“I have run away, hold me. When you have brought me back to my master Zoninus, you shall receive a gold coin,” read the collar of a Roman slave. Who holds the collar of Artificial Intelligence today? And, above all, does it still make sense to look for one?

In this technological interregnum, the European Union is conducting a complex regulatory overhaul, often labeled as Lawfare. Purging the analysis of superficial romanticism, this architecture does not represent a harmonious connection or a purely meditative counterbalance, but rather the product of a sharp friction between the need to preserve human dignity ^[1]Charter of Fundamental Rights of the EU (CFR), Article 1 and the terror of continental structural obsolescence, made explicit by the urgencies of competitiveness ^[2]The Draghi report on EU competitiveness.

The epistemological leap from generative architectures to Agentic Artificial Intelligences (Agentic AI) disintegrates the traditional perimeters of imputability. If an AGI assumes the capability to scale and govern assets with total autonomy ^[3]Jensen Huang Says AGI is Here, and It Can Run Companies Too!, https://www.mitsloanme.com/article/jensen-huang-says-agi-is-here-and-it-can-run-companies-too/, we witness a radical dissociation between capital and the exercise of enterprise. Faced with networks of neural micro-transactions capable of moving billions in fractions of a second ^[4]The Most Important Thing Jensen Huang Said at GTC 2026 Wasn’t About a Chip – eWeek, https://www.eweek.com/news/nvidia-inference-ai-economy-agents-gtc-2026/, human directors collide with a new frontier of culpa in vigilando: the material impossibility of monitoring hyper-fast and opaque synthetic choices ^[5]NVIDIA CEO Jensen Huang Claims AGI Achieved: What It Means and Market Impact – Phemex, https://phemex.com/blogs/jensen-huang-agi-claims-critics-market-signals. In this scenario, the law must abandon reassuring fictions and face engineering reality head-on.

I – The Pragmatism of the Segregated Estate

To decode the civil liability of Embodied AI, a leap backward led to the resurrection of the noxal action (actio noxalis). In Roman law, if a servus caused damage, the dominus could free himself by surrendering the culprit (noxal surrender) or by paying for the damages.

However, translating this legal institution to contemporary AI constitutes a dangerous ontological and economic fallacy. The Roman slave, albeit classified as res, was biologically human, endowed with free will and, crucially, mens rea (intentionality). The algorithm, although simulating complex reasoning with I.Q. scores of 101 ^[6]Spiceworks: Top Three LLMs Compared: GPT-4 Turbo vs. Claude 3 Opus vs. Gemini 1.5 Pro, https://www.spiceworks.com/tech/artificial-intelligence/articles/top-llms-compared/, optimizes mathematical functions. It does not “choose” to commit a tort. Treating AI as a slave amounts to falling into the “Mechanical Turk” trap ^[7]The Turk (automaton) – Wikipedia.

From an economic standpoint, allowing a “synthetic noxae deditio“—limiting the corporation’s liability to the mere value of the software or robot (like the Atlas models ^[8]Boston Dynamics Official Website: https://bostondynamics.com/)—would be systemic suicide. It would allow multi-billion dollar damages generated by autonomous algorithms to be externalized onto society.

Roman law offers us a far clearer lens: the Peculium (a segregated estate intended for the exercise of a business) and the actio institoria. Through the actio institoria, the principal (dominus negotii) held unlimited liability for obligations contracted by the business manager, by virtue of the mere appointment (having inserted him into his own production cycle). The machine is not punished; rather, the one who profits from it (the cheapest cost avoider) is held responsible. Translated to 2026, the modern legislator must structure compulsory insurance systems and compensatory guarantee funds: a bound digital peculium wherein strict liability is equipped with insurance ceilings proportional to the actual financial volumes moved by neural networks. The modern legislator must look to compulsory insurance systems and guarantee funds (a bound digital peculium), where strict liability is inextricably linked to the financial volumes moved by neural networks, internalizing the social costs of innovation.

The EU AI Act ^[9]European Parliament, P9_TA(2024)0138, Artificial Intelligence Act, europa.eu, https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 has mapped out risk levels, but the real civil game is played on the field of liability.

II – Product Liability and Deregulation

The European regulatory review (the Digital AI Omnibus) is not a moment of calm reflection, but the outcome of pressing industrial pushback. The postponement of compliance for high-risk systems to December 2027 ^[10]The Digital AI Omnibus: Proposed deferral of high risk AI obligations under the AI Act, … Continue reading masks a retreat aimed at lightening the burden on enterprises, risking the hollowing out of original safeguards ^[11]The EU AI Act: A Story without a Plot? – A Lack of Vision in the Digital AI Omnibus Proposal threatens constitutional AI Governance – ResearchGate, … Continue reading. However, the irreconcilable asynchrony between industrialists and guarantors led, precisely on April 28, 2026, to the dramatic failure of the political trilogue on the AI Omnibus. Without an agreement, the original August 2026 deadlines remain in force, leaving the industry (especially high-risk systems integrated into regulated products) in a jurisprudential void ^[12]EU AI Act Omnibus: The Trilogue Failed, What Happens to the August 2026 Deadline? – Modulos, https://www.modulos.ai/blog/ai-act-omnibus-trilogue-failed/.

Will this asynchrony collide with the entry into force of the Product Liability Directive (PLD) at the end of 2026?

The repeal of the AILD Directive ^[13]AI Commission Confirms that SEP Regulation, AI Liability Directive are Officially Scrapped, … Continue reading shifts the compensatory burden onto the PLD, which classifies the algorithm as a “product” subject to strict liability ^[14]Revised Product Liability Directive – European Parliament, https://www.europarl.europa.eu/RegData/etudes/BRIE/2023/739341/EPRS_BRI(2023)739341_EN.pdf. While for consumers the wall of the black box is torn down ^[15]Artificial Intelligence and Civil Liability – IRIS, https://www.iris.sssup.it/bitstream/11382/585714/1/IUST-STU%282025%29776426-EN.pdf, an existential crisis looms for the Open Source ecosystem: exposure to million-dollar lawsuits risks triggering an inexorable chilling effect on bottom-up innovation ^[16]Product Risks Today: How the new Product Liability Directive turns AI Act compliance into a question of liability – Freshfields, … Continue reading. The recent “GNOPPIX Case,” in which a European volunteer project relocated its servers to Japan precisely to escape the uncertain grip of “strict liability” applied to shared code, demonstrates how the PLD risks handing the market exclusively over to mega-vendors capable of absorbing insurance costs ^[17]EU Digital Laws Spark FOSS Liability Fears: The GNOPPIX Case – LinuxCareers, … Continue reading.

Even more naive is the claim to resolve the opacity of humanoid robots in the real world through disclosure obligations and “continuous physical markers” (e.g., LEDs or acoustic warnings). As demonstrated by HCI (Human-Computer Interaction), this does not generate awareness, but rather disclosure fatigue: users will ignore the signals just as they do with cookie banners. Worse still, it will trigger an Implicit Truth Bias: any interaction lacking the marker will be mistakenly deemed safe or “human,” leaving users vulnerable to tactical manipulations. Formal compliance is not synonymous with safety by design.

III – The Return of Defensive (Algorithmic) Medicine?

In the healthcare sector, the advent of AI robots promises exceptional progress, but the national legislative framework (Law 132/2025) reveals structural cracks dictated by the anxiety to preserve anthropocentrism at all costs. Requiring the physician to act as the “ultimate guarantor” (Human-in-the-Loop), forcing them to critically evaluate the output of a black box that cross-references millions of clinical parameters in milliseconds, is a legal utopia ^[18]È in vigore la legge italiana sull’intelligenza artificiale: le implicazioni per il settore sanitario, … Continue reading. The professional materially lacks the tools to refute the machine’s probabilistic calculation.

The core of the normative friction lies in the short circuit between Art. 7 of Law no. 132/2025 (which imposes on the doctor the final word on the AI’s “non-binding suggestions”) and the “double track” of the Gelli-Bianco Law (Law 24/2017). The Gelli-Bianco assigns contractual liability to the healthcare facility, leaving the employee physician with extra-contractual tort liability (with the shield of Art. 590-sexies of the Criminal Code valid only in the presence of codified guidelines).

If the Health Directorate mandates the use of an AI Decision Support system to cut costs and increase throughput (e.g., forcing a radiologist “with AI” to report Nx3 patients per day instead of N “without AI”), the workload falls outside the perimeter of human control. With equal time and human resources, the radiologist cannot physically re-examine the radiological voxels analyzed by the neural network for every single report. In the event of a false negative (for instance, a tumor lesion missed by the AI and endorsed by the doctor due to lack of time), the facility will respond civilly under the hospital contract (contratto di spedalità), but the physician will suffer the penal and extra-contractual aggression. They will not be able to invoke the shield of Art. 590-sexies of the Criminal Code, since there are currently no consolidated “clinical-assistance guidelines” on how and to what extent to interpret the errors of a generative or predictive intelligence.

The risk, driven by potential sanctions for gross negligence ^[19]AI e odontoiatri: dal 10 ottobre obbligo di informare i pazienti – Dentista Manager, https://www.dentistamanager.it/ai-e-odontoiatri-dal-10-ottobre-obbligo-di-informare-i-pazienti/, is the genesis of Algorithmic Defensive Medicine. Squeezed between the fear of deviating from hospital-certified best practices ^[20]Presidenza del Consiglio dei Ministri – Comitato Nazionale per la Bioetica, https://bioetica.governo.it/media/dpzffvau/p-157_2026-relazione-di-cura-nellera-della-ia.pdf and the impossibility of fully understanding the software, the clinician will develop either a dangerous automation bias (passively validating every error to avoid assuming direct risks) or an obstructive rejection of the technological medium. Even taking into account the interposition of the healthcare facility under the framework of the Gelli-Bianco Law ^[21]Law no. 24/2017, shifting algorithmic engineering defects onto the personal liability of the physician does not protect the patient; rather, it risks turning professionals into the scapegoats of multinational AI product developers ^[22]Legge 23 settembre 2025, n. 132 – Gazzetta Ufficiale, https://www.gazzettaufficiale.it/eli/id/2025/09/25/25G00143/sg ^[23]Il diritto penale dell’intelligenza artificiale nella legge 132/2025: frammentazione sistematica e tensioni costituzionali, … Continue reading.

IV – Strategic Realism and Dual-Use

Finally, the military sector’s exemption from the logic of the EU AI Act deserves examination. Dismissing it as a “forcing” of the Geneva Conventions ^[24]Lethal Autonomous Weapon Systems | United Nations Office for Disarmament Affairs, https://disarmament.unoda.org/en/our-work/emerging-challenges/lethal-autonomous-weapon-systems means ignoring its strategic realism. In an era of global hegemonic competition, binding the European Defense Industrial Base to exhausting human rights assessments would have meant continental tactical suicide.

The true paradox is not the absence of rules, but the bureaucratic illusion of being able to cleanly separate “civilian” from “military” use. AI is inherently Dual-Use. The dispute between Anthropic and the Pentagon is proof of this: faced with defense imperatives, the pacifist clauses of private policies evaporate ^[25]Europe’s AI Blind Spot: What the Anthropic-Pentagon Dispute Reveals, https://us.boell.org/en/2026/04/09/europes-ai-blind-spot-what-anthropic-pentagon-dispute-reveals ^[26]The Pentagon blacklisted Anthropic for opposing killer robots. Europe must respond., https://www.epc.eu/publication/the-pentagon-blacklisted-anthropic-for-opposing-killer-robots-europe-must-respond/. Without a stringent global treaty ^[27]Killer Robots: New UN Report Urges Treaty by 2026 | Human Rights Watch, https://www.hrw.org/news/2024/08/26/killer-robots-new-un-report-urges-treaty-2026, decision traceability ^[28]Additional Protocol I to the Geneva Conventions, Art. 36 risks remaining an empty shell. The moment European military infrastructures rely on foundational models developed abroad, the idea of maintaining “meaningful human control” becomes a legal fiction aimed at justifying the automated kill-chain ^[29]Lethal Autonomous Weapons Systems & International Law: Growing Momentum Towards a New International Treaty – ASIL, https://asil.org/insights/volume-29-issue-1/.

V – Conclusions

The rise of polyvalent and embodied AI compels us to abandon consolatory narratives. The European AI law is not an unshakable ethical beacon, but an imperfect shield forged in geopolitical and industrial emergency. The collar of the Roman slave ^[30]Il collare di uno schiavo — Google Arts & Culture belongs to a past where fault resided in the servant; today, the winning legal paradigm is not the abandonment of the robot, but execution against the corporate peculium.

Only by accepting the inevitability of strict liability, repudiating facade bureaucracies, and protecting human professionals, can we build a sustainable ecosystem. Preventive Legal Design ceases to be a bureaucratic frill and becomes the fundamental strategic asset for companies to govern risk and thrive in the century of synthetic cognition.

Avvocato Giovanni Paolo Sperti