163
Editorial Note:* This article serves as an introductory overview of a subject that is at once legally urgent, ethically contested, and institutionally underexamined. Given its breadth, each section necessarily operates as a thematic threshold rather than an exhaustive treatment. Future columns in this series will develop each dimension in depth: the algorithmic fairness problem and its implications for judicial practice; the specific governance gaps in the AI Act’s enforcement architecture; the bioethical reconfiguration required by opaque medical AI systems; and the question of liability in distributed human-algorithmic decision chains. The present text maps the terrain. What follows, in the columns ahead, will excavate it.*
Introduction: An Invisible Judge #
The integration of artificial intelligence into decision-making did not happen abruptly, nor did it arrive through a single institutional rupture. It evolved gradually — first in research, then in clinical and judicial practice.
I have been working for thirteen years in courts of first instance, primarily in civil proceedings. I know what makes a case file heavy: psychiatric assessments, forensic reports, social evaluations. I also know where technology ends and judgment begins, or at least that was true until recently. Artificial intelligence (hereafter AI) does not come to abolish that distinction. It blurs it. The critical question is no longer whether AI will enter the justice system — it already has — but who bears responsibility when its use leads to error. The defendant, the victim, the patient: none of them asks which model processed their data. Yet each of them expects someone to be accountable.
What Exactly Is Changing #
Machine learning algorithms are not an invention of the last five years. What is changing now is where they are applied: in domains where human judgment was never merely an option — it was the process itself.
In oncology, the Sybil model developed at MIT Jameel Clinic detects lung cancer risk up to six years before symptoms appear, achieving an AUC of 0.92–0.94 for one-year prediction, based solely on low-dose CT scans.1 We are speaking of medicine that diagnoses people who feel perfectly healthy and this is no longer considered remarkable; it is simply the direction the field has taken. Kuhn was right: a gap is not being filled, a new map is being drawn.2
In forensic medicine, digital tools are now entering the drafting of expert reports. They identify injury patterns, estimate time of death, and filter toxicological data. Research platforms such as InForensic at the University of Piraeus have already broken new ground. No official announcement was made that something had changed. It simply changed.
What the Law Says — and What It Leaves Unsaid #
The AI Act
The AI Act
Regulation (EU) 2024/1689, the AI Act, entered into force on 1 August 2024 and applies progressively.3 Its phased implementation reflects both the original risk-based architecture of the Regulation and the subsequent legislative adjustments intended to facilitate effective compliance. The first prohibitions applied from February 2025; obligations for general-purpose AI systems (GPAI) have been in force since August 2025. Following the provisional political agreement reached by the Council and the European Parliament on 7 May 2026 under the AI Omnibus package, the application of obligations for stand-alone high-risk AI systems listed in Annex III — including judicial and medical applications — is expected to be deferred until 2 December 2027, while obligations for high-risk AI systems embedded in regulated products under Annex I are expected to apply from 2 August 2028, subject to the adoption of the Digital Omnibus package.3,18
The Regulation is structured around risk tiers. What is deemed unacceptable, such as social scoring of citizens, is expressly prohibited. Systems falling within the high-risk category, including medical and judicial applications, are subject to strict requirements regarding transparency, documentation, and human oversight. Annex III, paragraphs 6–8, explicitly designates systems that process medical data or are used in judicial proceedings as high-risk.3 If a machine participates in decisions that affect lives, someone must be capable of sitting in the dock.
The GDPR: A Right That Was Already There
The GDPR: A Right That Was Already There
Regulation (EU) 2016/679 — the GDPR — has since 2018 contained a provision that is rarely invoked: Article 22 prohibits decisions taken solely on the basis of automated processing when those decisions produce significant legal effects.4 This is not a theoretical provision. In every forensic opinion or recidivism risk assessment that is co-produced by an algorithm, this article asserts its relevance — provided someone chooses to invoke it.
The Accountability That Finds No Address
The Accountability That Finds No Address
The most intractable legal problem posed by AI is not definitional. It is the question of responsibility. The physician says “the algorithm recommended it.” The manufacturer says “the physician should have exercised judgment.” The institution says “protocols were followed.” In this way, a complex web of distributed responsibility forms — one in which identifying a clearly accountable party becomes structurally difficult.
The traditional model was straightforward: a decision belonged to a named individual. Now the algorithm proposes, the professional signs, the institution provides cover, and the manufacturer has already fixed the parameters.5,6 Accountability does not disappear — it becomes fragmented and harder to locate. This diffusion creates structural gaps that obstruct the attribution of liability when errors occur.
Bioethical Tensions #
The four foundational principles of bioethics — autonomy, beneficence, non-maleficence, and justice are not invalidated by AI,5 but their application becomes substantially more complex once algorithmic systems are interposed. Autonomy for the patient or the accused presupposes understanding. Informed consent — a principle embedded in medical ethics for decades — requires that a person comprehend what they are agreeing to. But how does one explain to a patient the logic of an algorithm that even their own physician cannot fully articulate?7 Consent risks becoming a formality precisely where it matters most.
Beneficence and non-maleficence require us to understand how a system fails. AI systems trained predominantly on data from North American and Northern European patient populations may perform less reliably when applied to populations with distinct epidemiological, genetic, or healthcare-system profiles — a structural limitation acknowledged across the literature on algorithmic generalisability.8 Whether and to what extent this applies to Greek clinical settings remains an empirical question that has not yet been systematically studied — and that absence of study is itself a governance gap.9
Justice raises a question that remains unanswered: who will gain access to advanced AI? Large private hospitals and metropolitan courts, or public institutions and regional jurisdictions as well? If the technology is distributed unequally, the two-tiered system that follows will not concern healthcare alone — it will concern the law.
COMPAS: A Story That Is Not Only American #
The COMPAS system (Correctional Offender Management Profiling for Alternative Sanctions) was deployed across the United States to assess the risk of reoffending — a tool that informed decisions on pre-trial detention and sentencing. An independent analysis by ProPublica in 2016 found that the system flagged African American defendants as high risk at nearly twice the rate of white defendants with comparable criminal histories.10 Analogous forms of bias have been identified in healthcare management algorithms.11
This case does not reflect American exceptionalism. It reveals something deeper and harder to address through legislation: the concept of “fair” prediction is not mathematically neutral.
Northpointe challenged ProPublica with a 39-page technical memorandum, arguing that COMPAS achieves “predictive parity”, meaning that the same risk score corresponds to the same actual probability of reoffending, regardless of race.12 Chouldechova at Carnegie Mellon University then demonstrated with mathematical rigour why both sides are simultaneously correct — and why that is precisely the problem: when base rates of reoffending differ between groups, it is mathematically impossible to satisfy both error-rate equality and predictive parity at the same time.13 This is not a technical flaw correctable with better data. It is a fundamental incompatibility within the very concept of algorithmic fairness.
That incompatibility concerns every court considering the adoption of such tools, including Greek ones. The question every judicial officer must ask is not “Do I trust the technology?” but rather: “Which definition of justice has this algorithm encoded? And who decided that definition was the right one?” The mathematical impossibility identified by Chouldechova is not a technical defect awaiting correction. It is a constitutional question in algorithmic form: whose conception of equality is encoded, and by what authority. In a legal order grounded in Article 6 of the ECHR and Articles 20–21 of the EU Charter of Fundamental Rights, the answer cannot be left to software vendors or procurement committees. Every court considering the adoption of risk-assessment tools — including Greek courts — must first answer a question that no algorithm can answer on its behalf: which trade-off between error types is legally and morally defensible? Silence on that question is not neutrality. It is a decision made by default.
Human Dignity: Not as a Principle, but as Practice #
Article 1 of the EU Charter of Fundamental Rights declares that human dignity is inviolable. No one disputes this on paper. The question is what it means in practice when an AI system flags “high probability of non-accidental injury” in a child and the forensic physician disagrees. Does the physician sign off? Follow the system’s indication? And if they resist, what do they document, that they did their job, or that they “disregarded the tool’?
In oncology, the same dilemma takes a different form. A 58-year-old man asks his physician: “Am I going to die?” The algorithm returned a 15% five-year survival estimate. While algorithms can generate quantitative projections, communicating a prognosis still requires clinical judgment and the human relationship between physician and patient. The literature confirms that a prognosis that forecloses the horizon can hasten what we fear: depression following an unambiguous negative prediction demonstrably worsens clinical outcomes.14,15 Psychological factors — hope, resilience, the will to engage with treatment — profoundly shape the course of illness in ways that algorithmic models do not, and perhaps cannot, adequately capture.
What the law has not yet confronted is this: if a prognosis communicated on the basis of an algorithmic output demonstrably worsens a patient’s condition, who bears responsibility for the harm? The physician who repeated the figure? The system that generated it? The institution that deployed it without validation on comparable populations? Article 3 of the EU Charter — the right to physical integrity — and the established doctrine of informed consent in both medical law and the GDPR’s Article 22 provide the legal architecture for this question. What is missing is not the law. It is the institutional will to ask it.
Three Principles That Must Not Remain on Paper #
The following three principles are not philosophical declarations. They are questions that every institution must answer before deploying an AI system in a governance context.
First, dignity as a non-negotiable threshold. Any system that produces systematically inferior outcomes for particular groups — regardless of aggregate technical efficiency — fails the equality criterion. The AI Act states this explicitly in Articles 9–15,3 but legislation does not enforce itself. The postponement of high-risk obligations until 2027–2028 does not suspend the force of these principles — courts and hospitals are already obligated today to ask the same questions.
Second, human oversight with genuine substance. The forensic expert or judge who countersigns an algorithmic report must have the time to read it, the knowledge to evaluate it, and — critically — the institutional protection to dissent. Without that protection, the signature becomes a formality and oversight becomes an illusion.16
Third, the right to challenge the method, not merely the outcome. If a citizen cannot contest the reasoning by which an algorithm reached its conclusion, they face a witness who cannot be cross-examined. Article 22 of the GDPR4 provides the legal basis — but only if it is used.
The Gap That Has No Name #
The CEPEJ Ethical Charter (2018) articulates the right principles: respect for fundamental rights, non-discrimination, data quality, and human control.16 But it is not binding. And that is the problem: the Charter asks things of institutions that no institution is legally required to do. The AI Act addresses the binding deficit, but its enforcement depends on national supervisory authorities that, in many Member States, have yet to reach full operational capacity. In practice, the gap is not legal. It is organizational and political. Between what the Regulation mandates and what happens tomorrow morning in a courtroom, a considerable distance remains.
As of mid-2026, fewer than half of EU Member States have designated a fully operational national supervisory authority under the AI Act with the staffing and budget to conduct meaningful conformity assessments of high-risk AI systems. Meanwhile, AI-assisted tools are already in operational use: forensic platforms process evidence, risk instruments inform pre-trial decisions in several jurisdictions, and hospital procurement has proceeded without the conformity documentation that the AI Act will eventually require. The CEPEJ Ethical Charter of 2018 articulated the correct principles — and they have been widely cited and largely ignored. Institutional adoption requires not only legal obligation but budgetary commitment, professional training, and the protection of the individual practitioner who raises an objection. None of these are provided by a Charter.
The forensic community, the judiciary, and legal scholars cannot remain spectators while decisions are being made about tools they will be expected to use. If those decisions are made exclusively among software vendors, technical committees, and funding bodies, the tools will be installed without ever having been scrutinized by the people who will be required to put their names to them.17
Conclusion: A Question That Cannot Be Deferred #
There is one thing I know with certainty after thirteen years in the courts: every decision leaves a trace. In the file, in the memory of the accused, in the life of the person harmed. That trace is a signature, it belongs to a person who answers for it by name.
Artificial intelligence does not sign. That is not a philosophical observation — it is a legal gap. And while the AI Act and the GDPR provide the regulatory framework, no Regulation can substitute for the decision made every morning in a courtroom or an intensive care unit: the decision of a human being who stands before another human being and assumes the responsibility of judgment.
This does not mean rejecting the technology. It means something more demanding: knowing precisely what we delegate to it and what we hold inviolable. The line between the two will not draw itself. We will draw it — or it will not be drawn at all.
*Note: *This article reflects the personal scholarly views of the author. No conflicts of interest are declared. The research received no external funding.
1 Mikhael, P.G., Wohlwend, J., Yala, A., et al. (2023). Sybil: A validated deep learning model to predict future lung cancer risk from a single low-dose chest CT. Journal of Clinical Oncology, 41(12), 2191–2200. https://doi.org/10.1200/JCO.22.01345
2 Kuhn, T.S. (1962). The structure of scientific revolutions. University of Chicago Press.
3 European Parliament & Council of the EU. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, L 1689/1. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
4 European Parliament & Council of the EU. (2016). Regulation (EU) 2016/679 on the protection of natural persons with regard to the processing of personal data (General Data Protection Regulation). Official Journal of the European Union, L 119/1. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679
5 Beauchamp, T.L., & Childress, J.F. (2019). Principles of biomedical ethics (8th ed.). Oxford University Press.
6 Topol, E. (2019). Deep medicine: How artificial intelligence can make healthcare human again. Basic Books.
7 London, A.J. (2019). Artificial intelligence and black-box medical decisions: Accuracy versus explainability. Hastings Center Report, 49(1), 15–21. https://doi.org/10.1002/hast.973
8 Seyyed-Kalantari, L., Zhang, H., McDermott, M.B., Chen, I.Y., & Ghassemi, M. (2021). Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine, 27(12), 2176–2182.
9 Tilala, H.M., Chenchala, P.K., Choppadandi, A., et al. (2024). Ethical considerations in the use of artificial intelligence and machine learning in health care: A comprehensive review. Cureus, 16(6), e62443. https://doi.org/10.7759/cureus.62443
10 Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine bias. ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
11 Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. https://doi.org/10.1126/science.aax2342
12 Dieterich, W., Mendoza, C., & Brennan, T. (2016). COMPAS risk scales: Demonstrating accuracy, equity and predictive parity. Northpointe Inc. Research Department. https://go.volarisgroup.com/rs/430-MBX-989/images/ProPublica_Commentary_Final_070616.pdf
13 Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163. https://doi.org/10.1089/big.2016.0047
14 Satin, J.R., Linden, W., & Phillips, M.J. (2009). Depression as a predictor of disease progression and mortality in cancer patients: A meta-analysis. Cancer, 115(22), 5349–5361. https://doi.org/10.1002/cncr.24561
15 National Cancer Institute. (2024). Depression in cancer patients (PDQ®) — Patient version. https://www.cancer.gov/about-cancer/coping/feelings/depression-pdq
16 European Commission for the Efficiency of Justice (CEPEJ). (2018). European ethical charter on the use of artificial intelligence in judicial systems and their environment. Council of Europe.
17 Vardas, E.P., Marketou, M., & Vardas, P.E. (2025). Medicine, healthcare and the AI Act: Gaps, challenges and future implications. European Heart Journal — Digital Health, 6(4), 833–839. https://doi.org/10.1093/ehjdh/ztaf041
18 Council of the European Union. (2026). Artificial intelligence: Council and Parliament agree to simplify and streamline rules. Press Release, 7 May 2026. https://www.consilium.europa.eu/en/press/press-releases/2026/05/07/artificial-intelligence-council-and-parliament-agree-to-simplify-and-streamline-rules/
Have any thoughts?
Share your reaction or leave a quick response — we’d love to hear what you think!