The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions
Open accessThe DOI matches its Crossref record; 21 of the 22 references it cites were found in a registry or an index; 4 of the 13 sources its claims cite were checked for support.
Abstract
The last few years have seen a proliferation of principles for AI ethics. There is substantial overlap between different sets of principles, with widespread agreement that AI should be used for the common good, should not be used to harm people or undermine their rights, and should respect widely held values such as fairness, privacy, and autonomy. While articulating and agreeing on principles is important, it is only a starting point. Drawing on comparisons with the field of bioethics, we highlight some of the limitations of principles: in particular, they are often too broad and high-level to guide ethics in practice. We suggest that an important next step for the field of AI ethics is to focus on exploring the tensions that inevitably arise as we try to implement principles in practice. By explicitly recognising these tensions we can begin to make decisions about how they should be resolved in specific cases, and develop frameworks and guidelines for AI ethics that are rigorous and practically relevant. We discuss some different specific ways that tensions arise in AI ethics, and what processes might be needed to resolve them.
Overview
CoChat Summary
TL;DR
AI ethics has produced many overlapping high-level principles, but these principles are often too broad, ambiguous, and conflicting to guide practice. The authors argue the field should next focus on identifying and resolving the tensions that arise when principles are applied to concrete cases.
Key findings
- Principles for AI ethics have proliferated and substantially overlap: examples include the Asilomar AI principles signed by several thousand researchers, the ACM’s seven principles for algorithmic transparency and accountability, the UK House of Lords’ five principles, and Google’s AI ethics principles; many sets can be synthesised into five key principles—autonomy, beneficence, non-maleficence, justice, and explicability.
- Principles alone are limited: they are often ambiguous, highly general, and come into conflict in practice. For example, the UK House of Lords’ statement that AI with substantial impact on an individual’s life must generate a full and satisfactory explanation masks a tension between beneficence and explicability, since accurate diagnostic algorithms may not provide such explanations.
- The authors identify four key tensions: (1) using data to improve services vs. respecting privacy and autonomy; (2) using algorithms for more accurate decisions vs. ensuring fair and equal treatment; (3) increased personalisation vs. solidarity and citizenship; and (4) automation for convenience and empowerment vs. self-actualisation and dignity.
- Resolving tensions may require political trade-off judgements and public engagement where strict trade-offs exist, or further research and governance solutions where tensions are more contingent; the authors also note a tension between short-term implementation and long-term research.
Methods
- Design: conceptual/position paper in practical AI ethics, drawing comparisons with bioethics and reviewing the emergence and limits of AI ethics principles.
- Data or sample: no empirical dataset or systematic sample; the paper uses illustrative examples from AI ethics, industry, policy, and bioethics literature.
- Main analysis: argumentative synthesis proposing that the field focus on tensions, outlining why this is beneficial, identifying four central tensions, and discussing processes for resolving them.
Limitations
- Stated: the authors say there is unlikely to be an exhaustive list of all possible tensions, that most tensions cannot be resolved straightforwardly, and that it is unclear whether convergence on principles reflects deep consensus or merely shallow consensus.
- Evident: the paper is conceptual and does not present empirical data, a systematic review, or a formal framework for measuring or resolving tensions; its examples are illustrative rather than tested.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 1 Not valid · 7 Not checked · 40 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“While everyone might agree in principle that 'fairness' is important, there exist deep political disagreements about what exactly constitutes fairness [5].”
§ The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions > 3.1 Different Groups May Interpret Principles Differently
What the sources say
- Supportedcites [5] 5. Reuben Binns. 2017. Fairness in machine learning: Lessons from political philos- ophy. arXiv preprint arXiv:1712.03586 (2017). source
The Introduction (p. 2) states that 'fairness' in fair ML is 'best understood as a placeholder term for a variety of normative egalitarian considerations' and that 'while egalitarianism is a widely held principle, exactly what it requires is the subject of much debate,' followed by an overview of that debate. The Abstract (p. 1) and Introduction (p. 2) pose the same question in competing terms — equal probability of benefit versus minimising harms to the least advantaged — and note these concepts have been long 'formalising, differentiating and debating' by moral and political philosophers, with Section 3.1 (p. 6) listing competing egalitarian currencies (welfare, resources, capabilities, equal political status). This clearly supports the attributed point that there is deep, unresolved disagreement about what fairness requires; the 'everyone agrees in principle fairness is important' clause is the citing authors' framing, not judged here. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The most useful models also often quickly become very complex, introducing new issues around human interpretability [11,26].”
§ The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions > 5 WHICH TENSIONS?
What the sources say
- Partly supportedcites [11] 11. Finale Doshi-Velez and Been Kim. 2017. Towards a rigorous science of inter- pretable machine learning. arXiv preprint arXiv:1702.08608 (2017). sourceExperimental — accuracy not yet measured
The introduction (p. 1) supports the interpretability-issue half of the sentence: ML systems that 'outperform humans on specific tasks' are being deployed in complex applications, auxiliary criteria 'often cannot be completely quantified', and interpretability becomes 'a popular fallback'. But nothing in our copy states the other half the claim attributes here — that the most useful models 'quickly become very complex'; searches for a performance/accuracy-versus-complexity or black-box statement returned no such passage (our copy is not truncated). On the contrary, Section 2 (pp. 3-4) locates the need for interpretability in 'an incompleteness in the problem formalization' rather than in model complexity, and explicitly says 'Not all ML systems require interpretability', so the complexity-driven framing is the citing authors' own, not this paper's. (LLM, conf=0.68)
- Not checked yetcites [26] 26. Adrian Weller. 2017. Challenges for transparency. arXiv preprint arXiv:1708.01870 (2017).
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“For example, when Google DeepMind collaborated with the Royal Free Hospital, they encountered a conflict between protecting the privacy of patient data, and their goal of using AI to improve early diagnosis of acute kidney injury [20].”
§ The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions > 4.1 Bridging the Gap between Principles and Practice
What the sources say
- Partly supportedcites [20] 20. Julia Powles and Hal Hodson. 2017. Google DeepMind and healthcare in an age of algorithms. Health and technology 7, 4 (2017), 351-367. sourceExperimental — accuracy not yet measured
Our copy confirms the collaboration and the privacy problems that arose in it: the Abstract (p. 1) says the Royal Free project to assist in managing acute kidney injury met 'issues of privacy and power emerging as potent challenges', and 'Data, direct care and consent' records that millions of identifiable records were transferred without consent or notice. But two elements of the citing sentence are not what this paper says. First, it does not frame the difficulty as a conflict between privacy and the goal of using AI for early AKI diagnosis: 'A 'national algorithm' for AKI' stresses that the detection algorithm was NHS England's 2014 patient-safety alert algorithm and that DeepMind claimed Streams 'would not apply any of the machine learning or artificial intelligence techniques' but would act as a mere interface. Second, 'Assessing the damage > Caldicott guidelines' and 'Questioning direct care' argue the privacy failing stems from the Trust-wide data trove exceeding the AKI purpose (data that could have been limited to patients having the gateway blood test), not from an inherent tension with diagnosing AKI. (LLM, conf=0.70)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
None of the sources cited for this claim have been checked yet.
In the paper
“The central terms used in principles are often ambiguous, masking conceptual complexity and differences in interpretation across populations.”
§ The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions > 3.1 Different Groups May Interpret Principles Differently
states a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
In the paper
“However, the generality of most principles also limits their ability to guide practical action [3]).”
§ The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions > 3.2 Principles Are Highly General
What the sources say
- Not checked yetcites [3] 3. Tom L. Beauchamp. 1995. Principlism and its alleged competitors. Kennedy Institute of Ethics Journal 5, 3 (1995), 181-198.
states a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
17 more claims
Create a free account to see all 52 claims and to verify the 7 references still open.
- It may state two separate claims as one. 15
- Probably not a substantial claim. 24
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“First-party audits are now common at large tech companies, many of which have established auditing teams within their engineering departments.”
From Who Audits the Auditors? Recommendations from a field scan of the algorithmic auditing ecosystem (2022) · cites this paper as [63]
Experimental — accuracy not yet measuredThe check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“While not without faults, safety-critical and regulated industries such as aerospace and medicine have long traditions of auditable processes and design controls that have dramatically improved safety [77,81].”
From Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing (2020) · cites this paper as [81]
Experimental — accuracy not yet measured
Citations
Cites
Principles of Biomedical Ethics
(2012) · Thieves of Virtue · doi:10.7551/mitpress/9079.003.0009 · Cited by 13,453
Towards A Rigorous Science of Interpretable Machine Learning
Doshi-Velez, F., & Kim, B. (2017) · arXiv · doi:10.48550/arxiv.1702.08608 · Cited by 3,121 · cited 1×
“The most useful models also often quickly become very complex, introducing new issues around human interpretability (Doshi-Velez and Kim 2017; Weller 2017).”
BackgroundDancy, J. (2004) · doi:10.1093/0199270023.001.0001 · Cited by 1,677 · cited 1×
“But they have spurred substantial debate: some argue that we should put no weight on principles and focus entirely on the elements of specific cases [10], while others have advocated a more moderate view, whereby principles should be considered in close conjunction with analysis of ‘paradigm’ cases [17].”
BackgroundAbuse of Casuistry: A History of Moral Reasoning
(1988) · doi:10.1525/9780520352797 · Cited by 1,061
The Abuse of Casuistry: A History of Moral Reasoning
Jonsen, A. R., & Toulmin, S. E. (1988) · Cited by 773
Fairness in Machine Learning: Lessons from Political Philosophy
Binns, R. (2017) · arXiv · doi:10.48550/arxiv.1712.03586 · Cited by 676 · cited 2×
“‘justice’ for example are subject to substantial moral and political disagreement [5, 8].”
BackgroundClouser, K. D., & Gert, B. (1990) · Journal of Medicine and Philosophy · doi:10.1093/jmp/15.2.219 · Cited by 552 · cited 2×
“As [8] point out, the principle of ‘justice’ in bioethics does not say anything about what is just or unjust, leaving this to the agent to decide for themselves.”
BackgroundGoogle DeepMind and healthcare in an age of algorithms
Powles, J., & Hodson, H. (2017) · Health and Technology · doi:10.1007/s12553-017-0179-1 · Cited by 511 · cited 2×
“…Data Ethics and Innovation); • Researchers across disciplines whose work aims to inform these ways that AI ethics is put into practice: exploring the technical, philosophical, or legal aspects of using algorithms in ethical ways [20] or synthesising and translating such research into practice [27].”
BackgroundEthical governance is essential to building trust in robotics and artificial intelligence systems
Winfield, A. F. T., & Jirotka, M. (2018) · Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences · doi:10.1098/rsta.2018.0085 · Cited by 493 · cited 2×
“…Data Ethics and Innovation); • Researchers across disciplines whose work aims to inform these ways that AI ethics is put into practice: exploring the technical, philosophical, or legal aspects of using algorithms in ethical ways [20] or synthesising and translating such research into practice [27].”
BackgroundThe Intuitive Appeal of Explainable Machines
Selbst, A. D., & Barocas, S. (2018) · doi:10.2139/ssrn.3126971 · Cited by 392
Research works printed but not linked: 1
- Adel Tameem, 2018, International Conference on Machine Learning. 50--59
Cited by
The global landscape of AI ethics guidelines
Jobin, A., Ienca, M., & Vayena, E. (2019) · Nature Machine Intelligence · doi:10.1038/s42256-019-0088-2 · Cited by 6,220
“Such divergences and tensions illustrate a gap at the cross-section of principle formulation and their implementation into practice 133,134 .”
Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., Smith-Loud, J., Theron, D., & Barnes, P. (2020) · Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency · doi:10.1145/3351095.3372873 · Cited by 2,245
AI governance: themes, knowledge gaps and future agendas
Birkstedt, T., Minkkinen, M., Tandon, A., & Mäntymäki, M. (2023) · Internet Research · doi:10.1108/intr-01-2022-0042 · Cited by 309
Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims
Brundage, M., Avin, S., Wang, J., Belfield, H., Krueger, G., Hadfield, G., Khlaaf, H., Yang, J., Toner, H., Fong, R., et al. (2020) · arXiv · doi:10.48550/arxiv.2004.07213 · Cited by 306
“Further, they are amenable to various interpretations [13][14] and agreement on them can mask deeper disagreements [5].”
BackgroundResponsible AI—two frameworks for ethical design practice
Peters, D., Vold, K., Robinson, D., & Calvo, R. A. (2020) · IEEE Transactions on Technology and Society · doi:10.1109/tts.2020.2974991 · Cited by 293
“It is unlikely that the same priorities and solutions applicable to one domain, context, or project, will translate across to others [22], [25], [26].”
Background
Metadata
- Authors
- Whittlestone, J., Nyrup, R., Alexandrova, A., & Cave, S.
- Affiliations
- University of Cambridge (GB)
- Venue
- Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society
- Publisher
- ACM
- Published
- 2019-01-27
- Pages
- 195-200
- ISSN
- 3065-8365 (ISSN-L)
- OpenAlex
- W2956792022
- Cited by
- 399
- References
- 22
- Citations by year
- 2026: 99 · 2025: 90 · 2024: 71 · 2023: 40 · 2022: 39 · 2021: 36 · 2020: 19 · 2019: 5
- FWCI
- 25.87
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- Proceedings of the AAAI/ACM Conference on AI Ethics and Society (publishedVersion, cc-by)
- Funders
- Nuffield Foundation
- Leverhulme Trust
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Social Sciences › Business, Management and Accounting › Organizational Behavior and Human Resource Management › Ethics in Business and Education
- Keywords
- AI ethics, ethical principles, principle-based ethics, bioethics, ethical conflicts, applied ethics, common good, fairness, privacy, autonomy, moral norms, case-based decision making
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions. https://app.cochat.ai/papers/10.1145/3306618.3314289
- The paper
- Whittlestone, J., Nyrup, R., Alexandrova, A., & Cave, S. (2019). The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions. Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. https://doi.org/10.1145/3306618.3314289
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.