Source-linked AI summary
Ethics as a service: a pragmatic operationalisation of AI Ethics
Jessica Morley, Anat Elhalal, Francesca Garcia, Libby Kinsey, Jakob Mokander, Luciano Floridi
TL;DR
Existing hard governance mechanisms and translational tools do not adequately protect against AI harms or operationalise AI ethics effectively. The paper explores compromises and a customisable Ethics as a Service approach, concluding that shifting focus toward procedural regularity may make AI ethics more relatable to practitioners, although AI impacts cannot be entirely controlled through technical design.
Problem
Existing hard governance mechanisms provide insufficient protection from AI harms, while a gap remains between AI ethics principles and the practical design of AI systems.
Method
The paper examines the compromises required for AI ethics tools and provides theoretical grounding for a customisable Ethics as a Service approach.
Results
Existing translational tools and methods fail to operationalise AI ethics effectively because they are generally too flexible or too strict.
Takeaways & Limitations
Shifting AI ethics away from principles toward procedural regularity may make it more relatable to AI practitioners and support pro-ethical design if the necessary balance is achieved.
Takeaways & Limitations
The impacts of AI systems cannot be entirely controlled through technical design.
Abstract
from arXiv · showhide
As the range of potential uses for Artificial Intelligence (AI), in particular machine learning (ML), has increased, so has awareness of the associated ethical issues. This increased awareness has led to the realisation that existing legislation and regulation provides insufficient protection to individuals, groups, society, and the environment from AI harms. In response to this realisation, there has been a proliferation of principle-based ethics codes, guidelines and frameworks. However, it has become increasingly clear that a significant gap exists between the theory of AI ethics principles and the practical design of AI systems. In previous work, we analysed whether it is possible to close this gap between the what and the how of AI ethics through the use of tools and methods designed to help AI developers, engineers, and designers translate principles into practice. We concluded that this method of closure is currently ineffective as almost all existing translational tools and methods are either too flexible (and thus vulnerable to ethics washing) or too strict (unresponsive to context). This raised the question: if, even with technical guidance, AI ethics is challenging to embed in the process of algorithmic design, is the entire pro-ethical design endeavour rendered futile? And, if no, then how can AI ethics be made useful for AI practitioners? This is the question we seek to address here by exploring why principles and technical translational tools are still needed even if they are limited, and how these limitations can be potentially overcome by providing theoretical grounding of a concept that has been termed Ethics as a Service.
3. Alan Turing Institute, London
This section identifies the concept as ‘Ethics as a Service’ and situates it among AI, machine learning, data ethics, applied ethics, and business ethics.
- ‘Ethics as a Service’ is the concept identified in this section.
- The listed subject areas include Artificial Intelligence and Machine Learning.
- The listed subject areas also include Data Ethics, Applied Ethics, and Business Ethics.
1. Introduction
AI ethics has generated principle-based governance documents, but a persistent gap remains between abstract principles and the practical design and deployment of AI systems. The paper examines why existing translational tools are insufficient and develops theoretical grounding for Ethics as a Service as a possible response.
- Existing legislation and regulation provide insufficient protection from AI harms and do not sufficiently incentivise socially preferable or environmentally sustainable AI design.The limitation concerns protection for individuals, groups, society, and the environment.
- Principle-based ethics codes, guidelines, and frameworks were developed as important governance mechanisms, but highly abstract principles provide little protection without practical design guidance.Practitioners need guidance on how to design and deploy algorithms within ethical boundaries.
- A significant gap exists between the theory and practice of AI ethics, motivating efforts to help practitioners translate ethical principles into design decisions.Previous work identified tools and methods intended to help developers, engineers, and designers understand both what to do and how to do it.
- Existing translational tools and methods fail to operationalise AI ethics effectively because they are generally either too flexible or too strict.Flexible approaches are vulnerable to ethics shopping and ethics washing, whereas strict approaches may fail to account for disagreement about how principles should be interpreted or applied.
- The paper asks whether pro-ethical design remains viable despite these limitations and how AI ethics can be made useful for practitioners.It explores why principles and technical tools remain needed and how their limitations might be overcome through Ethics as a Service.
- The paper provides theoretical grounding for Ethics as a Service and discusses lowering abstraction, translational-tool limits, practical compromises, and its theoretical underpinning.The concept is grounded in Habermas’s discourse ethics and Floridi’s distributed responsibility, distinguishing it from a technocratic auditing interpretation.
2. Lowering the level of abstraction
AI ethics guidance has proliferated, but abstract principles vary across contexts and remain difficult to translate into actionable design practices. Translational tools lower abstraction yet provide only a partial solution and remain vulnerable to manipulation.
- More than 160 AI ethics documents exist, covering principles and concepts including beneficence, justice, transparency, fairness, privacy, sustainability, dignity, and solidarity.
- Vague statements such as “AI systems may be discriminatory” encourage broad and generic rather than deep and specific responses.
- Ethical principles vary across traditions, cultures, ideologies, systems, countries, times, and societies, making their practical content context-dependent.
- Principles remain valuable as foundations for ethical practice rather than as detailed, directly actionable instructions.
- Translational tools and methods bring ethical guidance toward the design level by translating the “what” of AI ethics into the “how” of technical specifications.
- The authors conclude that lowering abstraction is at best partial: tools leave issues unresolved and can be manipulated by reprehensible actors.
3. Limits of Principlism and Translational Tools
The paper identifies translational tools as limited because they are extra-empirical, often diagnostic rather than prescriptive, vulnerable to practitioner-defined standards, and prone to one-off compliance use. These limitations can obscure responsibility and fail to ensure ethical outcomes over time.
- Because tools may be manipulated or foster false security and complacency, the paper asks how AI ethics can be usefully operationalised for practitioners.
- Many translational tools diagnose problems such as dataset bias but provide little practical guidance on how practitioners should overcome them.
- When practitioners set diagnostic parameters themselves, objective critique can be lost and tools may optimise practitioner criteria rather than social preferability.
- Translational tools do not by themselves identify sufficient interventions or clarify which decision-makers are responsible for correcting injustices.
- Tools are often treated as one-off compliance tests, encouraging tick-box ethics and allowing ethical commitments to disappear before deployment.
- Ethical implications should instead be evaluated during validation, verification, and evaluation, including after deployment and possible revision.
4. A series of compromises
The paper argues that AI ethics is not futile, but requires a compromise between excessive flexibility and strictness, alongside repeatable, context-sensitive, publicly reasoned processes rather than one-off compliance.
- Current pro-ethical design remains difficult because no “Goldilocks Level of Abstraction” has yet balanced flexibility and strictness.
- Algorithmic systems have non-deterministic impacts across human and non-human actors, including uses and purposes beyond those originally stated.
- Useful operationalisation should use a repeatable process without reducing ethical application to a one-off tick-box exercise.
- AI ethics should be framed as a reflective development process that reveals practitioners’ subjectivity and biases within particular circumstances.
- Structured identification and transparent communication of tradeoffs can support resolutions that are imperfect but publicly defensible.
- The proposed approach combines inclusive principles, repeatable contextual translation into technical standards, and oversight across validation, verification, and evaluation.
5. Outlining Ethics as a Service
Ethics as a Service adapts the Platform as a Service model to distribute ethical responsibility between an independent advisory board and internal AI practitioners. The proposal aims to balance overly flexible and overly strict governance while retaining contextual judgement and repeated oversight.
- Ethics as a Service, based on Platform as a Service, is proposed as a compromise between overly flexible and overly strict ethical governance.
- The model could distribute responsibility across an independent multidisciplinary ethics board, a collaboratively developed ethical code, and AI practitioners.
- The ethics board would develop principles through discussion and negotiation involving individuals, businesses, and environments affected by company systems.
- The board would define contextual meanings, provide selectable translational tools, conduct ethical foresight, and support remediation when systems breach principles.
- AI practitioners would apply contextually defined principles, select appropriate tools, document decisions publicly, and justify how the process was followed.
- The proposal may overcome many current limitations in theory, but its practical effectiveness remains untested and continuous audits impose financial and administrative costs.
- A progressive level of governance proportional to technology-and-context risk is proposed to balance audit requirements with innovation incentives.
- The authors encourage partial or complete pilots with public reporting of successes and failures to build a commons of best ethical practice.
6. Conclusion
The conclusion argues that AI ethics may benefit from a customisable, procedurally regular approach that balances flexibility with contextual responsiveness. It also stresses that technical design cannot fully control AI impacts, requiring post-deployment re-evaluation and further empirical research.
- Applied ethics combines law, governance policies, practices, procedures, and contextual support to balance strictness with flexibility.
- AI ethics may similarly benefit from a customisable approach that balances flexibility and centralisation with contextual needs.
- A procedural focus could make AI ethics more relatable to practitioners by clarifying its parallels with quality-assurance processes such as safety testing.
- Ethics as a Service is presented as a way to highlight the need for careful consideration of each design decision and to advance discussion of AI ethics practice.
- Technical design cannot entirely control AI impacts because complex AI systems interact unpredictably and non-linearly with other agents and systems.
- The effects of pro-ethical design may remain unknown until deployment, making regular re-evaluation and further qualitative and empirical research necessary.