, video, sound and three-dimensional forms. For data in textual form, successful synthesis has been based on a recently-emerged technique refered to as large language models (LLMs). This may shortly be complemented by, and ultimately replaced by, small language models (SLMs). An LM implements a simplistic sense of grammatical structure by generating a likely next word in a sequence of words. The apparent authority of the generated text derives from a combination of:
Prominent text-oriented GenAI tools include ChatGPT, Microsoft's CoPilot front-end to the GPT LLM, Perplexity's interface to multiple LLMs, and the Chinese DeepSeek offering. Here is a simplified description of GenAI for text (s.2.1). Here is a longer-form description (s.2.2).
The smoothness and apparent authenticity of responses provided by these tools lulls even experienced professionals into suspending their disbelief and disengaging their critical faculties. Any calm analysis of the problems underlying the technology quickly identifies a raft of issues, such as:
Here is further discussion of potentially harmful attributes of GenAI (s.3), and here are some impacts, implications and risks that it is claimed GenAI gives rise to (Tables 5 and 6).
The interpretation of AI as currently practised includes a conditional fourth element requiring that the software implement actions that it has formulated. Computer-supported artefacts that embody actuators are commonly referred to as robots. The word 'robot' originated in Karel Capek's literary works at the end of World War I. Its usage since the 1940s implies both a machine with inherent computing capabilities, and a computer with sophisticated input/output devices. Isaac Asimov claims to have coined the term 'robotics' in 1942. He defined it in 1958, shortly after the notion of AI emerged, as "a science or art involving both artificial intelligence (to reason) and mechanical engineering (to perform physical acts suggested by reason)". Here is a review of the implications of the insightful literary device that is Asimov's 'Laws of Robotics'.
The robotics industry of the early 21st century concerns itself with robots used in secondary industry, services and medical applications. It claims a worldwide installed base of about 4 million robots, over 60% of them in the automotive and electrical / electronic industry sectors. The majority are in China, Korea and Japan, and European adoption exceeds that of the Americas. Most installed robots operate primarily autonomously, with a small percentage designed as 'cobots'. A small percentage are designed as 'cobots', signifying collaboration with human beings.
In some corners of the market, the application of some key principles are very evident. Robots can be particularly valuable where they do (a) things that humans can do, but that artefacts can do much more effectively or efficiently (e.g. speed and accuracy of computation and of movement), and (b) things that humans cannot do, or that have to be done in places where humans cannot go (e.g. control a Mars buggy in real time, enter Chernobyl or Fukishama, operate in the Mariana Trench). Here is further discussion of comparative robot advantage (s.4.1).
As indicated earlier, AI software can be implemented using any generation of software development tool. Robotics and AI have largely developed in parallel rather than intertwined, and it appears likely that the large majority of the installed base of industrial, services and medical robots does not use AI software at this stage. On the other hand, AI is significant in computer vision which is an important element of transport applications.
With rapid growth occurring in drones and 'self-driving' cars, it is important to consider the nature of artefact autonomy. There is widespread distaste about the possibility of some kinds of decisions being made by artefacts, while other, generally low-level and real-time functions are being delegated to software running in artefacts. For example, aspects of motor vehicle engine control and braking have been intermediated by pre-programmed electronics for several decades.
Moreover, the choice does not lie between control being exercised by either an autonomous machine or a self-sovereign human. There are in practice varying levels or degrees of autonomy. An artefact can be designed to act without reference to humans, or to act but inform a human supervisor, or to give notice to a human supervisor that an action will shortly be taken. The supervisor may have the power to suspend, cancel or interrupt an action, or to veto or override an imminent one. It is also possible to design a human-collaborative 'cobot' using decision support principles. In that case, the artefact, rather than taking action, instead provides recommendations, advises on options, or analyses options, and the human remains in control. Here is further discussion of degrees of autonomy in the context of drones and in artefacts generally
There is a world of robotic behaviour beyond industrial, transportation and logistics applications. Many lesser machines have actuators installed, such as an ATM's sliding tray for chip-cards. Commercial and administrative applications can be programmed to make decisions, and to directly trigger actions by devices that have a direct impact on the real world, perhaps as trivial as the despatch of an email, and the inclusion of an invoice or notice of a fine. Alternatively, it may be as significant as the blockage of a consumers' ability to make a payment, use public transport, or operate a motor vehicle. Affected parties, if they cannot discover a rational explanation for actions that have been taken, and if they can find no sentient being to deal with, face potentially serious constraints, and worrying times. Organisations' accountability for actions taken on their behalf is compromised. Here is a discussion of issues and regulatory options in the specific context of airborne drones (ss.4.4-4.5).
Robotics involves creating an artefact that can act in the real world, and then providing it with some kind of processing capability, perhaps at a level or of a kind that can be thought of as intelligence. The inverse approach is for a human to be provided with some artefactual support, whether in relation to actions in the real world or of a mental nature. Prostheses replace a human's physical capabilities, most commonly because they've been lost or materially reduced. Orthoses can be regarded as instead augmenting a human's physical capabilities. They may be external to the human body (e.g. walking sticks, decompression chambers, but also spacesuits, racing cars), or external but very adjacent (e.g. spectacles, conventional hearing aids, 'wooden legs', but also SCUBA equipment).
Two aspects are of particular relevance to this discussion. One is internal or implanted artefacts that replace lost or severely limited bodily function, usefully described as endo-prostheses (e.g. pacemakers, cochlear implants, implanted drug infusion devices) and endo-orthoses that enhance functionality (sprinters' artificial legs, racing wheelchairs). A second aspect is the (at this stage still only emergent) implantation of artefacts that replace or enhance mental rather than physical capabilities (e.g. rendering infra-red or ultra-violet wavelengths visible, or overlaying geographical contours over a view of a landscape or the visualisation of room-contents on the other side of an opaque wall).
Cyborgism, like robotics, is inherently linked to the fields of AI, action in the real and virtual worlds, and delegation of degrees of autonomy to artefacts. Here are segments offering further discussion of prostheses and orthoses (s.2.1), cyborgism generally (s.2.4), computationally-derived intelligence (s.3.2), prosthetes, orthots and cyborgs (s.2), cyborgism in the context of drones (s.5), and some of the implications of cyborgisation.
In the Introduction to this backgrounder, it was contended that 'artificial intelligence' embodies, and always has embodied, a serious and harmful misconception, and that a major shift in the discussion is long overdue. The first aspect of the misconception was argued to be that pursuing a 'grand challenge' to create machine intelligence that was like human intelligence was a valueless endeavour, because there is no shortage of human intelligence. The need is not for 'more of the same' intelligence, but for 'usefully different' intelligence.
A further problem is that the term 'artificial' carries with it unhelpful connotations of 'a copy of something real' / 'not as good as the real thing' / 'second-best'. A term such as 'artefactual intelligence' or 'artefact intelligence' avoids such overtones, and puts the focus on the purposes and outcomes of data processing performed by a manufactured thing.
Putting those ideas together, 'complementary artefact intelligence' has the following key attributes:
The term 'augmented intelligence' already exists, as a useful descriptor of the combination of human intelligence with artefact intelligence that has been designed to be complementary to human intelligence. These two notions provide a far more appropriate focal point for theory and practice than the hackneyed, ambiguous term 'artificial intelligence'.
On the other hand, the opportunity exists to link the three-part notion of augmented intelligence with the capacity for physical actions inherent in human effectors such as fingers and thumbs, and in artefacts' actuators. A human hand striking a hammerstone against a piece of flint long ago gave rise to a far more capable augmented person. Rather than conceiving of a robot as a standalone artefact, it can be thought of as a repository of artefactual capability that is complementary to human capability. The cobotics notion was originally an abbreviation of 'collaborative robotics'. A more rewarding focus would be 'complementary artefact capability', which when combined with human capability delivers 'augmented capability'.
Fuller discussion of complementary artefact capability is here (s.5.1), and of the full re-conception of the field is here (s.11).
The information infrastructure of 2025 is radically different from that at the turn of the century. The control that consumers had over their devices, their sofware and their data has been wrenched away from them, and into the hands of IT providers:
Onto this provider-dominated infrastructure, further features are being overlaid, as described in the main body of this Backgrounder. They include:
The intention of this backgrounder was to provide an accessible analysis of AI and associated technologies, as a basis for arguing for the abandonment of the old notion of AI. That enables the substitution of a cluster of linked terms that will stimulate a more appropriate design ethos for the technologies and their application.
Feedback on the presentation, the arguments and their accessibility would be greatly appreciated, to Roger.Clarke@xamax.com.au
See the annotated index of my collection of about 30 papers on AI and related topics, mostly in refereed journals, which have gained over 3,000 Google citations to date.
Roger Clarke is Principal of Xamax Consultancy Pty Ltd, Canberra. He is also a Visiting Professorial Fellow associated with UNSW Law & Justice, and a Visiting Professor in Computing in the College of Systems & Society at the Australian National University.
| Personalia |
Photographs Presentations Videos |
Access Statistics |
 |
The content and infrastructure for these community service pages are provided by Roger Clarke through his consultancy company, Xamax. From the site's beginnings in August 1994 until February 2009, the infrastructure was provided by the Australian National University. During that time, the site accumulated close to 30 million hits. It had passed 80 million by the end of 2025. Sponsored by the Gallery, Bunhybee Grasslands, the extended Clarke Family, Knights of the Spatchcock and their drummer, and her Dr Nurse site |
Xamax Consultancy Pty Ltd ACN: 002 360 456 78 Sidaway St, Chapman ACT 2611 AUSTRALIA Tel: +61 2 6288 6916 |
Created: 16 September 2025 - Last Amended: 3 October 2025 by Roger Clarke
This document is at www.rogerclarke.com/EC/AIBB.html
Mail to Webmaster - © Xamax Consultancy Pty Ltd, 1995-2026 - Privacy Policy