Roger Clarke's Web-Site

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

Roger Clarke's 'Principles for Responsible AI'

Principles for Responsible AI

Review Draft of 18 August 2019

Roger Clarke **

© Xamax Consultancy Pty Ltd, 2019

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

This document is at


Considerable public disquiet exists about Artificial Intelligence. Some concerns are ill-informed, some speculative, and some well-grounded. Public opposition could result not only in justified constraints on AI but also in kneejerk regulatory measures that create unnecessary obstacles to desirable forms and applications of AI. A set of 50 Principles for Responsible AI has been consolidated from a suite of proposals published by highly diverse organisation. The 50 Principles offer guidance for practitioners in the various phases of the AI supply chain.


1. Introduction

Artificial Intelligence (AI) excites a lot of people. And hyperbole about AI upsets a lot of other people. Engineers and computer scientists are busy sifting through the chaff to find the wheat, seeking to apply good ideas in order to achieve benefits without doing disproportionate harm. The hype and the hubbub are unhelpful distractions.

Of the concerns felt by the public, expressed on their behalf by advocacy organisations and amplified by the media, many are vague, and some are unfounded. In an analysis published in Clarke (2019a), I concluded that the core of the problem is that "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". That needs to be unpacked a little further, to identify the key factors.

The first issue is that artefacts (i.e. human-made objects) are increasingly sensor-computer-actuator bundles. They are becoming capable not only of computing and communicating, but also of taking autonomous action that directly affects the real world. So AI can directly affect people, without any humans intermediating in the process. Two further factors are that applications of AI make assumptions about the suitability of the inferencing process and about the data on which it depends. Those assumptions may not be adequately tested and may be faulty, perhaps generally, or perhaps only in specific contexts of use.

A further, key problem is that inferencing processes used in AI are typically opaque - even to the technology's originators and the designers of AI-based artefacts. This means that systems behave a-rationally, in the sense that there is no human-understandable explanation of how inferences were drawn. In such circumstances, processes cannot satisfy the requirements of replicability, auditability, and correctability. Combined with the inadequacies of product liability law, this means that organisations can escape accountability, i.e. the buck stops in the middle of nowhere. Given that the world appears to be going through a power-shift towards artefacts, it's only natural that there is push-back from the public.

Some of these problems are outside the scope of engineers and computer scientists; but others fall squarely within their responsibility. What guidance is available to inventors, designers and implementors about the negative impacts that need to be prevented or mitigated, and the risks that need to be managed?

2. A Consolidated Set of Principles

There's a very healthy marketplace for AI principles. Large IT companies and industry associations have felt the pressure of public concern about AI, and issued statements about how much care they're taking. Advocacy organisations have poked holes in those statements, and responded with much more substantial lists of what they think is needed. Recently, government agencies and regulators have clambered into the ring, most of them motivated by concern that innovation will be undermined if public acceptance of new technologies isn't assured, but some adopting consumer or human rights perspectives. Professional associations have pondered deeply (see, in particular, IEEE 2017), but continue to be very slow to deliver concise, actionable guidance for their members.

The purpose of this article is to present a set of 'Principles for Responsible AI'. I assembled the set on the basis of a suite of 30 such documents that had been previously published by a wide array of organisations. I then extracted the principles they contained, identified commonalities, and built a super-set that endeavours to encompass them all. Access is provided to the working papers, which lists and provides access to the documents, and shows the sources on which each Principle was based. This enables audits to be undertaken of the extraction process and the expressions chosen. Alternative interpretations can be made and alternative expressions can be formulated.

The resulting set of 50 Principles might have been redundant if any of the 30 sources already covered the field satisfactorily. The coverage proved to be generally very thin, however, with almost all of the sources containing less than half of the complete set, and many of them being very limited in scope. The only authoritative document that was found to have substantial coverage was that of the European Commission (EC 2019), and even that only scores 74%. The EC's document may, however, be more directly relevant than the 50 Principles presented here, for professionals whose focus is on goods and services specifically for EU countries, or for export, including to EU countries.

3. Applying the 50 Principles

Reading through the set of 50 Principles at the end of this article provides a basis for professionals to contemplate the impacts that their work may have. If a copy is pinned up on the wall, the Principles can be discussed within teams, and referred to when differences of opinion arise. This section suggests a few further aspects whereby they can be of use to engineers and computer scientists.

(1) Use them as principles, not as a recipe

A great deal of diversity is evident in AI technologies, in the artefacts that AI is built into, in the systems that incorporate the artefacts, and in the uses to which the systems are put. In such circumstances, it is infeasible to formulate a general recipe or an industry Standard that can be applied by technicians, far less a law. The appropriate form of guidance is a set of principles, which require interpretation and application by professionals.

An example of a general Principle that can't be expressed in a way that applies equally well in every situation is "Ensure human control over autonomous behaviour of AI-based technology, artefacts and systems" (3.2). A line needs to be drawn somewhere between real-time decision-making, such as the equilibration of a vehicle's steering on uneven surfaces [or the tuning of the fuel-mix delivered into an engine?], and more deliberative decisions, such as the guilt or innocence of a defendant. Similarly, "Implement safeguards to avoid, prevent and mitigate negative impacts and implications" (4.4) has to be operationalised in each particular context, because designs need to exhibit as much variety as the technologies, the artefacts and systems, and the circumstances in which they're applied.

(2) Manage conflict among the Principles and other interests

There are many conflicts between these people-protective Principles and the interests of other stakeholders. The organisations that sponsor the development of AI technology or AI-based artefacts or systems can usually protect their own interests, but many other organisations may be involved, and there are impacts on the economy, society as a whole, and the environment.

Conflicts even arise within the 50 Principles. For example, it can be difficult to both "Ensure ... performance of intended functions" (7.1) and "Ensure people's physical health and safety ('nonmaleficence')" (4.1). To invoke a classic in the field, in the very first of Asimov's short stories in his long series on the Laws of Robotics, 'Robbie the robot' caused (minor) harm to the robot's owner, because she was knocked breathless by the act of getting her out of the path of a tractor.

The resolution of conflicts involves value judgements. Some may fall within the designer's sphere of responsibility; but the big calls depend on executive judgement and consultation with stakeholders and regulators.

(3) Reflect where you are in the supply chain

The Principles need to be interpreted differently in different parts of the supply chain or industry network. A simple model of industry structure distinguishes research, invention, innovation, dissemination and application phases, with researchers producing AI technology, inventors using the ideas to produce AI-based artefacts, developers incorporating the artefacts within AI-based systems, purveyors of various kinds establishing an installed base of such systems, and user organisations applying them, resulting in impacts and implications.

The responsibilities of professionals playing roles higher up the supply-chain differ from the responsibilities of designers close to the applications end of the business. For example, "Deliver transparency and auditability" (7.) calls for purveyors and users to ensure understandability by affected stakeholders of each inference, decision and action that arises from AI-based systems. Higher up the chain, on the other hand, the Principle requires that inventors and innovators design transparency into AI technology, and ensure that developers and users of AI-based artefacts and systems can readily understand the nature of the underlying technology, and have the means available to them to devise ways of fulfilling their own transparency and auditability obligations.

(4) Reflect the particular form of AI

A further challenge arises from the clumping together of rather different things under the banner of 'AI'. Many of the Principles are of greater significance in relation to some forms of AI than others; and many would therefore be more directly useful to professionals if they were re-phrased, re-framed or customised to particular forms of AI.

For many people, the most obvious technological threat is robotics. Robotics is migrating well beyond factory floors and automated warehouses, including to aircraft and watercraft. The large majority of the Principles are directly applicable to designers of autonomous vehicles and their support systems, both in controlled environments such as mines, quarries and dedicated bus routes, and where the vehicle is on the open road or in public airspace. On the other hand, the Principle "Design as an aid, for augmentation, collaboration and inter-operability" (2.1) is readily understood in the context of high-order functions such as mission control, anticipation of changes in weather conditions and collision-risk detection; whereas considerable care is needed when it is applied to real-time control over a vehicle's attitude, position, course and collision-avoidance functions.

A second area usefully regarded as being within the AI field is cyborgisation, the process of enhancing individual humans by technological means, resulting in hybrids of a human and one or more artefacts. Many forms of cyborgisation fall outside the field of AI, such as spectacles, implanted lenses, stents, inert hip-replacements, SCUBA gear. Some enhancements qualify, however, by combining sensors and actuators with computational intelligence. Prominent examples include heart pacemakers (since 1958) and cochlear implants (since the 1960s, and commercially since 1978). Another is the kinds of replacement legs for above-knee amputees that contain software to sustain balance within the knee-joint. In this context, "Respect each person's autonomy, freedom of choice and right to self-determination" (3.4) is writ particularly large, i.e. free and informed consent is a pre-condition. On the other hand, "Ensure human review of inferences and decisions prior to action being taken" (3.5) requires careful interpretation, in order to balance the need for very prompt action against the need for individual self-determination. Another Principle requiring care in its application is "Avoid services being conditional on the acceptance of AI-based artefacts and systems" (3.7), which may be in direct conflict with the express desire of an applicant for a life-sustaining heart-pacemaker.

It might reasonably be expected that the medical implants field would provide leading exemplars of well-articulated guidance relevant to cyborgisation. However, studies conducted by investigative journalists and published in reputable media outlets have thrown considerable doubt on the effectiveness of the processes used in the field in protecting patients' interests and assuring quality and safety (ICIJ 2019).

Two other forms of AI to which the 50 Principles are applicable are the relatively established area of rule-based expert systems, and the still-immature field variously referred to as AI/ML or neural networking. Unlike predecessor approaches such as algorithmic / procedural programming, the rule-based expert systems field embodies no conception of either a problem or a solution. A rule-base merely describes a problem-domain. Techniques in the machine-learning field differ even more from earlier approaches, in that they do not necessarily begin with active and careful modelling of a real-world problem-solution, problem or even problem-domain. Rather than comprising a set of entities and relationships that mirrors the key elements and processes of a real-world system, a neural network model may be simply a list of input variables and a list of output variables (and, in the case of 'deep' networks, intermediary variables). The weightings assigned to each connection reflect the particular learning algorithm that was applied, and the characteristics of the training-set that was fed into it.

When these forms of AI are in play, a critical issue is how compliance can be achieved with the Principle "Ensure that people ... have access to humanly-understandable explanations of how [inferences, decisions and actions] came about" (6.3). Further challenges arise from the need to "Ensure that inferences are not drawn from data using invalid or unvalidated techniques" (7.6), "Test result validity ..." (7.7), and "Impose controls in order to ensure that the safeguards are in place and effective" (7.8). It is also unclear how neural networking techniques can enable performance of "audits of the justification, the proportionality, the transparency ..." (8.3). Given that these are all supported features of systems that AI/ML is seeking to supplant, it's problematical, not to mention morally dubious, for designers to argue that these Principles aren't applicable, merely because they're difficult to comply with.

The litany of problems arising with learning algorithms, training-sets, data accuracy, suitability and compatibility of data definitions, obscure and even non-existent explanations, lack of validation, lack of safeguards, and lack of controls to ensure that the safeguards are working, represents ample reason for public disquiet about when, how and even if AI/ML should be applied to human affairs.

This brief review has of course merely scratched the surface of the 50 Principles and their application to AI. It has, however, brought to light both the need for them to be not blindly but intelligently applied to the particular technology and in the particular context; and it has illustrated how public distrust of some AI technologies and AI-based artefacts and systems is not misguided, but grounded in reality.

(5) Apply the Principles within a Business Process

A set of principles is no more than a checklist. A further key element of the framework needed for assuring responsible AI is a process within which the Principles can be used to guide design. The Principles imply this by saying "conduct impact assessment ..." (1.4). Impact assessment is an approach that originated in environmental contexts. It adopts perspectives different from or additional to those of the organisation that drives the initiative. It is accordingly commonly conducted by other organisations, outside the boundaries of the driving organisation.

Within the driving organisation, the well-established tools of risk assessment and risk management might fill this role. However, their conventional forms are inadequate for the purpose. This is because they adopt the perspective only of the organisation that is sponsoring the activity. The process includes the identification of stakeholders, but their interests are reflected only to the extent that harm to them may result in material harm to the sponsoring organisation.

Responsible application of AI is only possible if stakeholder analysis is undertaken in order not only to identify the categories of entities that may be affected by the particular project, but also to gain insight into those entities' needs and interests. Note too that the notion of a 'stakeholder' goes beyond the users of and participants in the system. People may be affected by a system, and therefore have a stake in it, even if they are what are sometimes called 'usees', outside the system. This arises, for example, in the operation of credit-reporting agencies, tenancy and insurance claim data-pools, and criminal intelligence systems. Users' dependants, their local communities, population segments and whole economic regions can be usees as well. In the context of autonomous vehicles, stakeholders include passengers, occupants of other vehicles, pedestrians, vehicle-owners, transport-service-providers, individuals in occupations whose existence is threatened by the new approach (e.g. taxi-, courier- and truck-drivers), their dependants, their employers, and their unions.

Risk assessment processes need to be conducted from the perspectives of each stakeholder group, to complement that undertaken from the organisation's perspective. The various, and inevitably in part conflicting, information needs to be then integrated, in order to deliver a multi-stakeholder risk management plan (Clarke 2019b).

4. Conclusions

The 50 Principles that this article presents were derived by consolidating those proposed by a diverse assortment of organisations. The set therefore has at least some superficial validity as a proxy for what 'the public as a whole' thinks AI needs to look like. The brief analysis possible within this short article has shown ways in which the Principles can be applied, and can be fitted within a risk assessment process. It has also brought to the surface a few of the challenges that AI researchers, inventors, developers, purveyors and users must address in order to satisfy the needs of stakeholders, and overcome the concerns of the general public, and of legislators, regulators, financiers and insurers.

Once technologies and process standards have matured, the 50 Principles for Responsible AI will be redundant, replaced by more specific formulations for each category of what is currently termed 'AI' technology. However, given that the first 60 years of the AI field have seen so little progress in establishing such guidance, the set presented here may be 'as good as it gets' for some time to come.

Responsible AI Technologies, Artefacts, Systems and Applications
The 50 Principles

Also available is a 2-page PDF version, suitable for printing

1. Assess Positive and Negative Impacts and Implications

AI offers prospects of considerable benefits and disbenefits. All entities involved in creating and applying AI have obligations to assess its short-term impacts and longer-term implications, to demonstrate the achievability of the postulated benefits, to be proactive in relation to disbenefits, and to involve stakeholders in the process.

1.1 Conceive and design only after ensuring adequate understanding of purposes and contexts

1.2 Justify objectives

1.3 Demonstrate the achievability of postulated benefits

1.4 Conduct impact assessment, including risk assessment from all stakeholders' perspectives

1.5 Publish sufficient information to stakeholders to enable them to conduct their own assessments

1.6 Conduct consultation with stakeholders and enable their participation in design

1.7 Reflect stakeholders' justified concerns in the design

1.8 Justify negative impacts on individuals ('proportionality')

1.9 Consider alternative, less harmful ways of achieving the same objectives

2. Complement Humans

Considerable public disquiet exists in relation to the replacement of human decision-making by inhumane decision-making by AI-based artefacts and systems, and displacement of human workers by AI-based artefacts and systems.

2.1 Design as an aid, for augmentation, collaboration and inter-operability

2.2 Avoid design for replacement of people by independent artefacts or systems, except in circumstances in which those artefacts or systems are demonstrably more capable than people, and even then ensuring that the result is complementary to human capabilities

3. Ensure Human Control

Considerable public disquiet exists in relation to the prospect of humans being subject to obscure AI-based processes, and ceding power to AI-based artefacts and systems.

3.1 Ensure human control over AI-based technology, artefacts and systems

3.2 In particular, ensure human control over autonomous behaviour of AI-based technology, artefacts and systems

3.3 Respect people's expectations in relation to personal data protections, including their awareness of data-usage, their consent, data minimisation, public visibility and design consultation and participation, and the relationship between data-usage and the data's original purpose

3.4 Respect each person's autonomy, freedom of choice and right to self-determination

3.5 Ensure human review of inferences and decisions prior to action being taken

3.6 Avoid deception of humans

3.7 Avoid services being conditional on the acceptance of AI-based artefacts and systems

4. Ensure Human Safety and Wellbeing

All entities involved in creating and applying AI have obligations to provide safeguards for all human stakeholders, whether as users of AI-based artefacts and systems, or as usees affected by them, and to contribute to human stakeholders' wellbeing.

4.1 Ensure people's physical health and safety ('nonmaleficence')

4.2 Ensure people's psychological safety, by avoiding negative effects on their mental health, emotional state, inclusion in society, worth, and standing in comparison with other people

4.3 Contribute to people's wellbeing ('beneficence')

4.4 Implement safeguards to avoid, prevent and mitigate negative impacts and implications

4.5 Avoid violation of trust

4.6 Avoid the manipulation of vulnerable people, e.g. by taking advantage of individuals' tendencies to addictions such as gambling, and to letting pleasure overrule rationality

5. Ensure Consistency with Human Values and Human Rights

All entities involved in creating and applying AI have obligations to avoid, prevent and mitigate negative impacts on individuals, and to promote the interests of individuals.

5.1 Be just / fair / impartial, treat individuals equally, and avoid unfair discrimination and bias, not only where they are illegal, but also where they are materially inconsistent with public expectations

5.2 Ensure compliance with human rights laws

5.3 Avoid restrictions on, and promote, people's freedom of movement

5.4 Avoid interference with, and promote privacy, family, home or reputation

5.5 Avoid interference with, and promote, the rights of freedom of information, opinion and expression, of freedom of assembly, of freedom of association, of freedom to participate in public affairs, and of freedom to access public services

5.6 Where interference with human values or human rights is outweighed by other factors, ensure that the interference is no greater than is justified ('harm minimisation')

6. Deliver Transparency and Auditability

All entities have obligations in relation to due process and procedural fairness. These obligations can only be fulfilled if all entities involved in creating and applying AI ensure that humanly-understandable explanations are available to the people affected by AI-based inferences, decisions and actions.

6.1 Ensure that the fact that a process is AI-based is transparent to all stakeholders

6.2 Ensure that data provenance, and the means whereby inferences are drawn from it, decisions are made, and actions are taken, are logged and can be reconstructed

6.3 Ensure that people are aware of inferences, decisions and actions that affect them, and have access to humanly-understandable explanations of how they came about

7. Embed Quality Assurance

All entities involved in creating and applying AI have obligations in relation to the quality of business processes, products and outcomes.

7.1 Ensure effective, efficient and adaptive performance of intended functions

7.2 Ensure data quality and data relevance

7.3 Justify the use of data, commensurate with each data-item's sensitivity

7.4 Ensure security safeguards against inappropriate data access, modification and deletion, commensurate with its sensitivity

7.5 Deal fairly with people ('faithfulness', 'fidelity')

7.6 Ensure that inferences are not drawn from data using invalid or unvalidated techniques

7.7 Test result validity, and address the problems that are detected

7.8 Impose controls in order to ensure that the safeguards are in place and effective

7.9 Conduct audits of safeguards and controls

8. Exhibit Robustness and Resilience

All entities involved in creating and applying AI have obligations to ensure resistance to malfunctions (robustness) and recoverability when malfunctions occur (resilience), commensurate with the significance of the benefits, the data's sensitivity, and the potential for harm.

8.1 Deliver and sustain appropriate security safeguards against the risk of compromise of intended functions arising from both passive threats and active attacks, commensurate with the significance of the benefits and the potential to cause harm

8.2 Deliver and sustain appropriate security safeguards against the risk of inappropriate data access, modification and deletion, arising from both passive threats and active attacks, commensurate with the data's sensitivity

8.3 Conduct audits of the justification, the proportionality, the transparency, and the harm avoidance, prevention and mitigation measures and controls

8.4 Ensure resilience, in the sense of prompt and effective recovery from incidents

9. Ensure Accountability for Obligations

All entities involved in creating and applying AI have obligations in relation to due process and procedural fairness. The obligations include the entity ensuring that it is discoverable, and addressing problems as they arise.

9.1 Ensure that the responsible entity is apparent or can be readily discovered by any party

9.2 Ensure that effective remedies exist, in the form of complaints processes, appeals processes, and redress where harmful errors have occurred

10. Enforce, and Accept Enforcement of, Liabilities and Sanctions

Each entity's obligations in relation to due process and procedural fairness include the implementation of systematic problem-handling processes, and respect for and compliance with external problem-handling processes.

10.1 Ensure that complaints, appeals and redress processes operate effectively

10.2 Comply with external complaints, appeals and redress processes and outcomes, including, in particular, provision of timely, accurate and complete information relevant to cases


The two articles by the author provide access to c.70 references.

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

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

EC (2019) 'Ethics Guidelines for Trustworthy AI' High-Level Expert Group on Artificial Intelligence, European Commission, April 2019, at

ICIJ (2019) 'Implant Files' International Consortium of Investigative Journalists, 2019, at

IEEE (2017) 'Ethically Aligned Design: The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems', Version 2. IEEE, December 2017. at

Supporting Materials

The first two documents identify, extract and cite the 30 source documents, and the third relates the 30 source documents to the consolidated set of 50 Principles:

Ethical Analysis and Information Technology (8 Source Documents)

Principles for AI: A SourceBook' (22 Source Documents)

The 50 Principles Cross-Referenced to the Source-Documents


This paper has benefited from feedback from multiple colleagues, and particularly Prof. Graham Greenleaf and Kayleen Manwaring of UNSW, Robin Eckermann, and Peter Leonard.

Author Affiliations

Roger Clarke is Principal of Xamax Consultancy Pty Ltd, Canberra. He is also a Visiting Professor in Cyberspace Law & Policy at the University of N.S.W., 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: 1 August 2019 - Last Amended: 18 August 2019 by Roger Clarke - Site Last Verified: 15 February 2009
This document is at
Mail to Webmaster   -    © Xamax Consultancy Pty Ltd, 1995-2022   -    Privacy Policy