Roger Clarke's Web-Site

© Xamax Consultancy Pty Ltd,  1995-2024
Photo of Roger Clarke

Roger Clarke's 'The Re-Conception of IA'

The Re-Conception of AI:
Beyond Artificial, and Beyond Intelligence

Version of 2 January 2023

Published in IEEE Transactions on Technology & Society 4,1 (March 2023) 24-33

Roger Clarke **

© Xamax Consultancy Pty Ltd, 2022-03

Available under an AEShareNet Free
for Education licence or a Creative Commons 'Some
Rights Reserved' licence.

This document is at

The slide-set presented at ANU on 13 July 2023 is at

The slide-set presented at UNSW AI Institute on 29 August 2023 is at


The original conception of artificial intelligence (old-AI) was as a simulation of human intelligence. That has proven to be an ill-judged quest. It has led too many researchers repetitively down too many blind alleys, and embodies many threats to individuals, societies and economies. To increase value and reduce harm, it is necessary to re-conceptualise the field.

A review is undertaken of old-AI's flavours, operational definitions and important exemplars. The heart of the problem is argued to be an inappropriate focus on achieving substitution for human intelligence, either by replicating it in silicon or by inventing something functionally equivalent to it. Humankind instead needs its artefacts to deliver intellectual value different from human intelligence. By devising complementary artefact intelligence (CAI), and combining it with human intelligence, the mission becomes the delivery of augmented intelligence (new-AI). These alternative conceptions can serve the needs of the human race far better than either human or artefact intelligence can alone.

The proposed re-conception goes a step further. Inferencing and decision-making lay the foundations for action. Old-AI has tended to compartmentalise discussion, with robotics considered as though it were a parallel or at best overlapping field of endeavour. Combining the intellectual with the physical leads to broader conceptions of far greater value: complementary artefact capability (CAC) and augmented capability (AC). These enable the re-orientation of research to avoid dead-ends and misdirected designs, and deliver techniques that serve real-world needs and amplify humankind's capacity for responsible innovation.


1. Introduction

As a theme for a research project in the mid-1950s, 'Artificial Intelligence' (AI) was a brilliant idea. As a unifying concept for a field of endeavour, on the other hand, it has always been unsatisfactory. For decades, its delivery has fallen short of its promises. The recent phase of insufficiently sceptical conception, development and application of AI systems is coupling overblown expectations with collateral damage to people caught up in the maelstrom. The author's contention is that the original notion of AI is a mis-conception of both the need and the opportunity, resulting in mis-targeted research and mis-shapen applications. This article proposes alternative conceptions that have the capacity to bring technologies back under control, to greatly reduce their potential for harm, and to create new opportunities for constructive designs that can benefit society, the economy, and investors. The argument is pursued in several complementary ways. The article commences by revisiting the original conception of AI, and the notion's drift over time. This is followed by a review of various instantiations of systems within which AI has been embodied, and an outline of how the shift from procedural programming languages, via rule-based systems, to forms of machine-learning, has undermined decision-rationale.

After acknowledging AI's potentials, attention is turned to the generic threats that it embodies and the prospects of seriously negative impacts on individuals, the economy and societies, and hence organisations. Unsurprisingly, this gives rise to a great deal of public concern. Regrettably, the reaction of AI proponents and government policy agencies has not been to address the problems and impose legal safeguards, but instead to mount charm offensives and propose complex but ultimately vacuous ethical guidelines and soft law.

Building on that view of 'old-AI', alternative conceptions are outlined. The first, dubbed here 'new-AI', has a similar field of view, but avoids the existing and critical mis-orientation. The second conception, referred to as 'AC', is broader in scope, and represents a more comprehensive approach to the utilisation of technologies.

The article's first contribution is the combination of existing knowledge into an integrated whole that provides a clear rationale for change. Its second contribition is a cohesive framework that replaces old-AI with new conceptions. The intention is to present a somewhat sweeping proposition within the constraints of a journal article of modest length and hence with economy in the threads of history and technology that are pursued, examples that are proferred, and citations that are provided.

2. The Original Conception of AI

At the time the idea of AI was launched, the notion was based on:

"the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it" (McCarthy et al. 1955)

"The hypothesis is that a physical symbol system [of a particular kind] has the necessary and sufficient means for general intelligent action" (Simon 1996, in many variants)

The reference-point was human intelligence, and the intention was to create artificial forms of it using a [computing] machine.

Within a mere 5 years, however, the conjecture developed into a fervid belief, e.g.:

"Within the very near future - much less than twenty-five years - we shall have the technical capability of substituting machines for any and all human functions in organisations. ... Duplicating the problem-solving and information-handling capabilities of the brain is not far off; it would be surprising if it were not accomplished within the next decade"(Simon 1960)

Herbert Simon (1916-2001) ceased pursuing such propositions over two decades ago; but successor evangelists have kept the belief alive, such as this prognostication of 2005:

"By the end of the 2020s [computers will have] intelligence indistinguishable to biological humans" (Kurzweil 2005, p.25)

3. Drift in the Conception of AI

AI has had an unhappy existence. It has blown hot and cold through summers and winters, as believers have multiplied, funders have provided resources, funders have become disappointed by the very limited extent to which promises were delivered, and many believers have faded away again. One major embarrassment was the failure of Japan's much-vaunted Fifth Generation Project, a decade-long failure that commenced in 1982 (Pollack 1992).

Many histories have been written, from multiple viewpoints (e.g. Russell & Norvig 2003, pp. 16-28, Boden 2016). This article adopts an abbreviated view of the bifurcation of the field. The original conception can be seen as representing a grand challenge. Many AI practitioners subsequently divorced themselves from that challenge. They refer to the original notion as 'artificial general intelligence' or 'strong AI', which "aspires to" replicate human intelligence. The alternative stance, now widely adopted, is that AI work is "inspired by" human intelligence (Boden 2016, (Lieto & Radicioni 2016).

The last decade has seen a prolonged summer for AI technologists and promoters, who have attracted many investors. The author contends that the current euphoria about AI is on the wane, and another hard AI winter lies ahead. With much enthusiasm in evidence, and many bold predictions about the promise of AI, such scepticism is unconventional, and needs to be justified. The first step in doing so is to review some of the many, tangled meanings that people attribute to the term Artificial Intelligence.

4. AI Defined

People working in AI have had successes. An important example is techniques for pattern recognition in a wide variety of contexts, including in aural signals representing sound, and images representing visual phenomena. Applications of these techniques have involved the engineering of products, leading to more grounded explanations of how to judge whether an artefact is or is not a form of AI.

The following is an interpretation of many attempts to express an operational definition of AI. It is a paraphrase of multiple sources, including (Albus 1991), Russell & Norvig 2003 and McCarthy 2007), and not a direct quotation from any single source:

Intelligence is exhibited by an artefact if it:

  1. evidences:
    (a) perception, and
    (b) cognition,
    of relevant aspects of its environment;
  2. has goals; and
  3. formulates actions towards the
    achievement of those goals;

but also (for some commentators at least):

  1. implements those actions.

5. AI Instantiations

A complementary approach to understanding what AI is and does, and what it is not and does not do, is to consider the kinds of artefacts that have been used in an endeavour to deliver it. The obvious category of artefact is computers. The notion goes back to the industrial era, in the first half of the 19th century, with Charles Babbage as engineer and Ada Lovelace as the world's first programmer (Fuegi & Francis 2003). The explosion in innovation associated with electronic digital computers commenced about 1940 and is still continuing.

Computers were intended for computation. However, capabilities necessary to support computation are capable of supporting processes that have other kinds of objectives. A program is generally written with a purpose in mind, thereby providing the artefact with something resembling goals -- attribute (2) of the definition presented above. Computers can be applied to assist in the formulation of actions -- attribute (3) of the definition. They can even be used to achieve a kind of cognition, if only in the limited sense of categorisation based on pattern-similarity -- attribute (1b). Adding sensors that gather or create data representing the device's surroundings provides at least a primitive form of perception of the real world -- attribute (1a). So it is not difficult to contrive a simple demonstration of uses of a computer that satisfy the shorter version of the operational definition of AI.

When a computer is extended by means of actuators of various kinds, providing it with the capacity to act on the real world, the resulting artefact is both "a computer that does" and "a machine that computes", and hence what is commonly called a robot. The term was invented in the 1920s for a play (Capek 1923), and has been much used in science fiction, particularly by Isaac Asimov (1968, 1985). Industrial applications became increasingly effective in structured contexts from about 1960. Drones are flying robots, and submersible drones are swimming robots.

Other categories of artefact can run software that may satisfy the definition of AI. These include everyday things, such as bus-stops that feature display panels that impact people in their vicinity, projecting images selected or devised to sell something to someone. Their sensors may be cameras to detect and analyse faces, or detectors of personal devices that enable access to data about device-carriers' attributes.

Another category is vehicles. The author's car, a 1994 BMW M3, was an early instance of a vehicle whose fuel input, braking and suspension were supported by an electronic control unit. A quarter-century later, vehicles have far more layers of automation intruding into the driving experience. Driverless vehicles, of course, depend on software performing much more abstract functions than adapting fuel mix.

For many years, sci-fi novels and feature films have used as a stock character a robot designed to resemble humans, referred to as a humanoid. From time to time, it is re-discovered that greater rapport is achieved between a human and a device if the device evidences feminine characteristics. Whereas the notion 'humanoid' is gender-neutral, 'android' (male) and 'gynoid' (female) are gender-specific. The English translation of Capek's 1922 play used 'robotess' and more recent entertainments have popularised 'fembot'. This notion was central to an early and much-celebrated but also much-misunderstood AI, called Eliza (Weizenbaum 1966).

A promising, and challenging, category of relevance is cyborgs -- that is to say, humans whose abilities are augmented with an artefact, as simple as a walking-stick, or with simple electronics such as a health-condition alert mechanism, or a heart pacemaker. As more powerful and sophisticated computing facilities and software are installed, the prospect exists of AI integrated into a human, and guiding or even directing the human's effectors (Clarke 2005, 2011).

Various real-world exemplars can be readily argued to satisfy elements of the definition of AI. Particularly strong cases exist for industrial robots, and for driverless vehicles on rails, in dedicated bus-lanes and in mines. In the case of Mars buggies, the justification is not just economic but also functional. The signal-latency in Earth-with-Mars communications (about 20 minutes) precludes effective operation by an Earth-bound driver. Such examples satisfy the definitional criteria declared in section 4, because there is evidence of:

  1. (mechanistic forms of) both
    1. perception and
    2. cognition;
  2. goals (at least implicit / designed-in);
  3. (mostly pre-formulated) actions that work towards (mostly pre-programmed) goals; and
  4. actuators to implement those actions.

However, very charitable interpretation is needed to detect evidence of the second-order intellectual capacity that we associate with human intelligence (Dreyfus 1972, Weizenbaum 1976), such as:

This section has identified many flavours of AI applications. Some are in controllable environments in which engineered artefacts dominate; but others are in contexts that involve humans, human values and value-conflicts.

6. AI Techniques

Another approach to understanding AI is to consider the different ways in which the intellectual aspects of artefact behaviour are brought into being. During the first half-century of the information technology era, software was mostly developed using procedural or imperative languages. These involve the expression of a solution, which in turn requires a clear understanding (if not an explicit definition) of a problem (Clarke 1991). This approach uses genuinely-algorithmic languages, and the program comprises a sequence of well-defined instructions, including selection and iteration constructs.

Research within the AI field has variously delivered and co-opted a variety of software development techniques. One of particular relevance is rule-based expert systems. Importantly, a rule-set does not embody, and does not even recognise, a problem or a solution. Instead, it defines 'a problem-domain', a space within which cases arise. The idea of a problem is merely a perception by people who are applying some kind of value-system when they observe what goes on in the space.

Another relevant software development technique is AI/ML (machine-learning) and its most widely-discussed form, artificial neural networks (ANN). The original conception of AI referred to learning as a primary feature of (human) intelligence. The ANN approach, however, adopts a narrow focus on data that represents instances. There is not only no sense of problem or solution, but also no model of the problem-domain. There is just data, with meta-data provided or generated from within the data-set. It is implicitly assumed that the world can be satisfactorily understood, decisions can be made, and actions can be taken, on the basis of the similarity of a new case to the 'training set' of cases that have previously been reflected in the software. Other categories, such as Convolutional Neural Networks (CNN, Albawi et al. 2017) and Recurrent Neural Networks (RNN, Medsker & Jain 2001), appear to share those attributes with ANN.

Discussions of AI/ML frequently use the term 'algorithmic', as in the expression 'algorithmic bias'. This is misleading, because ANNs in particular are not algorithmic, but rather entirely empirical. ANNs are also not scientific in the sense of providing a coherent, systematic explanation of the behaviour of phenomena. Science requires theory about the world, directed empirical observation, and feedback to correct, refine or replace the theory. ANNs embody no theory about the world. The detachment between real-world needs and machine-internal processing, which emerged with rule-based systems, is pursued to completion by AI/ML. The absence of any reliable relationship with the real world underpins many of the issues identified below.

7. Areas of Specific Promise

So far, this review has considered the original conception of AI, a definition of AI, some key examples of AI embodied in artefacts, and some important AI techniques. Together, these provide a basis for consideration of AI's impacts.

The primary concern of this analysis is with downsides. There is no shortage of wide-eyed optimism about what AI might do for humankind -- although many claims are nothing more than vague marketing-speak. The more credible arguments relate to:

8. The Generic Threats Inherent in AI

AI embodies threats to human interests. There have been many expressions of serious concern about AI, including by theoretical physicist Stephen Hawking (Cellan-Jones 2014), Microsoft billionaire Bill Gates (Mack 2015), and technology entrepreneur Elon Musk (Sulleyman 2017). However, critics seldom make clear quite what they mean by AI, and their concerns tend to be long lists with limited structure; so progress in understanding and addressing the problems has been glacially slow. Some useful sources include (Dreyfus 1972, Weizenbaum 1976, Scherer 2016 esp. pp. 362-373, Yampolskiy & Spellchecker 2016, Mueller 2016, Duursma 2018).

This author associates the negative impacts and implications of AI with 5 key features (Clarke 2019a). The first of these is Artefact Autonomy. We are prepared to delegate to cars and aircraft straightforward, technical decisions about fuel mix and gentle adjustments of flight attitude. On the other hand, we are, and we need to be, much more careful about granting delegations to artefacts in relation to the more challenging categories of decision. In Table 1, I identify some key characteristics of challenging decisions, drawing in particular on Weizenbaum (1976) and Dreyfus (1972). Serious problems arise when authority is delegated to an artefact that lacks the capability to draw inferences, make decisions and take actions that are reliable, fair, or right, according to any of the stakeholders' value-systems.

Table 1: Characteristics of Challenging Decision-Categories


Where at least some autonomy is tenable, a framework is needed within which the degree of autonomy can be discussed and decided. Figure 1 draws on Armstrong (2010, p.14) and Sheridan & Verplank (1978, Table 8.2, pp.8.17-8.19), as interpreted by Robertson et al. (2019, Table 1).

Figure 1: Levels of Artefact Autonomy

Reproduced from Clarke (2014, Table 1, p.427)

Figure 1 distinguishes 6 levels of autonomy, 3 of which are of the nature of decision systems, and 3 of which are decision support systems. At levels 1-3, the human is in charge, although an artefact might unduly or inappropriately influence the human's decision. At levels 4-5, on the other hand, the artefact is primary, with the human having a window of opportunity to influence the outcomes. At level 6, it is not possible for the human to exercise control over the act performed by the artefact.

The second feature underlying concerns about AI is Inappropriate Assumptions about Data. There are many obvious data quality factors, including accuracy, precision, timeliness, completeness, the general relevance of each data-item, and the specific relevance of the particular content of each data-item (Wang & Strong 1996). Another consideration that is all too easy to overlook is the correspondence of the data with the real-world phenomena that the process assumes it to represent. That depends on appropriate identity association, attribute association and attribute signification (Clarke 2016).

It is common in data analytics generally, including in AI-based data analytics, to draw data from multiple sources. That brings in additional factors that can undermine the appropriateness of inferences arising from the analysis. Inconsistent definitions and quality-levels among the data-sources are seldom considered. The combination of the large numbers of problems with data-sets is one of the primary sources of what is commonly, if misleadingly, referred to as 'algorithmic bias' (Akter et al. 2021).

Data scrubbing (or cleansing) may be applied; but this is a dark art, and most techniques generate errors in the process of correcting other errors. Claims are made that, with sufficiently large volumes of data, the impacts of low-quality data, matching errors, and low scrubbing-quality automatically smooth themselves out. This may be a justifiable claim in specific circumstances, but in many cases it is a magical incantation that does not hold up under cross-examination (boyd & Crawford 2011).

Also far too common are Inappropriate Assumptions about the Inferencing Process that is applied to data. Each approach has its own characteristics, and is applicable to some contexts but not others. Despite the dangers inherent in mis-application of each particular technique, training courses and documentation seldom communicate much information about the limitations and risks. One issue is that data analytics practitioners frequently assume quite blindly that the data that they have available is suitable for the particular inferencing process they choose to apply. Data on nominal, ordinal and even cardinal scales is not suitable for the more powerful tools, because they require data on ratio scales. It is a potentially fatal flaw to assume that ordinal data (typically most / more / middle / less / least) is ratio-scale data. Mixed-mode data is particularly challenging. Meanwhile, the strategy adopted to deal with missing values is alone sufficient to deliver spurious results; yet the choice is often implicit rather than a rational decision based on a risk assessment. For any significant decision, assurance is needed of the data's suitability for use with each particular inferencing process.

The fourth feature is the Opaqueness of Inferencing Processes. As discussed above, some techniques used in AI are merely empirical, rather than being scientifically-based. Given that artificial neural networks are not algorithmic, no rationale, in the sense of a procedural or rule-based explanation, can be provided for the inferences that have been drawn. Inferences from ANNs are a-rational, i.e. not supportable by a logical explanation. Courts, coroners and ombudsmen demand explanations for the actions taken by people and organisations. Unless a decision-process is transparent, the analysis cannot be replicated, the process cannot be subjected to audit by a third party, the errors in the design cannot be even discovered let alone corrected, and guilty parties can escape accountability for harm they cause. The legal principles of natural justice and procedural fairness are crucial to civil behaviour, but they are under serious threat from some forms of AI.

The dilution of accountability is closely associated with the fifth major feature of AI, which is Irresponsibility. The AI supply-chain runs from laboratory experiment, via Industrial R&D, to artefacts that embody the technology, to systems that incorporate the artefacts, and on to applications of those systems, and deployment in the field. Successively, researchers, inventors, innovators, purveyors, and users bear moral responsibility for disbenefits arising from AI. On the other hand, the laws of many countries do not impose legal responsibility. The emergent pseudo-regulatory regime in Europe for AI actually absolves some of the players from incurring liability for their part in harmful technological innovation (Clarke 2022a).

These five generic features together mean that AI that may be effective in controlled environments (such as factories, warehouses, thinly human-populated mining sites, and distant planets) faces far greater challenges in unstructured contexts with high variability and unpredictability (e.g. public roads, households, and human care applications). In the case of human-controlled aircraft, laws impose responsibilities for collision-detection capability, for collision-avoidance functionality, and for training in the rational thought-processes to be applied when the integrity of location information, vision, communications, power, fuel or the aircraft itself is compromised. However, devising and implementing such capabilities in mostly-autonomous drones is very challenging.

9. Broad Areas of Negative AI Impact

To get to grips with what AI means for people, organisations, economies and societies, the generic threats identified in the previous section need consideration in a wide variety of contexts. Many people have reacted in paranoid manner to compelling imagery in sci-fi novels and films such as a robot apocalypse, attack drones and cyborgs with runaway enhancements.

Dystopian treatment of obscure technologies has a long history. The notion that industrial processes could develop to the point where humankind has ceded control to 'the machine', is completely dependent on it, and has no idea what to do when the machine stops performing its no-longer-understood functions, dates to 1909 (Forster 1909). Concerns about self-replicating, intelligent machines were expressed four decades earlier (Butler 1872). The need is for categories and examples grounded in real-world experience, and carefully projected into plausible futures using scenario analysis.

Table 2 suggests some natural outcomes, where AI delivers on its promise. As these threads continue to develop, they are likely to mutually reinforce, resulting in the removal of self-determination and meaningfulness from at least some people's lives.

Table 2: Readily-Foreseen Negative Impacts of A.I.


This discussion has had its main focus on impacts on individuals, societies and the economy. Organisations are also themselves affected, both directly and indirectly. AI-derived decisions and actions that prove to have been unwise undermine the reputations of and trust in organisations that have deployed AI. That will have inevitably negative impacts on adoption, deployment success and return on investment.

10. How to Reap Benefits but Mitigate Harms and Manage Risks

The previous sections argue that AI, as currently conceived, and as practised in both research and applications contexts, suffers from internal inconsistencies, and embodies considerable threats. What might be done about the problem?

The value of artefacts has been noted in dull, dirty and dangerous work, and where they are demonstrably more effective than humans. In such circumstances, the human race has a century of experience in automation. Careful engineering design, testing and management of AI components within such decision systems is doubtless capable of delivering benefits. However, development processes for impactful products require far greater investment in multi-stakeholder risk assessment than is apparent to date (Clarke 2022b).

Beyond those contexts are many others in which it is crucial that decision-making be reserved for humans. These involve factors that arise in important real-world decision-making, including, as indicated in Table 1, complexity, uncertainty, ambiguity, variability, fluidity, value-content, multiple stakeholders and value-conflicts.

In all such circumstances, the use of empirically-based AI, incapable of providing an underlying rationale, is subject to important provisos. It must be avoided entirely, or limited to level 1 of Figure 1 (Analyse Options), or its outputs must be handled by sceptical and careful decision-makers, and subjected to safeguards and mitigation measures, and to controls to ensure the safeguards and mitigation measures are functioning as intended, with liability accepted by its proponents where safeguards fail.

A considerable literature has emerged concerned with 'responsible application of AI' (Mueller 2016, Turner 2019, Clarke 2019b). Nominally, "Soft law's role in governance ... is not meant to diminish the need for regulations, but rather be considered an interim solution ..." (Gutierrez et al. 2021, p.168). Instead, despite the enormous threats AI brings with it, governments everywhere have been prevailed upon by AI proponents to avoid formal law and instead rely on self-regulation using Principles, with much stronger emphasis on quelling public concerns than on meaningful obligations, sanctions and enforcement.

The so-called 'precautionary principle' (Wingspread 1998), expressed as an ethical norm, requires that, where an action is suspected of causing harm, and where scientific consensus that it is not harmful is lacking, the burden of proof falls on those taking the action. In some contexts, particularly environmental law, it is an enforceable requirement that "When human activities may lead to morally unacceptable harm that is scientifically plausible but uncertain, actions shall be taken to avoid or diminish that potential harm" (TvH 2006). The precautionary principle needs to be applied to AI and the behaviour of autonomous artefacts. At the very least, there is a need for Principles for Responsible AI to be not merely talked about, and not merely the subject of soft law, but to actually be subject to co-regulatory regimes, including enforcement (Clarke 2019c, 2022a, Calo 2021). In addition, dangerous techniques and applications need to be subject to moratoria, until an adequate regulatory regime is in place.

11. A Re-Conception of the Field

More than enough experience has been gathered about Old-AI approaches, and more than enough evidence exists of their deficiencies, and the harm that they cause. The world needs an alternative way of conceiving and implementing 'smart' systems, to enable both the achievement of positive outcomes and the prevention, control and mitigation of harmful impacts. This final section draws on the above analysis, provides a rationale underlying re-conception of the field, and then articulates a set of ideas that the author contends can deliver what the world needs.

A starting-point is recognition of the fact that there are already 8 billion humans. So society gains very little by creating artefacts that think like humans. The idea of 'artificial intelligence' was misdirected, and has resulted in a great deal of wasted effort during the last 70 years. We don't want artificial; we want real. We don't want more intelligence of a human kind; we want artefacts to contribute to our intellectual endeavours. Useful intelligence in an artefact, rather than being like human intelligence, needs to be not like it, and instead constructively different from it.

One useful redefinition of AI would be as 'Artefact Intelligence' (Clarke 2019a, pp.429-431). This is not a new term. For example, it was defined in Takeda et al. (2002), as "intelligence that maximizes functionality of the artifact", with the challenge being "to establish intentional/physical relationship between humans and artifacts" (pp.1, 2). A similar term, 'artefactual intelligence', was described by (De Leon 2003) as "a measure of fit among instruments, persons, and procedures taken together as an operational system".

The valuable form is usefully called 'Complementary Artefact Intelligence' (CAC). I proposed the following key attributes in 2019 as (Clarke 2019a, p.430):

  1. Effective performance of intellectual functions that humans do poorly or not at all;
  2. Performance of those intellectual functions within systems that include both humans and artefacts; and
  3. Effective, efficient and adaptable interfacing with both humans and other artefacts.

The idea was further articulated by Schneiderman in this journal a year later (Schneiderman 2020, pp.116-118), using the expressions "shift from emulating humans to empowering people" and "extend abilities, empower users, enhance human performance". In 2021, Schneiderman added the epithet "humans in the group; computers in the loop" Schneiderman 2021, p.58]. His subsequent book has expanded on the themes (Schneiderman 2022).

The focus on decision support systems rather than decision systems is also exemplified by an application of Zicari's 'Z-Inspection' process for 'Trustworthy AI' (Zicari et al. 2021): "The main goal of the system was to support (not replace) radiologists in assessing pneumonia severity for COVID-19 patients in all hospitalization phases" (Allahabadi et al. 2022, p.10, emphasis added).

However, we need to lift our ambitions higher than Artefact Intelligence. Our focus needs to be on devising Artefact Intelligence so that it combines with Human Intelligence to deliver something new that is superior to each of them. The appropriate term for this construct is 'Augmented Intelligence' (New-AI).

The idea of augmented intelligence has a long and to some extent cumulative history. Ashby (1956) proposed the notion of "intelligence amplification", and Engelbart (1962) was concerned with "augmenting human intellect". More recently, (Zheng et al. 2017) provides some articulation of the process using the cumbersome term "Human-in-the-loop hybrid-augmented intelligence".

A trade press article has depicted augmented intelligence as "an alternative conceptualization of AI that focuses on its assistive role in advancing human capabilities" (Araya 2019). The following working definition is proposed:

Augmented Intelligence (New-AI) is the integration of Human Intelligence and Complementary Artefact Intelligence into a whole that is different from, and superior to, either working alone

The IEEE Digital Reality Initiative has referred to 'augmented intelligence' as using machine learning and predictive analytics, not to replace human intelligence, but to enhance it (IEEE-DR 2019). Computers are regarded as tools for mind extension in much the same way as tools are extensions of the body, or as extensions of the human capability of action. However, by describing the human-computer relationship using the biological term 'symbiotic', that paper risks artefacts being treated as equals with humans rather than conceiving them as being usefully different from and complementary to people. A further issue is that the paper expressly adopts the flighty transhumanism and posthumanism notions, which postulate that a transition will occur to a new species driven by technology rather than genetics.

A related line of thinking is in Wang et al. (2021), which provides a summation of another segment of the IEEE initiative. The authors of that paper propose that the appropriate focus is Symbiotic Autonomous Systems (SAS): "advanced intelligent and cognitive systems embodied by computational intelligence in order to facilitate collective intelligence among human-machine interactions in a hybrid society" (p.10). This approach is reductionist. Key features of the sociotechnical system perspective is that organisations comprise people using technology, that each affects the other, and that effective design depends on integration of the two (Abbas and Michael 2022). SAS stands at least in contrast with the notion of socio-technical systems, and even in conflict with it. SAS is framed in a manner that acknowledges artefacts as at least equal members in a "hybrid society", and arguably the superior partner. The sci-fi of Isaac Asimov and Arthur C. Clarke anticipated the gradual ceding of power by humans to artefacts; but it is far from clear that contemporary humans are ready to do that, and far from clear that technology is even capable of taking on that responsibility, let alone ready to do so.

Sci-fi-originated and somewhat meta-physical, even mystical, ideas deflect attention away from the key issues. They do not deliver what is needed to replace old-AI. We need to focus on artefacts as tools to support humans. Humans of the 21st century want Humans+, but by federating with artefacts rather than by uniting with them, or by becoming them, or by them becoming us.

A further term related to augmented intelligence can be detected in academic literature as far back as the 1970s, with increasing frequency from about 2010, but relatively small citation-counts. In Old-AI thinking, 'hybrid intelligence' refers to human intelligence serving the needs of artificial intelligence (Kamar 2016). Akata et al., on the other hand, defined it in a 2020 article in a manner consistent with the approach proposed here, as "the combination of human and machine intelligence, augmenting human intellect and capabilities instead of replacing them and achieving goals that were unreachable by either humans or machines" (Akata et al. 2020, p.19).

Re-conception of the 'A' in AI from Artificial to Augmented is alone capable of delivering great benefits. It deters product designers and application developers from limiting their focus to artefacts. It demands underlying techniques that support designs for socio-technical systems, within which artefacts must embody intellectual capabilities that are distinct from those of people, and complementary to them (Abbas et al. 2021). Attention must also be paid to the effectiveness and efficiency of interactions between the people and the artefacts.

The risk remains that designers will think only of the humans who directly interact with the artefact ('users'), and will overlook the interests of others who are affected by the operation of the socio-technical system ('usees' -- Clarke 1992, Fischer-Huebner &Lindskog 2001, Baumer 2015). Nonetheless, re-discovery of technology-in-use as the proper focus of design is welcome.

A further step remains, that extends beyond the intellectual realm. Inferences and decisions give rise to actions. Humans act through effectors like arms and fingers. Artefacts, in order to act in the world, are designed to include actuators such as robotic arms that exert force on real-world phenomena. In the same way that Artefact Intelligence is most valuable when it is designed to complement Human Intelligence, Artefact Actuators can be developed with the aim in mind to complement Human Effectors, resulting in Augmented Actors comprising a compound of humans and machines.

The capability of action arises from the combination of the intellectual and the physical. Human Capability arose not merely from the ability to oppose forefinger and thumb, but also from the intellectual realisation that something useful can be done with things held between the two body-parts. Similarly, artefacts' direct real-world impacts depend on the combination of Artefact Intelligence with suitable Actuators.

The final part of the re-conception is to prioritise Complementary Artefact Capability (CAC) that dovetails with Human Capability, resulting in a powerful form of synergy. By coordinating the intellectual plus the physical characteristics, of both humans and artefacts, we can achieve a combined capability of action superior to that which either human or artefact can achieve alone. This is usefully described as Augmented Capability (AC). Figure 2 provides a visual depiction of the proposed re-conception of the field.

Figure 2: The Proposed Re-Conception

12. Conclusions

The argument advanced in this article has been that the conception of old-AI is inappropriate and harmful. The question remains as to what role techniques associated with old-AI can play within the replacement conception of smart systems. To draw conclusions, each technique needs evaulation of its fit to the new contexts of Augmented Intelligence (new-AI), Complementary Artefact Intelligence (CAI), Complementary Artefact Capability (CAC) and Augmented Capability (AC).

I contend that the implications of this re-conception are profound, and extend across the entire value-chain from the laboratory to deployment in the field.

For real-world applications, the notion of humanlike or human-equivalent intelligence is a dangerous distraction. The focus must be on artefacts as tools, and on the delivery of purpose-designed artefact behaviour that dovetails with human behaviour to serve real-world objectives. Application designers must understand the available tools, and their characteristics, but very importantly, also the indications and contra-indications for their use. They must demand that tools come with clear statements of appropriate and inappropriate uses, and explanations of the conditions that are associated with, respectively, success and failure.

Further up the value-chain, providers of enabling facilities or middleware must appreciate the contexts in which developers will apply the available tools. They must critically evaluate ideas emerging from research laboratories. Only those ideas should be adopted that are consistent with the notions of complementary artefact intelligence and complementary artefact capability, and that are able to contribute to augmented intelligence and augmented capability.

The scope remains for academic research to continue its pursuit of the grand challenge of strong AI, emulating human intelligence at variously deep and functional levels. What must be avoided, however, is clinging to the mistaken notion that humanlike or human-equivalent intelligence has applicability to real-world problems.

A technique has merit if it is scientifically-based, blending theory about the real world with empirical insights; but if it is purely empirical in nature it must be viewed very sceptically and managed very carefully. If an artefact, or a human examining an artefact, can explain the rationale underlying the inferences it draws, and justify any decisions it takes or actions it performs, it has a role to play in socio-technical systems; otherwise, its fitness for purpose is seriously questionable.

If Artefact Intelligence replaces Human Intelligence, as part of a decision system, its applicability is limited to purely technical systems; whereas if it augments Human Intelligence, as part of a decision support system, it may also be a valuable contributor within socio-technical systems.

To the extent that a system includes actuators, if the Artefact Capability replaces Human Capability, it might be useful as part of a purely technical decision-and-action system; but its relevance to a system with substantial human or social elements is severely circumscribed. On the other hand, to the extent that the Artefact Capability complements Human Capability, the combined Augmented Capability represents a decision-and-action support system suitable for application to socio-technical systems.

This article has provided an interpretation of Artificial Intelligence (old-AI) together with an analysis of artefact autonomy, in order to propose a different conception of Augmented Intelligence (new-AI), and a shift in focus beyond even that, to a more comprehensive conception, Augmented Capability (AC).

Digital and intellectual technologies offer enormous potential. The author contends that the new-AI and Augmented Capability notions can quickly lead us towards appropriate ways to apply technologies while managing the similarly enormous threats that they harbour for individuals, economies and societies.

Reference List

Abbas R. & Michael K. (2022) 'Socio-Technical Theory: A review' In S. Papagiannidis (Ed), 'TheoryHub Book', TheoryHub, 2022, at

Abbas R., Michael K., Michael M.G., Perakslis C. & Pitt, J. (2022) 'Machine learning, convergence digitalization, and the concentration of power: Enslavement by design using techno-biological behaviors' IEEE Transactions on Technology and Society 3,2 (2022) 76-88, at

Abbas R., Pitt J. & Michael K. (2021) 'Socio-Technical Design for Public Interest Technology' Editorial, IEEE Trans. on Technology and Society 2,2 (June 2021) 55-61, at

Adensamer A, Gsenger R. & Klausner L.D. (2021) '"Computer says no": Algorithmic decision support and organisational responsibility' Journal of Responsible Technology 7--8 (October 2021) 100014, at

Akata Z. et al. (2020) 'A Research Agenda for Hybrid Intelligence: Augmenting Human Intellect With Collaborative, Adaptive, Responsible, and Explainable Artificial Intelligence' IEEE Computer 53,08 (August 2020) 19-28, at

Akter S., Dwivedi Y.K., Biswas K., Michael K., Bandara R.J. & Sajib S. (2021) 'Addressing Algorithmic Bias in AI-Driven Customer Management' Journal of Global Information Management 29, 6 (November-December 2021) 1-27, at

Albawi S., Mohammed T.A. & Al-Zawi S. (2017) 'Understanding of a convolutional neural network' Proc. Int'l Conf. on Engineering and Technology (ICET), 2017, 1--6, at Albawi/publication/319253577_Understanding_of_a_Convolutional_Neural_Network/lins/599ddb8045851574f4b35853/Understanding-of-a-Convolutional-Neural-Network

Albus J.S. (1991) 'Outline for a theory of intelligence' IEEE Trans Syst, Man Cybern 21, 3 (1991) 473--509, at pdf

Allahabadi H. et al. (2022) 'Assessing Trustworthy AI in times of COVID-19. Deep Learning for a multi-regional score conveying the degree of lung compromise in COVID-19 patients' IEEE Transactions on Technology and Society, Early Access, July 2022, at predicting stamp.jsp?tp=&arnumber=9845195

Araya D. (2019) '3 Things You Need To Know About Augmented Intelligence' Forbes Magazine, 22 January 2019, at

Armstrong A.J. (2010) `Development of a Methodology for Deriving Safety Metrics for UAV Operational Safety Performance Measurement' Report , Master of Science in Safety Critical Systems Engineering, Department of Computer Science, York University, January 2010, at

Asaro P.M. (2019) 'AI Ethics in Predictive Policing: From Models of Threat to an Ethics of Care' IEEE Technology and Society Magazine 38,2 (June 2019) 40-53, at

Ashby R. (1956) 'Design for an Intelligence-Amplifier' in Shannon C.E. & McCarthy J. (eds.) 'Automata Studies' Princeton University Press, 1956, pp. 215-234

Asimov I. (1968) 'I, Robot' (a collection of short stories originally published between 1940 and 1950), Grafton Books, London, 1968

Asimov I. (1985) 'Robots and Empire' Grafton Books, London, 1985

Baumer E.P.S. (2015) 'Usees' Proc. 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI'15), April 2015, at

Bennett Moses L. & Chan J. (2018) 'Algorithmic prediction in policing: assumptions, evaluation, and accountability' Policing and Society 28, 7 (2018) 806--822, at

Boden M. (2016) 'AI: Its Nature and Future' Oxford University Press, 2016

Borgesius F.J.Z. (2020) 'Strengthening legal protection against discrimination by algorithms and artificial intelligence' Intl J. of Human Rights 24, 10 (2020) 1572--1593, at

boyd D. & Crawford K. (2011) `Six Provocations for Big Data' Proc. Symposium on the Dynamics of the Internet and Society, September 2011, at

Butler S. (1872) 'Erewhon' Truebner & Co, 1872, in particular 'The Book of the Machines', pp. 190-221, at

Calo R. (2021) 'Artificial Intelligence and the Carousel of Soft Law' IEEE Transactions on Technology and Society 2, 4 (December 2021) 171-174

Capek K. (1923) 'R.U.R (Rossum's Universal Robots)' Doubleday Page and Company, 1923 (orig. published in Czech, 1918, 1921)

Cellan-Jones R. `Stephen Hawking warns artificial intelligence could end mankind' BBC News, 2 December 2014, at

Clarke R. (1991) 'A Contingency Approach to the Application Software Generations' Database 22, 3 (Summer 1991) 23-34, PrePrint at

Clarke R. (1992) 'Extra-Organisational Systems: A Challenge to the Software Engineering Paradigm' Proc. IFIP World Congress, Madrid, September 1992, PrePrint at

Clarke R. (2005) 'Human-Artefact Hybridisation: Forms and Consequences' Invited Presentation to the Ars Electronica 2005 Symposium on Hybrid - Living in Paradox, Linz, Austria, September 2005, PrePrint at

Clarke R. (2011) 'Cyborg Rights' IEEE Technology and Society 30, 3 (Fall 2011) 49-57, PrePrint at

Clarke R. (2014) 'Understanding the Drone Epidemic' Computer Law & Security Review 30, 3 (June 2014) 230-246, PrePrint at

Clarke R. (2016) 'Big Data, Big Risks' Information Systems Journal 26, 1 (January 2016) 77-90, PrePrint at

Clarke R. (2019a) 'Why the World Wants Controls over Artificial Intelligence' Computer Law & Security Review 35, 4 (Jul-Aug 2019) 423-433, PrePrint at

Clarke R. (2019b) 'Principles and Business Processes for Responsible AI' Computer Law & Security Review 35, 4 (Jul-Aug 2019) 410-422, PrePrint at

Clarke R. (2019c) 'Regulatory Alternatives for AI' Computer Law & Security Review 35, 4 (Jul-Aug 2019) 398-409, PrePrint at

Clarke R. (2022a) 'Responsible Application of Artificial Intelligence to Surveillance: What Prospects?' Information Polity 27, 2 (Jun 2022) 175-191, Special Issue on 'Questioning Modern Surveillance Technologies', PrePrint at

Clarke R. (2022b) 'Evaluating the Impact of Digital Interventions into Social Systems: How to Balance Stakeholder Interests' Working Paper for ISDF, Xamax Consultancy Pty Ltd, 1 June 2022, at

De Leon D. (2003) 'Artefactual Intelligence: The Development and Use of Cognitively Congenial Artefacts' Lund University Press, 2003

Dreyfus H.L. (1972) 'What Computers Can't Do' MIT Press, 1972; Revised edition as 'What Computers Still Can't Do', 1992

Duursma (2018) 'The Risks of Artificial Intelligence' Studio OverMorgen, May 2018, at

Engelbart D.C. (1962) 'Augmenting Human Intellect: A Conceptual Framework' SRI Summary Report AFOSR-3223, Stanford Research Institute, October 1962, at

Engstrom D.F., Ho D.E., Sharkey C.M. & Cue?llar N.-F. (2020) 'Government by Algorithm: Artificial Intelligence in Federal Administrative Agencies' NYU School of Law, Public Law Research Paper No. 20-54, April 2020, at

Fischer-Huebner S. & Lindskog H. (2001) 'Teaching Privacy-Enhancing Technologies' Proc. IFIP WG 11.8 2nd World Conference on Information Security Education, Perth, 2001, at

Forster E. M. (1909) 'The Machine Stops' Oxford and Cambridge Review, November 1909, at

Foster K.R., Vecchia P. & Repacholi M.H. (2000) 'Science and the precautionary principle Science 288, 5468 (2000) 979-981, at

Fuegi J. & Francis J. (2003) 'Lovelace & Babbage and the Creation of the 1843 'Notes'' IEEE Annals of the History of Computing 25, 4 (October--December 2003) 16-26, at

Gutierrez C.I., Marchant G.E. & Michael K. (2021) 'Effective and Trustworthy Implementation of AI Soft Law Governance' Editorial, IEEE Transactions on Technology and Society 2, 4 (December 2021) 168-170, at

IEEE-DR (2019) 'Symbiotic Autonomous Systems: White Paper III' IEEE Digital Reality, November 2019, at

Kamar E. (2016) 'Directions in Hybrid Intelligence: Complementing AI Systems with Human Intelligence' Proc. IJCAI, 2016, pp.4070--4073, at

Kurzweil R. (2005) 'The singularity is near' Viking Books, 2005

Lieto A. & Radicioni D.P. (2016) 'From Human to Artificial Cognition and Back: New Perspectives on Cognitively Inspired AI Systems' Cognitive Systems Research 39 (September 2016) 1-3, at

McCarthy J. (2007) 'What is artificial intelligence?' Department of Computer Science, Stanford University, 2007, at

McCarthy J., Minsky M.L., Rochester N. & Shannon C.E. (1955) 'A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence' Reprinted in AI Magazine 27, 4 (2006), at

Mack E. (2015) 'Bill Gates says you should worry about artificial intelligence' Forbes Magazine, 28 January 2015, at

Medsker L.R. & Jain L.C. (Eds.) (2001) 'Recurrent Neural Networks: Design and Applications' CRC Press, 2001

Mueller V.C. (ed.) (2016) 'Risks of general intelligence' CRC Press, 2016

Pollack A. (1992) ''Fifth Generation' Became Japan's Lost Generation' The New York Times, 5 June 1992, at

Robertson L.J., Abbas R., Alici G., Munoz A. & Michael K. (2019) 'Engineering-Based Design Methodology for Embedding Ethics in Autonomous Robots' Proc. IEEE 107, 3 (March 2019) 582-599, at

Russell S.J. & Norvig P. (2003) 'Artificial intelligence: a modern approach' 2nd edition, Prentice Hall, 2003, 3rd ed. 2009, 4th ed. 2020

Scherer M.U. (2016) 'Regulating Artificial Intelligence Systems: Risks, Challenges, Competencies, and Strategies' Harvard Journal of Law & Technology 29, 2 (Spring 2016) 354-400

Sheridan T.B. & Verplank W.L. (1978) 'Human and Computer Control for Undersea Teleoperators' MIT Press, 1978, at

Shneiderman B. (2020) 'Human-Centered Artificial Intelligence: Three Fresh Ideas' AIS Transactions on Human-Computer Interaction' 12,3 (2020) 109-124

Shneiderman B. (2021) 'Human-Centered AI' Issues in Science and Technology, Winter 2021, pp.56-61 at

Shneiderman B. (2022) 'Human-Centered AI' Oxford University Press, 2022

Simon H.A. (1960) 'The shape of automation' Reprinted in various forms, 1960, 1965, quoted in Weizenbaum J. (1976), pp. 244--245

Simon H. A. (1996) 'The sciences of the artificial' 3rd ed., MIT Press, 1996

Smith A. (2021) 'Face scanning and `social scoring' AI can have `catastrophic effects' on human rights, UN says' The Independent, 16 September 2021, at

Sulleyman A. (2017) 'Elon Musk: AI is a "fundamental existential risk for human-civilisation" and creators must slow down' The Independent, 17 July 2017, at

Takeda H., Terada K. & Kawamura T. (2002) 'Artifact intelligence: yet another approach for intelligent robots' Proc. 11th IEEE Int'l Wksp on Robot and Human Interactive Communication, September 2002, at

Turner J. (2019) 'Robot Rules: Regulating Artificial Intelligence' Palgrave Macmillan, 2019

TvH (2006) 'Telstra Corporation Limited v Hornsby Shire Council' NSWLEC 133 (24 March 2006), esp. paras. 113-183, at

Wang R.Y. & Strong D.M. (1996) 'Beyond Accuracy: What Data Quality Means to Data Consumers' Journal of Management Information Systems 12, 4 (Spring, 1996) 5-33, at

Wang Y. et al. (2021) 'On the Philosophical, Cognitive and Mathematical Foundations of Symbiotic Autonomous Systems (SAS)' Phil. Trans. Royal Society (A): Math, Phys & Engg Sci. 379(219x), August 2021, at

Weizenbaum J. (1966) 'ELIZA-a computer program for the study of natural language communication between man and machine' Commun. ACM 9, 1 (Jan 1966) 36--45, at

Weizenbaum J. (1976) 'Computer power and human reason' W.H. Freeman & Co., 1976

Yampolskiy R.V. & Spellchecker M.S. (2016) 'Artificial Intelligence Safety and Cybersecurity: a Timeline of AI Failures' arXiv, 2016, at

Zheng N., Liu Z., Ren P., Ma Y., Chen S., Yu S., Xue J., Chen B. & Wang F. (2017) 'Hybrid-augmented intelligence: collaboration and cognition' Frontiers of Information Technology & Electronic Engineering 18, 1 (2017) 153--179, at

Zicari R.V. et al. (2021) 'Z-Inspection: A Process to Assess Trustworthy AI' IEEE Transactions on Technology and Society 2, 2 (June 2021) 83-97, at


The original version of this paper was prepared in response to an invitation from Prof. Vladimir Mariano, Director of the Young Southeast Asia Leaders Initiative (YSEALI) of Fulbright University Vietnam, to deliver a Distinguished Lecture to YSEALI, broadcast around South-East Asia on 7 June 2022. My thanks to Prof. Mariano for the challenge to take "a long view of AI", its implications, and its directions. The paper further develops on my prior works in this area, which are cited. The presentation of the argument benefited from comments from two anonymous referees and the journal's editorial team.

Author Affiliations

Roger Clarke is Principal of Xamax Consultancy Pty Ltd, Canberra. He is also a Visiting Professor associated with the Allens Hub for Technology, Law and Innovation in UNSW Law, and a Visiting Professor in the Research School of Computer Science at the Australian National University.

xamaxsmall.gif missing
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 passed 65 million in early 2021.

Sponsored by the Gallery, Bunhybee Grasslands, the extended Clarke Family, Knights of the Spatchcock and their drummer
Xamax Consultancy Pty Ltd
ACN: 002 360 456
78 Sidaway St, Chapman ACT 2611 AUSTRALIA
Tel: +61 2 6288 6916

Created: 20 March 2022 - Last Amended: 26 January 2024 by Roger Clarke - Site Last Verified: 15 February 2009
This document is at
Mail to Webmaster   -    © Xamax Consultancy Pty Ltd, 1995-2022   -    Privacy Policy