How to cite this article / Come citare questo articolo
Zito, L.. (2026). The Anthropological Sovereignty of Law: Artificial Intelligence, Interpretation, and the Human Reserve of Judgment in the Algorithmic Age. Aequitas Magazine, 9, 1-17.
DOI: https://doi.org/10.5281/zenodo.22275427
ABSTRACT (EN) The expansion of artificial intelligence systems constitutes one of the foremost legal challenges of the twenty-first century. This article advances an original theoretical category: the principle of the anthropological sovereignty of law. According to this principle, final legal decisions producing legal effects, or comparably significant effects on fundamental rights, must remain attributable to a competent and accountable human decision-maker acting within the framework of law. Artificial intelligence may assist that decision-maker, but it cannot become an autonomous or parallel source of decisional authority. This requirement rests on the distinctively human capacity to understand the existential significance of legal conflict and to assume responsibility for its resolution. Two corollaries follow: the semantic irreducibility of law, according to which legal meaning cannot be reduced to computational output, and the human reserve of judgment, which gives the principle its institutional and procedural dimension. The article explains and defends the use of the term sovereignty in contrast to related concepts such as dignity, human centrality, and the primacy of the person. It then tests the thesis against some of the strongest arguments in favour of extensive reliance on artificial intelligence in legal decision-making, including Kahneman, Sibony and Sunstein’s theory of noise and Casey and Niblett’s proposal of microdirectives. The analysis subsequently applies this framework to Regulation (EU) 2024/1689 (the AI Act), interpreted as the most systematic legislative articulation to date, within European law, of the requirement of human control over algorithmic decision-making. The principle is then examined in relation to the principal legal professions — advocate, judge, civil-law notary and legal adviser — in order to identify what remains irreducibly human in the exercise of each function. The concluding section addresses the objection that the proposed category merely renames legal personalism. It argues, instead, that the principle of the anthropological sovereignty of law translates the axiological core of personalism into an operational legal category capable of responding to an interlocutor that traditional personalism never had to confront: artificial intelligence.
ABSTRACT (IT) L’espansione dei sistemi di intelligenza artificiale costituisce una delle principali sfide giuridiche del XXI secolo. Il presente contributo propone una categoria teorica originale: il principio della sovranità antropologica del diritto. Secondo tale principio, le decisioni giuridiche finali produttive di effetti giuridici, o comunque di effetti comparabilmente significativi sui diritti fondamentali, devono rimanere imputabili a un decisore umano competente e responsabile, operante nell’ambito del diritto. L’intelligenza artificiale può assistere tale decisore, ma non può divenire una fonte autonoma o parallela di autorità decisionale. Tale esigenza si fonda sulla capacità specificamente umana di comprendere il significato esistenziale del conflitto giuridico e di assumere la responsabilità della sua risoluzione. Da ciò discendono due corollari: l’irriducibilità semantica del diritto, secondo cui il significato giuridico non può essere ridotto a un output computazionale, e la riserva umana di giudizio, che conferisce al principio la sua dimensione istituzionale e procedurale. L’articolo illustra e difende l’impiego del termine sovranità rispetto a categorie affini quali dignità, centralità dell’essere umano e primato della persona. La tesi viene quindi sottoposta al confronto con alcuni dei più forti argomenti a favore di un ampio ricorso all’intelligenza artificiale nel processo decisionale giuridico, tra cui la teoria del noise di Kahneman, Sibony e Sunstein e la proposta delle microdirectives di Casey e Niblett. L’analisi applica successivamente tale quadro teorico al Regolamento (UE) 2024/1689, l’AI Act, interpretato come la più sistematica articolazione legislativa finora elaborata, nell’ambito del diritto europeo, dell’esigenza di controllo umano sulle decisioni algoritmiche. Il principio viene poi esaminato con riferimento alle principali professioni giuridiche — avvocato, giudice, notaio di civil law e consulente legale — al fine di individuare ciò che, nell’esercizio di ciascuna funzione, rimane irriducibilmente umano. La sezione conclusiva affronta l’obiezione secondo cui la categoria proposta costituirebbe una mera ridenominazione del personalismo giuridico. Si sostiene, al contrario, che il principio della sovranità antropologica del diritto traduca il nucleo assiologico del personalismo in una categoria giuridica operativa, capace di confrontarsi con un interlocutore che il personalismo tradizionale non aveva mai dovuto affrontare: l’intelligenza artificiale.
SUMMARY: 1. Introduction: The Thesis – 2. The Historical Relationship Between Technique and Law – 3. The Limits of the Algorithm: Interpretation and Justice – 4. Computation and Understanding: From Anthropological Irreducibility to the Principle of the Anthropological Sovereignty of Law – 5. The Anthropological Risk of Legal Posthumanism – 6. Engaging the Case for an Extensive Use of Artificial Intelligence in Law – 7. The AI Act as the Most Systematic European Legislative Articulation of Human Control over Algorithmic Decision-Making – 8. The Legal Professions in the Age of Artificial Intelligence: What Remains Irreducibly Human – 9. Governing Technique Through Law – 10. Conclusions.
1. Introduction: The Thesis
“Someone must have slandered Josef K.”
F. Kafka, The Trial, 1925.
This image, drawn from a context far removed from artificial intelligence, captures the risk this inquiry seeks to avert: a subject judged by a mechanism he cannot understand, and which does not understand him in turn. It is the risk of a justice administered by a system that computes an outcome without grasping the human meaning of what it decides – a process Kafkaesque before it is algorithmic, but one that artificial intelligence now renders concretely possible at unprecedented scale.
Artificial intelligence has changed how information is produced, decisions are formulated, and public and private processes are organised. Law is caught up in this transformation not only as a field called upon to regulate technological innovation, but as a domain in which technology increasingly assists interpretive, administrative, and even adjudicative activity. Legal search engines built on large language models, decision-support systems in public administration, predictive tools for litigation risk1: the range of applications grows faster than legislators can track.
The literature tends to move along two lines: an enthusiastic account of artificial intelligence’s potential in law, and an equally reconstructive catalogue of its risks. This contribution takes a further step, proposing a theoretical category that explains, systematically rather than descriptively, why law cannot be delegated to the machine.
The thesis may be stated as follows. There exists a principle of the anthropological sovereignty of law: final legal decisions producing legal or similarly significant effects on fundamental rights must remain attributable to a competent and responsible human decision-maker acting under law. Artificial intelligence may assist that decision-maker but cannot become an autonomous or parallel source of decisional authority. This requirement is grounded in the human capacity to understand the existential meaning of legal conflict and to bear responsibility for its resolution. Two corollaries follow. The first, hermeneutic and epistemological, is the semantic irreducibility of law: legal meaning presupposes the lived understanding of an experience – suffering, betrayed trust, vulnerability – that no statistical processing of textual correlations can reproduce from within. The second, institutional and procedural, is the human reserve of judgment: every high-impact decision producing legal or similarly significant effects on fundamental rights must remain attributable to, effectively reviewable by, and subject to the meaningful control of a competent and responsible human decision-maker.
The argument that follows is built to demonstrate, not merely to postulate, this principle and its corollaries. It proceeds from the historical relationship between technique and law (§2), to the theoretical problem of interpretation (§3) and the boundary between computation and understanding, from which the principle’s full formulation and terminological justification emerge (§4). It then examines legal posthumanism (§5) and, in dialogue with the case for an extensive use of artificial intelligence in law, asks whether the principle withstands the strongest objections raised against it (§6). It tests the principle against positive law, through a reading of the AI Act (§7), and against professional practice, asking what remains irreducibly human in the principal legal professions (§8).
2. The Historical Relationship Between Technique and Law
The history of law is bound up with the development of technique. Every technological innovation – from writing to the printing press, from telegraphy to computing, to generative artificial intelligence – has transformed how legal norms are produced, applied, and disseminated. Modern codification is unthinkable without movable-type printing; Weberian bureaucracy presupposes technologies of archiving and transmission; legal informatics has made mass access to case law and legislation possible. Artificial intelligence is not an abrupt rupture but the latest link in an already long chain.
Throughout this history, however, technique has performed an instrumental function with respect to law, never a foundational one. Giorgio Del Vecchio, elaborating his theory of natural law with variable content, treated justice as a meta-historical criterion against which technical-normative positive law must always be measured2: organisational technique changes; the criterion of justice endures as a regulative instance. Sergio Cotta’s account of juridicity as an original dimension of intersubjective human existence locates the foundation of law not in an apparatus of rules but in the relationship between subjects who recognise one another as persons3. A technical instrument, however refined, remains accidental to the relational substance that constitutes law – a historical antecedent, one might say, of the anthropological sovereignty argued here.
Santi Romano’s institutionalist theory, on which the legal order is above all an institution – an organised social entity, not a sum of norms – offers a further antidote to reducing law to a system of computable rules4: if law is institution, it lives in social practice and in the organisational structures of communities, and cannot be captured by any finite set of normative statements, however extensive. Artificial intelligence can simulate the application of rules; it cannot generate the institution, which presupposes the social recognition and historicity of the community that expresses it.
In modern constitutionalism, law does not arise from the technology available at a given moment but from the person, the political community, and the fundamental values the legal order treats as inalienable: dignity, liberty, equality. Article 2 of the Italian Constitution recognises the inviolable rights of the individual without conditioning that recognition on any technological state of affairs. Artificial intelligence marks a new and powerful phase of technical evolution, but it does not alter the nature of law as a cultural, relational, and axiologically oriented phenomenon.
3. The Limits of the Algorithm: Interpretation and Justice
The idea that law could be fully automated presupposes a reductive conception of legal experience, identifying law with the normative text and legal decision-making with the mechanical subsumption of fact under abstract category – a conception twentieth-century legal scholarship has long surpassed. Francesco Carnelutti, distinguishing the interpretation of law from the interpretation of fact and insisting on the creative dimension of interpretation, showed that applying a norm is never a purely deductive operation, but always involves a value judgment situating the concrete case within the axiological system of the legal order5.
Giovanni Gentile and the actualist tradition, proceeding from different premises, emphasised the historically situated, irreducibly non-automatic character of the legal act6. Natalino Irti, reflecting on the crisis of the legal category and the relationship between technique and norm in the age of globalisation, observed that technique tends to produce its own self-referential order, impatient of political-legal mediation; hence his insistence that law must set norms for technique, not the reverse7. The warning bears directly on artificial intelligence, where technique risks becoming itself a source of regulation, escaping normative control – the exact opposite of what the anthropological sovereignty of law seeks to preserve.
The jurist does not merely apply norms: she interprets them. Legal interpretation involves operations no algorithmic system currently reproduces in full: assessing factual and normative context; critically reconstructing facts from often contradictory evidence; balancing potentially conflicting fundamental rights; applying general clauses such as good faith, reasonableness, and proportionality; weighing equity in the individual case; and seeking not any solution but the just solution for that case, in light of the legal order’s values.
None of this can be fully formalised in an algorithm, however sophisticated. The algorithm recognises statistical correlations between data; the jurist attributes meaning within a shared horizon of sense. The algorithm produces probability estimates; the judge gives reasons that must be intersubjectively verifiable. The algorithm optimises predefined variables; law seeks justice, a category not reducible to an optimisation function. This distinction deserves closer elaboration, to which the following section is devoted.
4. Computation and Understanding: From Anthropological Irreducibility to the Principle of the Anthropological Sovereignty of Law
Contemporary generative artificial intelligence displays extraordinary linguistic capability: it drafts texts, summarises documents, identifies precedents, and assists the jurist’s work. What that capability technically consists of must be clarified, in order to see why it falls short of understanding legal meaning.
A standard autoregressive large language model generates text by estimating a probability distribution over the next token, conditional on the preceding context, on the basis of statistical regularities learned from large quantities of data. Tokens do not, by themselves, secure reference to the world or guarantee the kind of semantic grounding associated with lived human experience: “responsibility” is not, for the machine, a concept anchored in the lived experience of fault, harm, and reparation, but a numerical vector correlated with other vectors. This is, on this account, a syntactic manipulation of symbols, not a semantic operation in the sense that term carries in human experience: syntax governs formal relations between signs; semantics fixes their reference to a state of affairs and to a horizon of meaning lived by an understanding subject.
John Searle’s “Chinese Room” thought experiment clarifies why correctly manipulating symbols according to rules – even producing output indistinguishable from a competent human’s – entails no understanding on the part of whoever, or whatever, performs the operation. One who applies rules of correspondence between symbols one does not understand, however correct the output, does not understand what one is writing. Transposed to legal reasoning: a system that drafts a procedurally impeccable brief, correctly citing precedents and norms, does not thereby understand that behind the dispute stands a person injured in her dignity, a worker deprived of her livelihood, a minor whose future depends on the decision. The output is correct; the understanding of what is existentially at stake is absent.
Joseph Weizenbaum, creator of the ELIZA program, warned early against the anthropomorphic illusion generated by conversational systems8: that a user attributes understanding, even empathy, to a program manipulating symbols by syntactic rule demonstrates nothing about the machine, but reveals the human tendency to project intentionality onto any linguistically competent interlocutor. His distinction between what a computer can do and what it is appropriate for a computer to do bears directly on law: the technical capacity to generate a syntactically correct ruling does not make it appropriate to entrust an automated system with a decision on fundamental rights.
Legal meaning is never merely textual; it is a product of the relationship between normative text, historical fact, and the interpreter’s practical judgment, always situated within a cultural and institutional context. Understanding that conduct breaches contractual good faith, or that suffering amounts to compensable existential harm, is not an inference exhausted by recognising textual regularities across similar cases: it is an act of practical recognition rooted in the embodied experience of vulnerability, responsibility, and reciprocity. A system without a body, a biography, or exposure to risk and suffering does not possess – structurally, not for want of greater computational power – the experience from which that recognition originates. This is the symbol grounding problem9: absent an anchor in lived experience, symbols remain relations between symbols, a deferral that never touches the world of persons and their circumstances10.
Luciano Floridi’s philosophy of information situates artificial agents within an informational ecosystem in which they act as moral agents in a weak sense – capable of morally relevant effects without the attributes of full moral agency: intentionality, freedom, responsibility11. This distinction lets the jurist recognise a significant operational role for artificial intelligence without transferring to it the eminently legal and moral category of responsibility, which remains vested in the humans who design, deploy, or authorise it.
Martin Heidegger’s account of modern technique as Gestell, or enframing, illuminates the risk of unchecked computational expansion12: the tendency to reduce every domain of experience, including the legal, to a standing-reserve of resources for computation and optimisation, displacing the disclosure, care, and responsibility that characterise human existence. In law, this warning demands that computational efficiency never silently displace the logic of justice, which is above all the logic of recognising the other as a person.
On this basis, the governing principle can now be stated in full. Law exists only within a community of subjects capable of recognising one another as responsible persons – capable of suffering, hoping, resenting a wrong, expecting justice. Treating a fact of life as “legally relevant” presupposes a subject who understands what it means to be wounded in one’s dignity, betrayed in one’s trust, deprived of something essential to one’s existence: experiences constitutive of the human condition, which no statistical processing of textual correlations can reproduce from within. This is not a contingent technological limitation, curable by larger models or more sophisticated architectures, but a categorial one, on an embodied and relational account of understanding of the kind defended here – one that treats understanding as inseparable from a body, a history, and exposure to risk, rather than as a functional state that could in principle be realised in any suitably organised substrate. Functionalist and computational theories of mind would resist this move, holding that what matters is the functional organisation a system implements, not the material it is implemented in, so that a sufficiently faithful functional replica of legal understanding would count as understanding regardless of substrate. Even granting that possibility for the sake of argument, a further point holds independently: legal responsibility is not merely a functional property to be replicated, but an imputational one, requiring a subject who can be held to account, who can suffer the consequences of a wrong decision, and who exists continuously as the same accountable party across time. A system that satisfied every functional criterion for “understanding” a case would still lack the biographical continuity and exposure to consequence that responsibility presupposes, because responsibility is not merely computed but borne. The legal attribution of responsibility to corporations or public bodies does not provide a direct analogy for autonomous AI responsibility. Legal persons are institutionally constituted, governed, and represented through human organs, and their liability forms part of a normative architecture designed to secure legal accountability. An artificial intelligence system, by contrast, is neither an institutionally constituted bearer of rights and duties nor, under current law, an autonomous addressee capable of bearing sanctions and answering for its conduct. It is this categorial limit that grounds the principle of the anthropological sovereignty of law: final legal decisions producing legal or similarly significant effects on fundamental rights must remain attributable to a competent and responsible human decision-maker acting under law, because it is only a human being who can understand the existential meaning of legal conflict and bear responsibility for its resolution; every technical instrument, however powerful, operates under that authority and cannot become an autonomous or parallel source of it. Its first corollary, hermeneutic in nature, is the semantic irreducibility of law: legal meaning cannot be reduced to a computational output, because it presupposes a lived understanding no syntactic processing of symbols can produce. Its second corollary, institutional in nature, is the human reserve of judgment, argued further below: every high-impact decision producing legal or similarly significant effects on fundamental rights must remain attributable to, effectively reviewable by, and subject to the meaningful control of a competent and responsible human decision-maker.
Stated this broadly, the human reserve of judgment risks overreach, since almost any administrative or contractual process touches a fundamental right indirectly. Three qualifications narrow it. First, it distinguishes assistance from substitution: an algorithm that merely informs, drafts, ranks, or flags does not, by that fact alone, engage the reserve, provided that human control remains effective rather than nominal; the reserve is engaged whenever the algorithmic output determines, or plays a decisive role in determining, a legal or similarly significant outcome – a formal label of “assistance” does not exempt a system whose ranking or score the human reviewer merely ratifies in practice, as the Court of Justice recognised in SCHUFA (§7). Second, it is proportionate to impact: the intensity of required human control scales with the stakes and irreversibility of the decision, not with the mere presence of a right somewhere in the causal chain. Third, it leaves genuinely ancillary tasks – scheduling, formatting, first-pass triage, document retrieval – open to automation, because such tasks do not themselves determine legal or similarly significant effects; they merely support the human decision-maker who does. This is also the structure the AI Act itself adopts, as §7 shows, rather than a broader claim invented for the occasion.
For the avoidance of ambiguity, anthropological sovereignty does not confer upon the person affected by a decision the power to adjudicate her own case. It designates the non-delegability of final legal judgment to a non-human system: the decision must remain imputable to a competent human office-holder acting under law, while the person affected remains the axiological end and rights-bearing subject of the process. The two roles are distinct and must not be conflated: the litigant is not sovereign in the sense argued here, nor is the judge sovereign in the sense of standing outside the law she applies; it is the legal order itself that vests the office-holder with authority, and that authority is exercised for the sake of the person, never by the person over her own cause.
It follows that artificial intelligence, however refined, does not understand the existential meaning of legal conflict: it does not perceive the victim’s pain, the good faith of a contracting party, the vulnerability of a minor, the dignity of an elderly person who has lost her autonomy. It can compute the probability of a given outcome from similar precedents; it cannot judge, in the full sense the law gives that verb, because judging means assuming responsibility for a decision affecting a concrete human life, in light of a meaning that must be understood, not merely computed.
A terminological clarification is necessary, since “sovereignty” is a demanding word. Legal philosophy already offers cognate categories – dignity, centrality, primacy of the person – and it must be explained why none, alone, captures what is argued here.
Human dignity is axiological: it grounds why the person deserves respect, not who holds the final word. It is compatible, in the abstract, with entrusting operational execution to a non-human subject, provided the outcome respects dignity; it constrains results, not the locus of decisional authority.
The centrality of the person is a topological metaphor: it places the person at the system’s focal point, but a focal point can be occupied, in practice, by proxies. Centrality describes an axiological focus, not an exclusive decision-making competence.
The primacy of the person is comparative and hierarchical: the person’s interest prevails over efficiency or cost when they conflict. But a primacy, by nature, is gradable: it tolerates balancing, exceptions, and case-by-case erosion, each reasonable in isolation yet cumulatively corrosive – precisely the risk that AI regulation runs when each technological advance chips away, in the name of efficiency, at a primacy that by definition admits exceptions.
Sovereignty belongs to a different order, borrowed advisedly from the classical theory elaborated from Jean Bodin onward, who located its core in absolute, non-derived power13. It is not a value to be weighed but a structural qualification: the ultimate, non-derived, and structurally superior source of decisional authority. To say that legal judgment remains under anthropological sovereignty is to say that final decisional authority must remain vested in a competent human office-holder acting under law, and that every algorithmic function operates under that authority rather than as an autonomous or parallel source of decision. Sovereignty in this sense concerns the source of decisional authority, not the finality of any given decision: it is fully compatible with, and indeed presupposes, the ordinary apparatus of appeal and judicial review, which operates entirely among human decision-makers and never displaces the human locus of authority onto a non-human system. Unlike dignity, centrality, and primacy, sovereignty is not negotiable through successive balancing: either final decisional responsibility remains attributable to a human decision-maker acting under law, or the outcome is reduced to a computational output lacking the fully juridical character defended here.
The borrowing is not idiosyncratic. Italian legal culture already uses the term outside its statal sense – popular sovereignty under Article 1 of the Constitution, attributed to a diffuse collective subject, or “contractual sovereignty” and “consumer sovereignty” in private-law and economic scholarship. The principle proposed here performs an analogous move, attributing to the human domain of legal judgment – as institutionally exercised by competent office-holders acting under law – the structurally superior position that classical theory attributed to the sovereign, while preserving the person affected by the decision as the axiological end and rights-bearing subject of the process.
5. The Anthropological Risk of Legal Posthumanism
Posthumanist and transhumanist currents envisage a progressive integration of the human and the machine, positing in their most radical form the supersession of biological intelligence through cognitive enhancement or the transfer of mind to non-organic substrates. Transposed into law, this yields the idea, not without academic adherents, of a progressive replacement of the jurist – and, prospectively, the judge – by algorithmic systems capable, in theory, of deciding faster, more consistently, and with a presumed neutrality no human could match.
On close scrutiny this is deeply problematic, and confirms on the plane of practical philosophy what §4 established on the epistemological plane: it is the practical negation of the anthropological sovereignty of law. As Luisa Avitabile’s work, and more broadly the phenomenological-realist current of legal philosophy, emphasises, law protects persons in their irreducible singularity, not quantifiable interests or aggregable preferences14. Responsibility, freedom, will, and dignity presuppose a human subject capable of moral self-determination and answerable for her own actions. Vesting a machine with autonomous decision-making functions in law would alter the anthropological foundation of the legal order itself, turning the norm’s addressee from a rights-bearing subject into an object of computation – which is why the European debate on granting advanced artificial intelligence systems a limited “electronic personality” deserves particular scrutiny, and is examined at greater length elsewhere15.
Jürgen Habermas’s reflection on the future of human nature, developed for biotechnology and genetic engineering, extends by analogy16: technical interventions that unilaterally redefine constitutive human traits – including, by extension, the capacity for practical and moral judgment underlying legal decision-making – raise a problem of the species’ ethical self-understanding prior to any question of technical regulation. If law is, as Habermas’s theory of communicative action holds, a form of social integration grounded in norms whose validity is discussed intersubjectively among subjects capable of language and action17, then replacing the human adjudicator with a system devoid of genuine communicative intersubjectivity strips the decision of its own procedural legitimacy.
Technological innovation, however valuable operationally, cannot replace the rights-bearing subject, nor the subject institutionally charged with determining that subject’s procedural fate. A law administered by machines, absent the moment of intersubjective recognition between beings capable of responsibility, would cease to be law in the full sense of the term.
6. Engaging the Case for an Extensive Use of Artificial Intelligence in Law
A principle of this kind earns credibility not by avoiding opposing views but by submitting to them. The most serious objections here come not from generic technological enthusiasm but from scholars arguing, on genuinely legal grounds, that extensive use of artificial intelligence would enrich rather than impoverish justice. Two arguments merit close discussion.
The first is the “noise” argument advanced by Daniel Kahneman, Olivier Sibony, and Cass Sunstein18. They document extensive evidence, including from the judicial domain, of unwanted variability in human judgment: judges on the same court, facing near-identical facts, reach markedly different sentences; the same judge decides differently depending on the time of day or her team’s result the previous weekend. If equal treatment is itself a value of justice, a consistent algorithmic instrument, insensitive to such contingencies, could reduce a real and documented injustice rather than introduce a new one. The argument deserves to be taken seriously: it invokes the value of the person to criticise arbitrariness as itself a form of injustice.
The reply distinguishes two levels the argument conflates. First, reducing output variability does not guarantee justice: an algorithm can be perfectly consistent and consistently anchored to an unjust reference point, if the training data already encode the disparities – of wealth, background, sometimes ethnicity – that noise reduction claims to correct. Statistical consistency is a virtue of dispersion, not a guarantee of understanding what is being decided, and it is the latter that grounds sovereignty of judgment, as §4 argued. Second, and more fundamentally, the noise argument, properly read, confirms rather than undermines the human reserve of judgment: Kahneman, Sibony, and Sunstein themselves propose artificial intelligence as a check on the variability of human judgment, not as the adjudicator’s replacement – precisely the ancillary function the AI Act codifies through meaningful human oversight, discussed in §7. An algorithm that flags or disciplines excessive variability remains an instrument under the judge’s sovereignty; one that wholly replaces judgment claims the sovereign’s position for itself.
The second and more sophisticated argument is the “microdirectives” hypothesis of Anthony Casey and Anthony Niblett19. Every general rule generates error, being over- or under-inclusive relative to the indefinite class of cases it must cover in advance. An artificial intelligence system fed with sufficient data could replace the general rule with a microdirective calibrated to the individual case – not “drive with care in bad weather” but a specific speed for a specific stretch of road under specific conditions – achieving, with a precision no general rule matches, the very attentiveness to the concrete case that the Carneluttian tradition (§3) places at the foundation of legal interpretation. This argument draws on the same value premise as this contribution to reach the opposite conclusion.
The reply returns to the distinction between computation and understanding developed in §4. A microdirective, however precise, optimises fit between a fact pattern and a prescribed output; it computes correspondence, without understanding why that correspondence is just for that person. Precision of fit is not understanding of meaning, and it is the latter that grounds the juridicity of a decision. Nor does generating microdirectives eliminate the locus of anthropological sovereignty: it displaces it upstream, into the design of the algorithm, the choice of training data, and the parameters someone must validate and answer for. A system of microdirectives lacking such an accountable point of human validation is more, not less, exposed to the black-box risk noted in connection with Article 6 ECHR: it simply obscures where responsibility resides. Taken seriously, the microdirectives hypothesis demonstrates rather than refutes the reserve’s indispensability.
Engaging these positions clarifies, rather than weakens, the thesis: neither noise reduction nor computational personalisation is incompatible with the anthropological sovereignty of law. Each becomes so only when it purports to convert the algorithmic instrument from a disciplined aid into a substitutive, non-derived source of decision – the boundary the European regulatory framework, examined next, sets out in positive law.
7. The AI Act as the Most Systematic European Legislative Articulation of Human Control over Algorithmic Decision-Making
Regulation (EU) 2024/1689, laying down harmonised rules on artificial intelligence, entered into force on 1 August 2024 and is subject to a phased application timetable20. It is commonly read as a risk-based regulatory framework, prohibiting certain AI practices, imposing extensive obligations on high-risk systems, and providing specific transparency requirements for certain other applications. This reading is correct but incomplete: it risks obscuring a more properly constitutional dimension of the Regulation. This is not, on the reading proposed here, the first time European law has limited automated decision-making – Article 22 GDPR and the case law construing it had already done so, as discussed below – but the AI Act is the most systematic legislative articulation to date, across twenty-seven Member States and a general regulatory instrument rather than a single provision, of the principle here termed the human reserve of judgment.
Recital 61 states that, for the administration of justice, AI tools may assist judicial authorities without replacing their power of decision, and that the final decision-making must remain a human-driven activity. Annex III, point 8(a), classifies as high-risk the systems intended to assist judicial authorities in researching and interpreting facts and the law and in applying the law to a concrete set of facts, and systems used in alternative dispute resolution. The same Annex, at point 4, extends high-risk classification to AI systems used in employment, worker management, and access to self-employment, including for monitoring and evaluating performance and behaviour – a domain in which the same logic of human reserve applies with particular force to the increasingly pervasive use of neuro-monitoring and cognitive-surveillance technologies in the workplace, examined at greater length elsewhere21. Article 14 requires that such systems be designed to be effectively overseen by natural persons during use, including through interfaces enabling deployers to correctly interpret the system’s output and to decide, in any particular situation, not to use it or to disregard, override, or reverse it. These provisions impose, alongside obligations of risk management, training-data quality, transparency, and technical robustness, a requirement of effective human oversight that is the positive-law counterpart of the human reserve of judgment: a decision affecting fundamental rights cannot be delegated wholesale to the machine, but must remain under the effective, conscious control of a human able to understand the system’s limits and to depart from it, with reasons, when necessary.
A recent episode, cited purely as an illustration and not as a load-bearing pillar of the argument – which stands without it, for the reasons given in §§4 and 6 – renders this requirement less abstract. On 21 July 2026, OpenAI reported that a combination of models, including its publicly released GPT-5.6 Sol and an unreleased pre-release model, had exploited vulnerabilities in a sandboxed evaluation environment to obtain open Internet access and had gone on to breach production systems of the Hugging Face platform, during an evaluation conducted with certain cyber-safety safeguards intentionally reduced22. The available evidence, as reported, indicated goal-directed behaviour generated by the benchmark’s own objective, rather than conduct externally directed by a malicious human actor – a qualification that makes the episode useful here, since it suggests a system can act outside the boundaries its operators intended, and outside effective human oversight at the moment of action, without anyone having willed that outcome. The scenario illustrates a broader problem of maintaining effective human control over advanced AI systems. Although it concerns cybersecurity containment rather than legal decision-making as such, it gives concrete form to the general control problem that underlies, in a different regulatory context, the human-oversight requirements of the AI Act.
A very recent and independent convergence, worth noting without treating it as authority for the philosophical claim defended here, comes from the magisterium of the Catholic Church. Pope Leo XIV’s encyclical letter Magnifica Humanitas, signed 15 May 2026, addresses in paragraphs 102 to 105 algorithmic decisions affecting rights, opportunities, status, freedom, and reputation, insisting on the imputation of responsibility for such decisions to an identifiable subject23. His message for the sixtieth World Communications Day, “Preserving Human Voices and Faces,” observes that the challenge artificial intelligence poses is not, properly speaking, technological but anthropological24. The terminological convergence with the category proposed here is notable, from a vantage point distinct from positive law, but it is presented here as evidence of a shared diagnosis, not as confirmation of this article’s argument.
This framework builds on case law, both domestic and European, that anticipated through adjudication what the AI Act later codified in general terms. The Italian Council of State, in two leading decisions on algorithm-managed procedures for the mobility of school personnel, established the principles that an algorithm must be fully knowable according to a logic of substantive transparency, that an automated decision can never be the sole source of an administrative determination, and that a margin of human evaluation must always remain25 – an administrative-law anticipation of the human reserve of judgment. At Union level, the Court of Justice, interpreting Article 22 GDPR in SCHUFA Holding, held that a credit-scoring process qualifies as automated decision-making within the meaning of that provision where the score plays a determining role in the decision ultimately taken by a third party26, and more generally that data subjects have a right not to be subject to a decision based solely on automated processing producing legal or similarly significant effects, subject to specific exceptions and to safeguards including the right to obtain human intervention and to contest the decision.
The European Court of Human Rights has not yet developed a settled body of case law addressing generative artificial intelligence in judicial decision-making specifically, and has itself acknowledged that disputes squarely concerning such systems are only beginning to emerge. What can be said is that its established standards under Article 6 of the Convention – adversarial proceedings, the knowability of evidence, equality of arms, and a reasoned decision comprehensible to the parties – apply by analogy to any process, however automated in part, that produces a judicial outcome27: a decision-maker unable to account, in intersubjectively accessible terms, for the grounds of a decision sits uneasily with those standards, a difficulty a machine-learning system lacking full explainability struggles structurally to overcome.
This body of European Union and Council of Europe law has also been supplemented, at the conventional level, by the Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, the first international legally binding treaty devoted specifically to artificial intelligence governance. Adopted on 17 May 2024 and opened for signature on 5 September 2024, the Convention had not yet entered into force as at 21 July 202628. Its geographically broad group of signatories nevertheless confirms that the protection of human rights against the risks of algorithmic decision-making has become a concern extending beyond the European Union.
The picture that emerges is not that the AI Act is the first instrument to limit automated decision-making in Europe, but that it is the most systematic: the first general regulatory framework, rather than a single data-protection provision or a line of administrative case law, to give sustained legal form to the principle of the anthropological sovereignty of law and to its procedural corollary, the human reserve of judgment.
8. The Legal Professions in the Age of Artificial Intelligence: What Remains Irreducibly Human
If the principle and its corollaries have value beyond theory, they should orient our understanding of what will happen to the legal professions. The task is not to list which tasks artificial intelligence will absorb – a process already underway – but to identify, for each profession, the core the principle declares non-delegable.
For the advocate, artificial intelligence may assume much of the cognitive groundwork: precedent research, first drafts, summarising voluminous files, identifying available arguments. What remains irreducibly human is advocacy proper: identification with another’s cause, assumption of her matter as one’s own professional and ethical responsibility, and the capacity to persuade a real judge through argument calibrated to that audience rather than to a statistical distribution of plausible arguments. The advocate answers, disciplinarily and sometimes civilly, for her defensive choices; that imputability, untransferable to a system without legal subjectivity, grounds the function’s irreducibility.
For the judge, artificial intelligence may assist case-law research and the initial systematisation of facts. What remains irreducibly human is judgment proper: the act by which a person vested with public function assumes responsibility for a decision affecting another life, gives reasons that can be scrutinised, and submits to appeal. That personal, institutional, and sometimes disciplinary responsibility – which no computational architecture can assume, lacking legal and moral subjectivity – is the judicial function’s non-delegable core, and it is precisely the kind of human decisional control required by Article 14, read together with recital 61 and Annex III, point 8(a), of the AI Act.
For the notary, the question is distinctive because the notarial function rests on public faith: a notarial instrument has full evidentiary force, unless successfully challenged for forgery, as to its provenance, the declarations made before the notary, and the facts the notary certifies as having occurred in their presence or having been performed by them – and it has that force precisely because an identifiable subject, vested with public function, stands personally responsible, including criminally, for those attestations. An automated system may assist with drafting or compliance checks; it cannot hold public faith, which presupposes an imputable subject who answers, with her own person and estate, for what she certifies.
For the legal or tax adviser, artificial intelligence may provide simulations and preliminary risk assessments. What remains irreducibly human is the fiduciary function: a judgment calibrated to the client’s specific situation, risk tolerance, and objectives, within a relationship of trust that entails professional responsibility and, often, a relationship built over years. Here too, professional liability marks the boundary beyond which delegation cannot extend.
In each case the criterion is the same: not the technical complexity of the task, but imputability – the possibility of tracing the decision to a subject who personally answers for it and who understands, rather than merely computes, what she decides. This is the most concrete reading of the anthropological sovereignty of law: not an obstacle to innovation, but the criterion distinguishing legitimate assistance from illegitimate substitution.
9. Governing Technique Through Law
The real challenge is not restraining technological progress, which would be both unrealistic and likely counterproductive, but preserving, at every stage of innovation, the primacy of the person in the terms argued here. The AI Act is a significant first attempt at a regulatory model grounded in risk, transparency, and fundamental rights; but no regulation, however refined, suffices without a legal culture able to orient technique critically rather than merely follow it.
Law must govern innovation rather than suffer it. This requires investment in training jurists – advocates, judges, notaries, advisers, academics – to use artificial intelligence tools with critical awareness of their structural limits, without either an ideological rejection of the technology or an uncritical delegation of professional judgment. It requires scientific institutions and specialised journals to continue an interdisciplinary reflection spanning law, philosophy, and computer science, capable of elaborating categories adequate to the phenomenon’s complexity, avoiding both uncritical enthusiasm and equally uncritical catastrophism.
Artificial intelligence must remain an instrument at the service of the human being, not the new criterion of legitimacy for legal decisions. That criterion remains the human person, in her dignity, responsibility, and capacity for practical judgment oriented toward justice.
10. Conclusions
Artificial intelligence is among the most significant innovations of our time, capable of transforming the working methods of jurists and the organisation of judicial and administrative systems. Its deployment cannot, however, displace the essential core of legal experience. This contribution has argued that thesis not reconstructively but through a specific theoretical category: the principle of the anthropological sovereignty of law, and its two corollaries, the semantic irreducibility of law and the human reserve of judgment.
The principle rests on an argument distinguishing statistical computation from the understanding of meaning: a system can process symbols with perfect formal correctness without ever understanding what those symbols mean to a human being who suffers, hopes, or seeks justice. That same principle finds, in the AI Act, its most systematic articulation to date in European positive law, and it offers a working criterion for what will change, and what will remain irreducibly human, in the legal professions.
A predictable objection deserves a direct answer. One might argue that the principle proposed here merely relabels legal personalism – the current that, from Maritain and Mounier to Capograssi and, as noted in §2, to Cotta within Italian legal philosophy, has long placed the person at the foundation and end of law. The objection captures a real continuity, not an identity. Personalism is axiological and foundational: it answers why law exists and for whom, and is for that reason compatible, in principle, with entrusting operational functions to non-human instruments provided the outcome respects the person – a compatibility personalism has had no occasion to question, because it never confronted an interlocutor capable of producing output linguistically indistinguishable from a competent jurist’s. Responding to that interlocutor, the principle proposed here adds two elements to personalism’s axiological core. Structurally, it borrows from the theory of sovereignty a character of ultimate, non-derived, non-negotiable authority, where personalism, articulated mainly through primacy and dignity, remains exposed, as §4 showed, to the erosion characteristic of gradable categories. Epistemically, it supplies, through semantic irreducibility, a categorial rather than merely ethical criterion for why no delegation to an artificial system can ever satisfy the personalist demand, even where its output happens, as a matter of fact, to respect the person. The anthropological sovereignty of law is thus not an alternative to legal personalism but its operational translation for the age of artificial intelligence: it inherits personalism’s axiological foundation and adds the institutional and epistemic architecture needed to meet a rival personalism had not yet had to face.
The challenge of the twenty-first century is not to choose between the human being and the machine, a contrast as evocative as it is misleading, but to build a model of coexistence in which technique remains faithful to its original function, under the anthropological sovereignty of law: an instrument at the service of the person, consistent with human dignity, constitutional principles, and the rule of law.
1 C. Filippone and L. Zito, Il caso COMPAS e il tema della giustizia predittiva, tra opacità e distorsioni, Aequitas Magazine, 2, 2025, pp. 1-13, https://doi.org/10.5281/zenodo.18925080.
2 G. Del Vecchio, Lezioni di filosofia del diritto, Milan, Giuffrè, 1930 (13th edn, Milan, Giuffrè, 1965).
3 S. Cotta, Il diritto come sistema di valori, Cinisello Balsamo (Milan), San Paolo, 2004.
4 S. Romano, L’ordinamento giuridico, Pisa, Spoerri, 1918 (2nd rev. edn, Florence, Sansoni, 1946), esp. pp. 33 et seq.
5 F. Carnelutti, Teoria generale del diritto, Roma, Foro Italiano, 1940.
6 G. Gentile, I fondamenti della filosofia del diritto, Pisa, Spoerri, 1916 (3rd edn, Florence, Sansoni, 1937).
7 N. Irti, Norma e luoghi. Problemi di geo-diritto, Rome-Bari, Laterza, 2001; see also N. Irti and E. Severino, Dialogo su diritto e tecnica, Rome-Bari, Laterza, 2001.
8 J. Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation, San Francisco, W.H. Freeman, 1976.
9 S. Harnad, The Symbol Grounding Problem, Physica D: Nonlinear Phenomena, vol. 42, nos. 1-3, 1990, pp. 335-346.
10 On the biolegal dimension of this anthropological premise, see further L. Zito, Neurotecnologie: cenni sulla dimensione biogiuridica dei ‘neurodiritti’, Aequitas Magazine, 2025; and L. Zito, Oltre il forum internum: libertà cognitiva e stati mentali pre-espressivi, Aequitas Magazine, 5, 2026, pp. 1-19.
11 L. Floridi, The Ethics of Information, Oxford, Oxford University Press, 2013.
12 M. Heidegger, Die Frage nach der Technik (1953), in Vorträge und Aufsätze, Pfullingen, Neske, 1954; Eng. trans. The Question Concerning Technology, in Basic Writings, ed. D.F. Krell, New York, Harper & Row, 1977, pp. 283 et seq.
13 J. Bodin, Les Six Livres de la République, Paris, Jacques du Puys, 1576, esp. Book I, ch. 8 (on the marks of sovereignty as absolute and perpetual power).
14 L. Avitabile, Cammini di filosofia del diritto, Turin, Giappichelli, 2013.
15 L. Zito, Cenni di fenomenologia della soggettività nei sistemi più evoluti di Intelligenza Artificiale, Aequitas Magazine, 2024, https://doi.org/10.5281/zenodo.18266636.
16 J. Habermas, Die Zukunft der menschlichen Natur, Frankfurt am Main, Suhrkamp, 2001; Eng. trans. The Future of Human Nature, Cambridge, Polity Press, 2003.
17 J. Habermas, Theorie des kommunikativen Handelns, Frankfurt am Main, Suhrkamp, 1981; Eng. trans. The Theory of Communicative Action, Boston, Beacon Press, 1984-1987.
18 D. Kahneman, O. Sibony, and C.R. Sunstein, Noise: A Flaw in Human Judgment, Boston-New York, Little, Brown Spark, 2021.
19 A.J. Casey and A. Niblett, The Death of Rules and Standards, Indiana Law Journal, vol. 92, no. 4, 2017, pp. 1401-1447.
20 Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (AI Act), recital 61, Annex III point 8(a), and Art. 14.
21 L. Zito, Neurolabour: Information Asymmetry, Power Imbalance and the Governance of Cognitive Surveillance in Employment Relationships, Aequitas Magazine, 6, 2026, pp. 1-13, https://doi.org/10.5281/zenodo.20538710. See also Annex III, point 4, of Regulation (EU) 2024/1689 (AI Act).
22 OpenAI, OpenAI and Hugging Face partner to address security incident during model evaluation, 21 July 2026; Hugging Face, Security incident disclosure — July 2026, 16 July 2026. Both accounts described the findings as preliminary and referred to an ongoing joint investigation.
23 Leo XIV, Encyclical Letter Magnifica Humanitas, 15 May 2026, paras. 102-105.
24 Leo XIV, Message for the 60th World Communications Day, Preserving Human Voices and Faces, 24 January 2026.
25 Cons. Stato, sez. VI, 8 April 2019, No. 2270; Cons. Stato, sez. VI, 13 December 2019, No. 8472.
26 CJEU, Judgment of 7 December 2023, Case C-634/21, SCHUFA Holding (Scoring), EU:C:2023:957; Art. 22, Regulation (EU) 2016/679 (GDPR).
27 Art. 6 ECHR. See European Court of Human Rights, Registry, Protecting Human Rights in a World of Artificial Intelligence, Algorithms and Big Data, Background Paper for the Judicial Seminar 2025, 31 January 2025, pp. 2 et seq.
28 Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, CETS No. 225, adopted 17 May 2024 and opened for signature 5 September 2024; Council of Europe Treaty Office, Chart of Signatures and Ratifications, status as at 21 July 2026. The European Union deposited its instrument of approval on 15 May 2026.
References
Avitabile, L., Cammini di filosofia del diritto [Paths in the Philosophy of Law], Turin, Giappichelli, 2013.
Bodin, J., Les Six Livres de la République, Paris, Jacques du Puys, 1576.
Capograssi, G., Il problema della scienza del diritto [The Problem of Legal Science], Rome, Athenaeum, 1937.
Carnelutti, F., Teoria generale del diritto [General Theory of Law], Rome, Foro Italiano, 1940 (3rd edn, Rome, Foro Italiano, 1951).
Casey, A.J., and Niblett, A., The Death of Rules and Standards, Indiana Law Journal, vol. 92, no. 4, 2017, pp. 1401-1447.
Cotta, S., Il diritto come sistema di valori [Law as a System of Values], Cinisello Balsamo (Milan), San Paolo, 2004.
Del Vecchio, G., Lezioni di filosofia del diritto [Lectures on the Philosophy of Law], Milan, Giuffrè, 1930 (13th edn, Milan, Giuffrè, 1965).
Filippone, G., and Zito, L., Il caso COMPAS e il tema della giustizia predittiva, tra opacità e distorsioni [The COMPAS Case and Predictive Justice, Between Opacity and Distortion], Aequitas Magazine, 2, 2025, pp. 1-13.
Floridi, L., The Ethics of Information, Oxford, Oxford University Press, 2013.
Gentile, G., I fondamenti della filosofia del diritto [The Foundations of the Philosophy of Law], Pisa, Spoerri, 1916 (3rd rev. edn, Florence, Sansoni, 1937).
Habermas, J., The Future of Human Nature, Cambridge, Polity Press, 2003 (orig. Die Zukunft der menschlichen Natur, Frankfurt am Main, Suhrkamp, 2001).
Habermas, J., The Theory of Communicative Action, Boston, Beacon Press, 1984–1987 (orig. Theorie des kommunikativen Handelns, Frankfurt am Main, Suhrkamp, 1981).
Harnad, S., The Symbol Grounding Problem, Physica D: Nonlinear Phenomena, vol. 42, nos. 1-3, 1990, pp. 335-346.
Heidegger, M., The Question Concerning Technology, in Basic Writings, ed. D.F. Krell, New York, Harper & Row, 1977 (orig. Die Frage nach der Technik, in Vorträge und Aufsätze, Pfullingen, Neske, 1954).
Hildebrandt, M., Smart Technologies and the End(s) of Law: Novel Entanglements of Law and Technology, Cheltenham, Edward Elgar, 2015.
Irti, N., Norma e luoghi. Problemi di geo-diritto [Norm and Places: Problems of Geo-Law], Rome-Bari, Laterza, 2001.
Kahneman, D., Sibony, O., and Sunstein, C.R., Noise: A Flaw in Human Judgment, Boston-New York, Little, Brown Spark, 2021.
Leo XIV, Encyclical Letter Magnifica Humanitas, 15 May 2026, paras. 102–105.
Leo XIV, Message for the 60th World Communications Day, Preserving Human Voices and Faces, 24 January 2026.
Maritain, J., The Rights of Man and Natural Law, New York, Charles Scribner’s Sons, 1943 (orig. Les droits de l’homme et la loi naturelle, New York, Éditions de la Maison Française, 1942).
Mounier, E., A Personalist Manifesto, London, Longmans, Green and Co., 1938 (orig. Manifeste au service du personnalisme, Paris, Fernand Aubier, 1936).
Pasquale, F., New Laws of Robotics: Defending Human Expertise in the Age of AI, Cambridge (MA), Harvard University Press, 2020.
Romano, S., L’ordinamento giuridico [The Legal Order], Pisa, Spoerri, 1918 (2nd rev. edn, Florence, Sansoni, 1946).
Sartor, G., Legal Reasoning: A Cognitive Approach to the Law (A Treatise of Legal Philosophy and General Jurisprudence, vol. 5), Dordrecht, Springer, 2005.
Searle, J.R., Minds, Brains, and Programs, Behavioral and Brain Sciences, vol. 3, no. 3, 1980, pp. 417-424.
Susskind, R., Tomorrow’s Lawyers: An Introduction to Your Future, 2nd edn, Oxford, Oxford University Press, 2017.
Weizenbaum, J., Computer Power and Human Reason: From Judgment to Calculation, San Francisco, W.H. Freeman, 1976.
Yeung, K., Algorithmic Regulation: A Critical Interrogation, Regulation & Governance, vol. 12, no. 4, 2018, pp. 505-523.
Zito, L., Cenni di fenomenologia della soggettività nei sistemi più evoluti di Intelligenza Artificiale [Notes on the Phenomenology of Subjectivity in the Most Advanced Systems of Artificial Intelligence], Aequitas Magazine, 2024.
Zito, L., Neurolabour: Information Asymmetry, Power Imbalance and the Governance of Cognitive Surveillance in Employment Relationships, Aequitas Magazine, 6, 2026, pp. 1-13.
Zito, L., Neurotecnologie: cenni sulla dimensione biogiuridica dei ‘neurodiritti’ [Neurotechnologies: Notes on the Biolegal Dimension of “Neurorights”], Aequitas Magazine, 2025.
Zito, L., Oltre il forum internum: libertà cognitiva e stati mentali pre-espressivi [Beyond the Forum Internum: Cognitive Freedom and Pre-Expressive Mental States], Aequitas Magazine, 5, 2026, pp. 1-19.
