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

By Reconceiving AI, We Can Manage Its Risks

Version of 25 July 2026

Roger Clarke **

© Xamax Consultancy Pty Ltd, 2025-26

This document is at http://rogerclarke.com/EC/RAIM.html

A great deal has been written about the threats inherent in various forms of Artificial Intelligence (AI). Most public attention is currently on machine learning techniques for drawing inferences (AI/ML) and Generative AI. The sources of threat in these forms of AI can be usefully grouped into five areas.

Two related areas are inappropriate assumptions about data, and about inferencing processes. Techniques for processing data depend on the input having particular attributes. Yet a great deal of the practice of AI involves the processing of data that was collected for one purpose, but is applied to another. Data is frequently expropriated, and applied by very different organisations, in very different contexts. Data is consolidated from multiple sources. Far too little attention is paid to the meanings of the data-items that are being relied upon, and to the (in)compatibility of data acquired from disparate sources. Far too little account is taken of the limited quality assurance that was applied to the data when it was collected, and the infeasibility of improving the quality of data whose provenance is distant and poorly understood.

The resulting data-mounds are then mangled to draw inferences of various kinds. Some techniques are based on models of parts of the real world, in the hope that the results will bear a near-enough relationship with that partial world. Other techniques have abandoned theory and the rational modelling that theory enables. So-called 'unsupervised learning' empowers artefacts to 'make of the data whatever they make of the data'. Reliance on correlation alone is being championed as being superior to the scientific approaches of the last five centuries, which interleave theories and models drawn from observations with directed experimentation and purposeful measurement of outcomes.

A third set of issues arises from the increasing opaqueness of AI techniques. The algorithmic approaches commonly adopted between 1960 and 2000 involve a problem-definition that is either express or inferrable from an express solution design. The rationale underlying any inference, any decision and any action is either readily available or can be reconstructed. Rule-based expert systems compromise that transparency, but do not destroy it. With contemporary, purely numerical approaches, on the other hand, there is no rationale. Neither the artefact nor its user can provide a humanly-understandable explanation.

Generative AI, of which ChatGPT was the pioneer, goes beyond drawing inferences from large volumes of data, and uses its data-holdings to produce new instances of data on request. Generative AI applied to textual data evidences 20 potentially harmful attributes. The technique embodies a simplistic form of syntactical analysis that purports to offer semantic content, but lacks any appreciation of content, context, audience or impact. This has resulted in it being described as 'a stochastic parrot'. It offers an at best very limited capability to be interrogated in relation to its sources and their meaning and relevance. It projects and amplifies existing biases, errors, planted information, rumours and 'fake news'. It generates plausible-sounding statements some of which are demonstrably wrong, or are nonsensical (commonly referred to as 'hallucinations'). The problem is compounded by the infeasibility of tracing back to the sources of the misrepresentations.

The fourth problem-area is the explosion in artefact autonomy that is accompanying the vogue management notion of 'digitalisation'. This features dependence on automated processing of data-mounds, and reduced staff-counts. Intrinsic to it is the abandonment of rationality. Artefacts are being delegated to infer, decide and act to an unprecedented degree. The arrogance of the Robodebt assault on welfare-recipients was the tip of the iceberg. Robodebt used no AI techniques, just (simplistic) algorithms. The rationale was pilloried in public at the end of 2016. Once the logic were exposed to the court, in late 2019, the $2 billion fraud collapsed. The next Robodebt scheme will use AI/ML and GenAI techniques. Those techniques do not and cannot provide rational explanations. With 'nothing to see here', the next Robodebt fraud is much more likely to survive challenge. Corporations, small business and the public are at dire risk of arbitrary decisions by governments; and small business and the public face the same threat from corporations.

The fifth group of risk-sources derives from the combination of opaqueness with automation. In the absence of humanly-understandable explanations for decisions, affected parties cannot submit counter-argument and evidence, and reviewers, tribunals and courts cannot resolve disputes. This undermines accountability. No party in the supply chain can be held liable for harm caused by AI-based inferencing, decisions and actions. This particularly affects the less powerful, but is also has enormous implications for regulators, courts, policy agencies and parliaments, and for risk management and the insurance industry.

Superficially, it seems good for business enterprises if they can escape liability, and easier for government agencies if they have no need to justify administrative decisions. On the other hand, such technology-driven laissez faire represents the loss of social licence, and would quickly result in the collapse of consumer and citizen loyalty, and of public confidence in its institutions. Law enforcement and military agencies, and their suppliers, contractors and strategic partners, are largely immune from the constraint of public opprobrium. On the other hand, a breakdown in public trust will be seriously harmful to the rest of the business sector, to the broader economy, and to society and polity.

That bleak picture emerges from unsceptical faith by organisations in AI technology. However, an AI-induced trust deficit is not inevitable. An alternative scenario can be built by recognising the unsuitability of the notion of 'Artificial Intelligence', and re-shaping the field of endeavour to align with the needs of people, organisations, society, economy and polity.

For the first 70 years, the focus has been on inventing intelligence that is 'similar to human intelligence, but not real'. What would actually benefit society is a form of intelligence designed into artefacts that is 'usefully different' from human intelligence. Such software would perform intellectual functions that humans do poorly or not at all, perform them within socio-technical systems that include both humans and artefacts, and interact effectively with both humans and other artefacts. A suitable descriptor is 'complementary artefactual intelligence'.

Claims about AI are couched in terms of competition with humans, substitution of humans, and what the entertainment world conceives as robot dominance and wild theorists call 'a technological singularity' that will replace homo sapiens with roboticus sapiens. Switching the focus from AI to 'complementary artefactual intelligence' lays the foundations for synergy between artefacts and humans. We can apply the principles of decision support systems in which artefacts are a tool, or perhaps an artefactual assistant. The notion of a fruitful combination of human and artefactual intellectual abilities orginated at the same time as AI, and has been brewing quietly. The term used for it has long been 'augmented intelligence' -- humans working smarter, with the support of intellectual tools.

This alternative, positive scenario has a further feature. The reconception of intellectual endeavours needs to extend to physical action. There are currently about 4 million robots in the world. Consumer appliance robots are proliferating, and many functions are performed in robot-like manner in trains and boats and planes, and cars. Each generally operates for periods of time largely independently of humans, exercising a significant delegation to perform tightly engineered actions within constrained contexts.

Within the category of 'robot', the term 'cobot' was coined to signify active artefacts designed for collaboration with a person. A human hand striking a hammerstone against a piece of flint long ago gave rise to a far more capable augmented person. Rather than a robot being conceived as a standalone artefact, it can be thought of as a repository of artefactual capability that is complementary to human capability. The cobotics notion is consistent with the idea of 'complementary artefactual capability'. When combined with human capability, it delivers 'augmented capability'.

As the excessive enthusiasm for current forms of AI begins to wane, we can do better than usher intellectual technologies back to an Arctic Winter for another quiet, dark decade or two. We can abandon the naive notion of AI, and thank it, but bury it. We can reorient technology towards complementary artificial intelligence and capability, embedded in intellectual and physical tools, co-working-by-design with humans. There is then no need to encroach on the entertainment industry's love affair with super-intelligences and the robot apocalypse. We can get on with improvements to human life and organisational effectiveness and efficiency.

Acknowledgements

This paper is preceded by an introductory piece on the various meanings of AI, at https:rogerclarke.com/EC/BAI.html, and supported by a fuller Opinion Piece, at https:rogerclarke.com/EC/AIOP.html, and an index at https:rogerclarke.com/EC/AIC.html. These provide access to underlying references numbering in the hundreds.

Author Affiliations

Roger Clarke is Principal of Xamax Consultancy Pty Ltd, Canberra. He is also a Visiting Professor in Computing in the College of Systems & Society at the Australian National University and a Visiting Professorial Fellow associated with UNSW Law & Justice.

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: 2 November 2025 - Last Amended: 25 July 2026 by Roger Clarke
This document is at www.rogerclarke.com/EC/RAIM.html
Mail to Webmaster   -    © Xamax Consultancy Pty Ltd, 1995-2026   -    Privacy Policy