The thesis aims to assess whether Directive (EU) 2024/2853 on liability for defective products is suitable to address potential harm connected to artificial intelligence (AI) systems. The research is situated within an operational analysis of product liability and offers one of the first systematic and comprehensive readings of the new Directive as applied to AI systems. The increasing deployment of AI systems - often software-based, opaque, interactive, and embedded within complex value chains - has made the recalibration of the liability regime both urgent and widely advocated in legal scholarship. While the Directive remains structurally aligned with Directive 85/374/EEC - preserving a defect-based strict liability regime and pursuing the same objectives of harmonisation and fair risk allocation - it introduces targeted adjustments for the digital age, intended to overcome outdated definitions, excessive evidentiary burdens, and procedural constraints. The outcomes of this reform have not yet been comprehensively examined. The thesis is structured around three research questions: (i) which solutions the new Directive provides to overcome the previous limitations in its application to AI systems; (ii) whether such solutions effectively achieve the reform’s objectives of harmonisation and fair risk distribution; and (iii) which de iure condendo approaches may resolve recurring interpretative problems. The exegesis of the Directive is conducted according to the methodological canons of Italian private law, with particular emphasis on systematic and teleological interpretation, complemented by historical and comparative analysis, especially with reference to Italy, France, and Germany. While not intended as an exhaustive commentary on prior doctrinal and case-law debates, the thesis revisits key interpretative turning points of the former liability regime. In evaluating the effectiveness of the reform, the thesis adopts the European Commission’s legislative assessment criterion of effectiveness—understood as coherence between stated objectives and actual outcomes—and incorporates insights from law and economics as well as from data science and computer science literature. Following the articulated research questions and methodological framework, the thesis reaffirms the strict liability nature of product liability, clarifies the attribution criteria (defect, damage, and causal link between defect and damage), and delineates the role and limits of evidentiary presumptions. The research reaches two principal conclusions. First, it assesses the effectiveness of Directive (EU) 2024/2853 with respect to AI systems as only partial: although the introduced adjustments improve definitional clarity and access to evidence, significant shortcomings remain, particularly concerning the identification of the proper defendant and the governance of presumptions. Second, the thesis proposes a contextualisation of performance expectations that anchors the assessment of defectiveness to the function/application dyad, allowing courts to infer defect where, within the relevant operational context, the system exhibits error or incident rates that are consistently higher than those of comparable systems or of a human operator, based on verifiable data.
La tesi ha l’obiettivo di verificare se la Direttiva (UE) 2024/2853 sulla responsabilità per danno da prodotti difettosi sia idonea a fronteggiare possibili danni connessi a sistemi di intelligenza artificiale. Il lavoro si inserisce nell’analisi operativa della responsabilità da prodotto e offre uno dei primi tentativi organici di lettura complessiva della Nuova Direttiva applicata ai sistemi di IA. La crescente diffusione di sistemi di IA, prodotti a base software, opachi, interagenti e inseriti in catene del valore complesse, ha infatti reso urgente, nonché ampiamente invocata dalla letteratura, la rimodulazione del regime. Infatti, se da un lato la direttiva si pone in continuità con la precedente Direttiva 85/374/CEE – in quanto mantiene la responsabilità oggettiva centrata sul difetto e persegue gli stessi fini di armonizzazione ed equa ripartizione dei rischi – dall’altro introduce correttivi mirati per l’era digitale, volti a superare l’obsolescenza delle definizioni, la gravosità dell’onere probatorio e varie restrizioni all’azione. Gli esiti di tale riforma non sono stati ancora discussi con completezza. La tesi si struttura attorno a tre domande di ricerca: (i) quali soluzioni la nuova direttiva offre per superare i limiti precedenti nell’applicazione ai sistemi di IA; (ii) se tali soluzioni siano efficaci rispetto agli obiettivi della riforma e i fini dell’armonizzazione e di equa ripartizione dei rischi; (iii) quali approcci de iure condendo possano risolvere problemi ermeneutici ricorrenti. L’esegesi della nuova direttiva è condotta sui canoni metodologici propri del diritto privato italiano, con particolare ricorso all’interpretazione sistematica e teleologica, integrati da un’analisi storica e comparativa, in particolare su Italia, Francia e Germania. In tale esercizio, senza proporsi come commentario esaustivo del dibattito dottrinale e giurisprudenziale pregresso, si dà conto di snodi fondamentali nell’interpretazione del regime di responsabilità precedente. Al fine di valutare l’applicazione delle norme così riformate, la tesi mutua il criterio di valutazione dell’efficacia dalle valutazioni legislative della Commissione europea come coerenza tra obiettivi enunciati e risultati reali, e ricorre ad argomenti di analisi economica del diritto e dalla letteratura della scienza dei dati e informatica. Nel seguire le domande di ricerca e l’impianto metodologico adottato, la tesi ribadisce la natura di responsabilità oggettiva della responsabilità da prodotto, precisa i criteri di imputazione (difetto, danno e nesso causale tra difetto e danno) e delimita ruolo e limiti delle presunzioni. Soprattutto, la tesi giunge a due conclusioni principali. Valuta, anzitutto, l’efficacia della Direttiva (UE) 2024/2853 rispetto ai sistemi di IA come solo parziale: i correttivi introdotti migliorano definizioni e accesso alla prova, ma permangono criticità nell’individuazione del convenuto e nel governo delle presunzioni. In secondo luogo, propone una circostanza delle prestazioni che ancora il giudizio di difettosità alla diade funzione/ambito di applicazione e consente di inferire il difetto quando il sistema presenti, nel contesto pertinente, tassi di errore o incidenti stabilmente superiori ai comparabili o all’operatore umano sulla base di dati verificabili.
L'EFFICACIA DELLA DIRETTIVA (UE) 2028/2853 NELL'APPLICAZIONE AI PRODOTTI DI INTELLIGENZA ARTIFICIALE
De Gregorio, Fabrizio
2026
Abstract
The thesis aims to assess whether Directive (EU) 2024/2853 on liability for defective products is suitable to address potential harm connected to artificial intelligence (AI) systems. The research is situated within an operational analysis of product liability and offers one of the first systematic and comprehensive readings of the new Directive as applied to AI systems. The increasing deployment of AI systems - often software-based, opaque, interactive, and embedded within complex value chains - has made the recalibration of the liability regime both urgent and widely advocated in legal scholarship. While the Directive remains structurally aligned with Directive 85/374/EEC - preserving a defect-based strict liability regime and pursuing the same objectives of harmonisation and fair risk allocation - it introduces targeted adjustments for the digital age, intended to overcome outdated definitions, excessive evidentiary burdens, and procedural constraints. The outcomes of this reform have not yet been comprehensively examined. The thesis is structured around three research questions: (i) which solutions the new Directive provides to overcome the previous limitations in its application to AI systems; (ii) whether such solutions effectively achieve the reform’s objectives of harmonisation and fair risk distribution; and (iii) which de iure condendo approaches may resolve recurring interpretative problems. The exegesis of the Directive is conducted according to the methodological canons of Italian private law, with particular emphasis on systematic and teleological interpretation, complemented by historical and comparative analysis, especially with reference to Italy, France, and Germany. While not intended as an exhaustive commentary on prior doctrinal and case-law debates, the thesis revisits key interpretative turning points of the former liability regime. In evaluating the effectiveness of the reform, the thesis adopts the European Commission’s legislative assessment criterion of effectiveness—understood as coherence between stated objectives and actual outcomes—and incorporates insights from law and economics as well as from data science and computer science literature. Following the articulated research questions and methodological framework, the thesis reaffirms the strict liability nature of product liability, clarifies the attribution criteria (defect, damage, and causal link between defect and damage), and delineates the role and limits of evidentiary presumptions. The research reaches two principal conclusions. First, it assesses the effectiveness of Directive (EU) 2024/2853 with respect to AI systems as only partial: although the introduced adjustments improve definitional clarity and access to evidence, significant shortcomings remain, particularly concerning the identification of the proper defendant and the governance of presumptions. Second, the thesis proposes a contextualisation of performance expectations that anchors the assessment of defectiveness to the function/application dyad, allowing courts to infer defect where, within the relevant operational context, the system exhibits error or incident rates that are consistently higher than those of comparable systems or of a human operator, based on verifiable data.| File | Dimensione | Formato | |
|---|---|---|---|
|
tesiphd_completa_DeGregorio.pdf
accesso aperto
Licenza:
Creative Commons
Dimensione
1.76 MB
Formato
Adobe PDF
|
1.76 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/365267
URN:NBN:IT:UNICATT-365267