ETHICAL ARTIFICIAL INTELLIGENCE
Navigating the Landscape for Fairness, Accountability, and Human Control of Automated Systems
Nuriddin Khudoyberdiev
LL.M., Penn State Law, The Pennsylvania State University, USA; LL.M. and LL.B., Tashkent State University of Law, Uzbekistan.
n.khudoyberdiyev.law@gmail.com
Abstract
Artificial intelligence (AI) has matured from a research curiosity into general-purpose infrastructure that mediates hiring, lending, sentencing, medical triage, and the public information environment. The same systems that promise efficiency gains also encode historical inequities and concentrate decisional power in opaque computational pipelines. This article examines three interlocking ethical imperatives—fairness, accountability, and human control—drawing on the formal definitions of algorithmic fairness and the impossibility theorems that constrain their joint satisfaction1, and on the emerging regulatory landscape comprising the EU AI Act2, the UNESCO Recommendation, the NIST AI RMF, and the Republic of Uzbekistan’s 2025 AI Law3. Through case studies of COMPAS, Amazon’s discontinued resume tool, the Dutch SyRI system, and generative-AI harms, the article argues that ethical AI is an iterative socio-technical practice rather than a one-time engineering achievement.
Keywords: AI ethics; algorithmic fairness; accountability; human oversight; EU AI Act; UNESCO; NIST AI RMF; generative AI; Uzbekistan AI law.
1. Introduction
Artificial intelligence has become a cornerstone of contemporary technological development, restructuring industries, economies, and everyday social interaction. Statistical pattern recognition has moved within a generation from a specialist tool to a substrate of public administration, commerce, medicine, education, and law enforcement. The promise of intelligent automation is shadowed by documented failures: biased recidivism scores4; recruitment systems that reproduce gender exclusion; facial-recognition tools that misidentify minority groups at much higher rates5; and, since 2022, generative models capable of producing fluent misinformation, non-consensual imagery, and synthetic political speech at near-zero cost6.
This article proceeds from the premise that ethical considerations in AI are best understood as three interlocking domains. Fairness concerns the distribution of benefits and burdens across persons and groups. Accountability concerns the assignment of responsibility within complex socio-technical systems in which design, training, deployment, and use are separated across firms, jurisdictions, and time7. Human control concerns the preservation of human agency in domains where delegation to automation is technically feasible but normatively contested.
The analysis is situated within a regulatory landscape that did not exist when the first wave of AI-ethics scholarship was written: the UNESCO Recommendation adopted by 193 Member States; the NIST AI Risk Management Framework and its Generative AI Profile; the EU AI Act, the world’s first horizontal AI statute; and the Republic of Uzbekistan’s November 2025 Law on Relations Arising from the Use of AI, the first such statute in Central Asia. Sections 2–9 develop the methodology, the three substantive domains, four case studies, and the integrative discussion.
2. Methods
The investigation employs a deliberately pluralist methodology because the relevant questions cross disciplinary boundaries that no single approach fully respects. Technical questions belong to computer science; normative questions to moral and political philosophy; institutional questions to law and political economy. The study integrates four strands: a structured review of foundational scholarship on algorithmic fairness from Dwork and colleagues through Hardt, Price, and Srebro to Chouldechova and Kleinberg, Mullainathan, and Raghavan8; a typology of bias drawing on Barocas and Selbst and on Buolamwini and Gebru9; an analysis of transparency techniques including LIME and SHAP; and a comparative review of five regulatory instruments—UNESCO, OECD, NIST, the EU AI Act, and the Uzbek 2025 AI Law. Four case studies—COMPAS, Amazon’s resume tool, SyRI, and generative-AI harms—were selected for documentary richness and capacity to illuminate distinct ethical dimensions.
3. Fairness in AI Algorithms
Fairness in algorithmic decision-making is not a single property but a family of related concepts10. A useful taxonomy distinguishes group fairness (equalizing statistical properties across protected groups) from individual fairness (similar individuals receive similar treatment under a task-relevant metric). Five group-fairness definitions dominate practice. Demographic parity requires equal positive-prediction rates across groups. Equalized odds requires equal true- and false-positive rates11. Predictive parity requires equal positive predictive value. Counterfactual fairness requires that a prediction would not change in a counterfactual world with a different protected attribute12. Individual fairness requires similar treatment of similar individuals.
A foundational impossibility result, established independently by Chouldechova and by Kleinberg and colleagues, shows that calibration within groups and equal error rates cannot be jointly satisfied except in two pathological cases—equal base rates across groups, or a perfect classifier13. The COMPAS controversy is the canonical illustration: scores were calibrated within race but exhibited unequal error rates, and the cause was the differential base rate of re-arrest, itself a product of structural inequalities in policing. The choice of fairness criterion is therefore a moral and political choice, not merely a technical one.
Bias enters AI systems through at least six channels: historical bias (training data accurately reflecting past discrimination); representation bias (under-sampled minorities, as in the order-of-magnitude error gaps documented in commercial facial analysis)14; measurement bias (proxy variables that differ across groups); aggregation bias (single models that perform poorly on subgroups); deployment bias (model used in a context different from training); and feedback-loop bias (predictions that shape future training data). Mitigation operates pre-processing, in-processing, or post-processing, supplemented by organizational practices such as datasheets, model cards, and red-teaming15.
4. Transparency, Accountability, and the Global Regulatory Landscape
Transparency is the precondition for accountability, but contemporary AI systems—large neural networks and foundation models with billions of parameters—are not naturally transparent. Two post-hoc explanation methods dominate practice: LIME, which fits a local interpretable surrogate around a prediction16, and SHAP, which assigns each feature a contribution based on cooperative game theory17. Both produce unstable or misleading explanations when features are correlated, and an explanation of a model output is not the same as a justification of a model-aided decision: a SHAP plot does not tell a rejected applicant what they could have done differently, nor does it discharge an adjudicator’s duty to give reasons.
Four regulatory instruments anchor the contemporary landscape. The UNESCO Recommendation on the Ethics of AI, adopted on 23 November 2021 by 193 Member States18, articulates four core values and ten principles, including human oversight, transparency, accountability, and fairness, and is implemented through Readiness Assessment and Ethical Impact Assessment tools. The NIST AI Risk Management Framework, released in January 202319 with a Generative AI Profile in 2024, is voluntary and structured around four functions—GOVERN, MAP, MEASURE, MANAGE—that align with seven trustworthiness characteristics.
The EU Artificial Intelligence Act, the first comprehensive horizontal AI statute, entered into force on 1 August 202420. Prohibitions on unacceptable-risk practices applied from 2 February 2025, general-purpose AI obligations from 2 August 2025, with full application on 2 August 2026. The Act prohibits social scoring by public authorities, manipulative subliminal techniques, real-time remote biometric identification in public spaces, untargeted scraping for facial-recognition databases, emotion inference in workplaces and schools, biometric categorization of sensitive characteristics, and predictive policing based solely on profiling21; penalties can reach €35 million or 7 per cent of global turnover. High-risk systems—listed in Annex III and covering employment, education, essential services, law enforcement, migration, and the administration of justice—trigger lifecycle risk management, data-governance, technical documentation, logging, human oversight, conformity assessment, and post-market monitoring obligations, with Fundamental Rights Impact Assessments required of some deployers22. General-purpose AI providers must publish technical documentation and a training-data summary; models with “systemic risk” face additional evaluation and incident-reporting duties.
On 1 November 2025, the Senate of Uzbekistan approved the Law on Relations Arising from the Use of AI Technologies, making Uzbekistan the first Central Asian state with dedicated AI legislation23. The law builds on Presidential Decree PF-132 of 2024 and Resolution PP-358 approving the AI Strategy until 2030. It defines AI; establishes that algorithmic decisions affecting human rights cannot be made without human involvement—aligning with the EU oversight norm; prohibits AI that harms life, health, freedom, honour, or dignity; and mandates labelling of AI-generated images, audio, and video. Registered cases involving illegal AI-processed material rose from 1,129 in 2023 to 3,553 in 2024, supplying empirical motivation for the human-rights framing of the statute. Open questions concern enforcement capacity, by-law detail, and redress mechanisms. Beyond these anchors, the OECD AI Principles, the Council of Europe Framework Convention on AI, and the UN Secretary-General’s Governing AI for Humanity report24 complete the contemporary landscape.
5. Meaningful Human Control
Human control is not a single state but a spectrum25. Human-in-the-loop (HITL) architectures require explicit human approval before consequential outputs and are appropriate for high-stakes, low-frequency decisions; human-on-the-loop (HOTL) systems act autonomously under supervisory monitoring; human-in-command (HIC) architectures preserve overall human authority over the design and life cycle of the system. The EU AI Act constrains fully autonomous configurations for most consequential applications26.
Formal oversight, however, is not meaningful oversight. The Dutch SyRI affair is illustrative: nominally human-reviewed risk scores generated harms that the District Court of The Hague found to violate Article 8 of the European Convention on Human Rights27. Caseworkers lacked the information, time, and procedural support needed to challenge model outputs. The literature on adjustable autonomy responds by proposing systems whose level of autonomy is dynamically calibrated to context. The automation paradox—operators of highly reliable systems become less able to intervene when intervention is needed—requires deliberate counter-design: maintained competence, calibrated confidence, contestability interfaces, and override audits.
6. Generative AI and Emerging Ethical Challenges
The release of ChatGPT in November 2022 and the subsequent proliferation of large language and multimodal models introduced ethical challenges the first wave of AI-ethics literature did not anticipate28. Large language models predict plausible continuations and lack a mechanism for distinguishing well-grounded claims from hallucinations—fluent and confident outputs that are factually inaccurate or invented. Mitigations—retrieval-augmented generation, confidence calibration, and tool use—reduce but do not eliminate the problem, so verification duties must rest with users and deploying institutions.
Synthetic media—deepfakes—are now accessible at near-zero marginal cost. Documented harms include non-consensual intimate imagery (overwhelmingly targeting women and adolescents), financial fraud through voice cloning, political disinformation in 2024 elections, and the “liar’s dividend” that allows authentic evidence to be denied. Mitigation layers include provenance technologies such as C2PA content credentials, detection systems, and legal measures such as the EU AI Act’s machine-readable marking requirement29 and Uzbekistan’s mandatory labelling provision. A second-order issue concerns the concentration of frontier-model capability in a small number of firms and jurisdictions, raising distributive concerns that fairness metrics alone cannot address.
7. Case Studies
7.1 COMPAS. ProPublica’s 2016 analysis of more than 10,000 Broward County defendants reported that COMPAS misclassified Black defendants who did not reoffend as high-risk at nearly twice the rate of comparable white defendants30. Northpointe replied that the tool satisfied predictive parity within race. Both observations were mathematically correct; the dispute was the canonical illustration of the impossibility theorem under unequal base rates. The case informed the EU AI Act’s classification of justice-administration AI as high-risk31.
7.2 Amazon’s Resume Tool. Amazon’s internal resume-screening system, in development from 2014, was found in 2015 to systematically downgrade resumes from women: it penalized the word “women’s,” downgraded graduates of two all-women’s colleges, and reproduced the demographics of Amazon’s engineering workforce32. Neutralizing explicit gender markers did not eliminate correlated proxies, and the project was scrapped. The episode illustrates historical bias and the limits of self-regulation. Regulatory responses include New York City Local Law 144 and Annex III of the EU AI Act.
7.3 The Dutch SyRI Case. The System Risk Indication (2014–2020) combined data from multiple government databases to produce welfare-fraud risk scores, disproportionately targeting low-income immigrant neighbourhoods. On 5 February 2020, the District Court of The Hague held that SyRI violated Article 8 of the ECHR for want of safeguards against arbitrary interference with private life33. The judgment informed the EU AI Act’s provisions on high-risk AI in public services and on Fundamental Rights Impact Assessments. A parallel toeslagenaffaire (childcare-benefits) algorithm precipitated the resignation of the Dutch cabinet in 2021.
7.4 Generative AI Harms. In Mata v. Avianca (2023), attorneys submitted six judicial citations entirely fabricated by ChatGPT and were sanctioned34. An AI-generated image of a Pentagon explosion briefly moved U.S. equity markets in 2024. Deepfakes targeted political figures in multiple national elections; European public-opinion data report 40 per cent of respondents worried about AI misuse in elections and 31 per cent believing AI had influenced their vote. Non-consensual “nudification” applications have emerged as one of the most severe and least tractable harms.
8. Discussion
Several themes follow. First, fairness is not reducible to a single technical metric; the impossibility theorem shows that choosing one criterion is, mathematically and morally, refusing others. A defensible regime specifies the procedural conditions under which the choice is made and documented—a proceduralist intuition reflected in both the EU AI Act and the UNESCO Recommendation. Second, bias mitigation is ongoing rather than terminal; lifecycle monitoring, audits, and structured documentation are constitutive of ethical deployment35. Third, transparency is necessary but not sufficient: LIME and SHAP make individual predictions inspectable but cannot substitute for the institutional standing, evidence, and remedy that give explanation practical force.
Fourth, accountability frameworks must accommodate distributed, transnational AI value chains36; the EU Act’s tripartite distinction among providers, deployers, and users and UNESCO’s mutual-recognition logic point toward cross-border governance. Fifth, human control must be designed, not presumed; the SyRI affair is the cautionary tale and adjustable-autonomy literature the constructive response. Sixth, generative AI requires its own governance vocabulary—provenance infrastructure, platform liability, and democratic deliberation about capability release.
9. Conclusion
Ethical AI is not a state to be achieved but a practice to be sustained. It depends on the continuing collaboration of engineers, ethicists, lawyers, regulators, and the public, and on institutional arrangements adequate to the scale and velocity of the technology. The instruments of the past five years—the UNESCO Recommendation, the NIST AI RMF, the EU AI Act37, the Council of Europe Framework Convention, and the Uzbek AI Law of 2025—do not exhaust the work that remains, but they constitute a recognizable beginning. The mathematical impossibility theorems ensure that fairness will remain contested38; the distributed character of AI value chains ensures that accountability will remain a coordination problem; and the rise of generative systems ensures that the governance agenda will continue to expand. The task ahead is to operationalize present commitments, extend their coverage to new generative challenges, and ensure that the populations most affected by AI retain a meaningful voice in the institutions that govern them.
References
Al-kfairy, M., Mustafa, D., Kshetri, N., Insiew, M., & Alfandi, O. (2024). Ethical challenges and solutions of generative AI: An interdisciplinary perspective. Informatics, 11(3), 58.
Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016, May 23). Machine bias. ProPublica.
Barocas, S., Hardt, M., & Narayanan, A. (2023). Fairness and machine learning: Limitations and opportunities. MIT Press.
Barocas, S., & Selbst, A. D. (2016). Big data’s disparate impact. California Law Review, 104, 671–732.
Beigang, F. (2023). Reconciling algorithmic fairness criteria. Philosophy & Public Affairs, 51(2), 166–190.
Benjamin, R. (2019). Race after technology: Abolitionist tools for the new Jim Code. Polity Press.
Buolamwini, J., & Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of Machine Learning Research, 81, 1–15.
Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163.
Coeckelbergh, M. (2020). AI ethics. MIT Press.
Council of Europe. (2024). Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225).
Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press.
District Court of The Hague. (2020). NJCM et al. v. The State of the Netherlands (SyRI), judgment of 5 February 2020, ECLI:NL:RBDHA:2020:865.
Dwork, C., Hardt, M., Pitassi, T., Reingold, O., & Zemel, R. (2012). Fairness through awareness. In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference (pp. 214–226).
Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin’s Press.
European Commission, High-Level Expert Group on AI. (2019). Ethics guidelines for trustworthy AI. Brussels: European Commission.
Hagendorff, T. (2024). Mapping the ethics of generative AI: A comprehensive scoping review. arXiv preprint arXiv:2402.08323.
Hardt, M., Price, E., & Srebro, N. (2016). Equality of opportunity in supervised learning. Advances in Neural Information Processing Systems, 29, 3315–3323.
Kleinberg, J., Mullainathan, S., & Raghavan, M. (2017). Inherent trade-offs in the fair determination of risk scores. In Proceedings of the 8th Innovations in Theoretical Computer Science Conference.
Kusner, M. J., Loftus, J., Russell, C., & Silva, R. (2017). Counterfactual fairness. Advances in Neural Information Processing Systems, 30, 4066–4076.
Law of the Republic of Uzbekistan on Amendments and Additions to Certain Legislative Acts in Connection with Regulating Relations Arising from the Use of Artificial Intelligence Technologies. (2025). Adopted by the Senate, 1 November 2025.
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30, 4765–4774.
Mata v. Avianca, Inc., 22-cv-1461 (PKC), 2023 WL 4114965 (S.D.N.Y. June 22, 2023).
Methnani, L., Aler Tubella, A., Dignum, V., & Theodorou, A. (2021). Let me take over: Variable autonomy for meaningful human control. Frontiers in Artificial Intelligence, 4, 737072.
Mitchell, M. (2019). Artificial intelligence: A guide for thinking humans. Farrar, Straus and Giroux.
Mitchell, M., Wu, S., Zaldivar, A., et al. (2019). Model cards for model reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 220–229).
National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1.
National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1.
New York City Local Law 144 of 2021 (Automated Employment Decision Tools).
OECD. (2019, revised 2024). Recommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449).
O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Crown.
Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 (Artificial Intelligence Act). OJ L, 12 July 2024.
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why should I trust you?”: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference (pp. 1135–1144).
Russell, S. (2019). Human compatible: Artificial intelligence and the problem of control. Viking.
Salih, A., et al. (2025). A perspective on explainable artificial intelligence methods: SHAP and LIME. Advanced Intelligent Systems.
Selbst, A. D., Boyd, D., Friedler, S. A., Venkatasubramanian, S., & Vertesi, J. (2019). Fairness and abstraction in sociotechnical systems. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 59–68).
UNESCO. (2021). Recommendation on the Ethics of Artificial Intelligence. Paris: UNESCO.
United Nations Secretary-General’s High-Level Advisory Body on AI. (2024). Governing AI for humanity: Final report.
Chouldechova, 2017. Kleinberg, Mullainathan, & Raghavan, 2017.↩︎
EU AI Act, Reg. (EU) 2024/1689.↩︎
Uzbek AI Law, 1 Nov. 2025.↩︎
Angwin et al., 2016, Machine bias (ProPublica).↩︎
Buolamwini & Gebru, 2018, Gender shades.↩︎
Hagendorff, 2024.↩︎
Selbst et al., 2019, Fairness and abstraction.↩︎
Dwork et al., 2012, Fairness through awareness. Hardt, Price, & Srebro, 2016, Equality of opportunity.↩︎
Barocas & Selbst, 2016, Big data’s disparate impact.↩︎
Barocas, Hardt, & Narayanan, 2023, Fairness and machine learning.↩︎
Hardt, Price, & Srebro, 2016, Equality of opportunity.↩︎
Kusner et al., 2017, Counterfactual fairness.↩︎
Chouldechova, 2017. Kleinberg, Mullainathan, & Raghavan, 2017.↩︎
Buolamwini & Gebru, 2018, Gender shades.↩︎
Mitchell et al., 2019, Model cards.↩︎
Ribeiro, Singh, & Guestrin, 2016, LIME.↩︎
Lundberg & Lee, 2017, SHAP.↩︎
UNESCO Recommendation on the Ethics of AI, 2021.↩︎
NIST AI RMF 1.0, 2023.↩︎
EU AI Act, Reg. (EU) 2024/1689.↩︎
EU AI Act, Reg. (EU) 2024/1689, Art. 5.↩︎
EU AI Act, Reg. (EU) 2024/1689, Art. 27.↩︎
Uzbek AI Law, 1 Nov. 2025.↩︎
UN HLAB on AI, Governing AI for humanity, 2024.↩︎
Methnani et al., 2021, Variable autonomy.↩︎
EU AI Act, Reg. (EU) 2024/1689, Art. 14.↩︎
District Court of The Hague, SyRI judgment, 5 Feb. 2020.↩︎
Hagendorff, 2024.↩︎
EU AI Act, Reg. (EU) 2024/1689, Art. 50.↩︎
Angwin et al., 2016, Machine bias (ProPublica).↩︎
EU AI Act, Reg. (EU) 2024/1689, Annex III.↩︎
Barocas, Hardt, & Narayanan, 2023, Fairness and machine learning.↩︎
District Court of The Hague, SyRI judgment, 5 Feb. 2020.↩︎
Mata v. Avianca, 2023.↩︎
Mitchell et al., 2019, Model cards.↩︎
Crawford, 2021, Atlas of AI.↩︎
EU AI Act, Reg. (EU) 2024/1689.↩︎
Chouldechova, 2017. Kleinberg, Mullainathan, & Raghavan, 2017.↩︎
