logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-01182023-154118


Tipo di tesi
Tesi di laurea magistrale
Autore
PARENTE, ROBERTO
URN
etd-01182023-154118
Titolo
Civil Liability Risks, advance technology, neural network modeling, appraisal, AI peril
Dipartimento
GIURISPRUDENZA
Corso di studi
DIRITTO DELL'INNOVAZIONE PER L'IMPRESA E LE ISTITUZIONI
Relatori
relatore Prof. Bertolini, Andrea
Parole chiave
  • advance technology
  • aI peril
  • appraisal
  • civil liability risks
  • neural network modeling
Data inizio appello
01/02/2023
Consultabilità
Non consultabile
Data di rilascio
01/02/2063
Riassunto
Regulating artificial intelligence systems requires, firstly, defining it and predicting its risks.

Since it has heterogeneous applications, its regulation needs to be technology specific and product oriented. Today, at his core, artificial intelligence systems identify correlations and makes predictions based on patterns.

The lack of foreseeability with regard to the potential behavior of artificial intelligence applications lead to difficulty in establishing a legal nexus of causation between the injured party and the wrongdoer, this in turn hampers the attribution of legal responsibility to a specific liable party.

Five categories of risk related to artificial intelligence technologies are foreseeable for users and service providers: reputational, legal, compliance/regulatory, strategic/financial, and operational.

Class-action suits and other litigation are almost certain to arise in the coming years as AI/ML adoption increases. At the outset new risks mitigation approaches are to develop, namely, governance, assurance, and risk modeling practices.

Much of the potential concern about AI/ML applications stems from classic predictive inference models that are optimised to make predictions primarily or solely on correlations in the datasets, which the models then employ in making predictions. As a result they are not concerned with causation as they are trained to find correlation.

In this paper it is argued that Civil Liability AI-based risks can be classified of known unknown nature as most of artificial intelligence entities are currently narrowly created to replace a human performing a specific narrow task or a predetermined range of tasks. Therefore, we have far greater information about the statistical likelihood of artificial intelligence causing harm than we do about widespread risks. This suggests that the artificial intelligence risk is indeed manageable.

Predicting civil liability risks related to artificial intelligence applications can be a complex task, as it often involves trying to quantify the potential financial impact of potential legal claims, assessing a range of factors including the potential harm caused by the artificial intelligence system, the likelihood of that harm occurring, and the extent to which the harm was foreseeable.

New techniques and methodologies are investigated to make the risk modeling more efficient and accurate than traditional heuristics of today actuarial practices.

One theory for assessing, predicting civil liability risks triggered by artificial intelligence applications might be to use parametric approaches rely on a collection of techniques for the automatic discovery of meaningful features representations that are optimal for a particular task within datasets through the use of neural networks designed in a hierarchal manner (Deep Learning).

In the civil liability risks modeling problem a modern approach might be to fit a neural network composed of a hierarchy of feature layers to the civil liability dataset. The shallow layers of the network learn to detect simple features of the input data, such as artificial intelligence application or user occupancy, while the deeper layers of the network learn more complicated combinations of these features, such as a Court’s Judges or existing jurisprudence.

Several specialised layers (convolutional, recurrent and embedding layers) have been introduced in literature and adopted in practice allowing for representation learning tailored specifically to unstructured data.

In the liability risks regression problem the feature vectors might comprise textual data, for example description of risks by lawyers, artificial intelligence engineers, and data architects, or jurisprudence from Lower Courts, Courts of Appeal, Supreme Courts, and Claim handlers’ notes, could be incorporated into the pricing and reserving process using recurrent neural and word embeddings networks.
File