The Governance of Big Data and Artificial Intelligence in Network Industries

The Complementary Roles of the EU Data Regulatory Framework and AI for EU Regulatory Framework

Different Dimensions of AI Regulations

EU initiatives on ethical and security frameworks concerning AI are not internationally isolated but are to be considered against the background of OECD AI standardization initiatives (OECD, 2019a, p. 3; ITF, 2021b, pp. 13f.). Five specific AI characteristics are considered particularly relevant: complexity, transparency/opacity, continuous adaptation, autonomous behavior, and big data taking into account the interaction between AI regulations and existing sectoral product safety legislation (European Commission, 2020c; 2021c, pp. 33–37). In addition, the General Data Protection Regulation (GDPR), ⁴ as it applies to personal data, plays an important role to guarantee AI security from the perspective of damages caused by data processing that infringes the GDPR (European Union Agency for Cybersecurity ENISA, 2020, pp. 8–10). Security is a principle which must be guaranteed when processing personal data and is thereby considered as a data protection by design instrument (recital 83, Article 5, Article 32 GDPR).In contrast, the goal of the proposed Artificial Intelligence Liability Directive (European Commission, 2022d) is to improve the functioning of the internal market focusing on aspects of compensation claims based on non-contractual civil liability for damages caused by the implementation of AI systems (European Parliament, 2023).

With regard to transparency/opacity requirements in the context of automated decision making, the GDPR is of relevance. Of particular relevance is Art. 22 (1): “The data subject shall have the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her.” Exemptions are possible under very specific conditions laid down in Art. 22 (2), if the data subject’s rights and freedoms and legitimate interests are guaranteed respectively the data subject`s provides explicit consent. According to Articles 13–15, data controllers must be able to explain how an AI system makes decisions that have a significant impact on individuals. The question arises as to what extent a full disclosure of the algorithm can be enforced (EPFL IRGC, 2018, p. 16; Chivot, Castro, 2019, p. 7).

The proposed EU regulation safeguarding fundamental rights, safety and liability differentiates between prohibited AI practices, high-risk AI systems, and non-high-risk AI systems (European Commission, 2021d, Articles 5, 6, 7, and 2021c, pp. 47-49), particularly with regard to their impact on the fundamental rights to human dignity, privacy, and data protection equality and non-discrimination. According to TITLE II, Prohibited Artificial Intelligence Practices, Article 5 (European Commission, 2021d, pp. 12f., 43f.), all those AI systems are prohibited which contravene EU values, particularly those that violate fundamental rights. Such rights include the prohibition of the manipulation of persons through subliminal techniques beyond a person’s consciousness, rules against the exploitation of specific vulnerable groups such as children or persons with disabilities, and the prohibition of AI-based social scoring for general purposes and carried out by public authorities. The use of “real time” remote biometric identification systems in publicly accessible spaces for the purpose of law enforcement is also prohibited, with exemptions only in very specific situations of security violations.

According to TITLE III, Classification of AI Systems as High-Risk, Article 6 (European Commission, 2021d, pp. 13f., p. 45) AI systems likely to pose a high risk to fundamental rights and safety are differentiated from non-high-risk AI systems. Classification rules for high-risk AI systems refer to AI systems intended to be used as a safety component or which are themselves products that pose a risk of harm to health and safety or risk an adverse impact on fundamental rights. ⁵ High-risk AI systems are classified in different areas of the society such as biometrics, education and vocational training, employment and workers management, law enforcement or access to essential private and public services. Of particular relevance from the perspective of network industries is the classification of AI systems to be used as safety components in the management and operation of critical infrastructures such as, road traffic and the supply of water, gas, heating and electricity as high-risk.

High-risk AI systems fall under the regulatory requirements of high-quality data documentation and traceability, transparency, human oversight, accuracy, and robustness to mitigate the risk to fundamental rights and safety posed by AI that is not covered by other existing legal frameworks such as the GDPR or the (proposed) AI Liability Directive (European Commission, 2022d) taking the final responsibility to the person who designed and implemented the AI system. The EU has established a system for registering stand-alone, high-risk AI applications in a public, EU-wide database. Key participants across the AI value chain include providers and users of AI systems (European Commission, 2021d, p. 12). According to Recital (53) “It is appropriate that a specific natural or legal person, defined as the provider, takes the responsibility for the placing on the market or putting into service of a high-risk AI systems, regardless of whether that natural or legal person is the person who designed or developed the system” (European Commission, 2021d, p. 31). ⁶

According to Recital (48) “High-risk AI systems should be designed and developed in such a way that natural persons can oversee their functioning. For this purpose, appropriate human oversight measures should be identified by the provider of the system before its placing on the market or putting into service. In particular, where appropriate, such measures should guarantee that the system is subject to in-built operational constraints that cannot be overridden by the system itself and is responsive to the human operator, and that the natural persons to whom human oversight has been assigned have the necessary competence, training and authority to carry out that role” (European Commission, 2021d, p. 30).

The proposed new EU AI regulatory framework pursues a risk-based approach to ensure that the regulatory interventions are proportionate avoiding oversized regulatory basis, inefficient duplication of regulatory interventions as well as inadequate regulatory instruments. Regarding proper regulatory basis the goal is to be effective in drawing the borderlines between prohibited AI practices, high-risk AI systems and non-high-risk AI systems. Due to the dynamic evolution of AI systems a particular challenge is to draw the borderline between regulated high-risk and unregulated non-high-risk AI systems within a forward looking regulatory decision-making process. Since AI algorithms are of important relevance within high-risk AI systems the goals of regulation granting fundamental rights, safety and liability should be pursued avoiding unnecessary or inadequate interventions (European Commission, 2020d, pp. 16–18).

AI, Big Data, and the Search for Innovative Algorithms

AI-Powered Big Data Virtual Networks

The IoT is driven by the complementary roles of physical networks and virtual networks. Virtual network providers develop big data value chains according to the necessities of the IoT applications, combining QoS-differentiated bandwidth capacities with the different complementary components such as sensor networks and geo-positioning services, and different forms of data processing activities (Knieps, 2017).

The transition from 4G to 5G represents a disruptive innovation based on fixed-mobile convergence, ultra-high bandwidth capacities, a large variety of heterogeneous QoS of bandwidth capacities, and the possibility of ultra-low latency guarantees of data packet transmission. The concomitant characteristics of 5G as an application-agnostic new General Purpose Technology (GPT) rest on the potential for pervasive use within a large and open set of downstream application sectors driven by complementary innovative activities widely dispersed throughout the economy (Knieps, Bauer, 2022; Parcu et al., 2022). The focus is on the requirements of 5G and concomitant 5G-based big data virtual networks, which also satisfy the strong QoS requirements of bandwidth capacities of the tactile Internet.

The heterogeneity of AI algorithms and machine learning depends on the big data value chains of big data virtual networks. The focus is on the heterogeneity of data value chains and concomitant heterogeneity of AI algorithms. Data are not homogeneous, such that depending on the underlying IoT applications (use cases), the degree of aggregation of data and time-sensitivity of data varies such that time-insensitive data, real-time streaming process of real-time data, as well as interactive processing of massive scaled data, may become relevant. This impacts the whole data value chain as well as the AI design necessities required for a particular IoT application. The complementary requirements for AI differ depending on the goal, which depends on the use case under consideration.

The important role of AI systems based on big data magnifies the importance of data sharing and open data for AI applications in many areas of the app economy such as transport, agriculture, financial services, marketing and advertising, science, health, criminal justice, security in the public sector, applications using augmented and virtual reality, etc. (OECD, 2019b, pp. 47–80). ⁷ The Joint Technical Committee of the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC) (Subcommittee SC 42, Artificial intelligence) listed 24 application domains indicating a large variety of fields such as home/service robotics, media and entertainment, the public sector, education, health care, energy, manufacturing (including logistics), mobility, and transportation (ISO/IEC, 2021, pp. 7–9).

Examples for heterogeneous AI-powered big data virtual networks are based on heterogeneous data collection, processing and actuator requirements driven by heterogeneous AI algorithms characteristics. An important class of examples are driven by deep learning algorithms illustrating the insights into various big data analytics in the context of smart grids (Syed et al., 2020).

Big Data and AI-Driven Use Cases in Transportation Industries

In recent years, big data in transport has gained particular relevance (OECD/ITF, 2015; OECD/ITF, 2016), with a large potential for reporting mobility data (ITF, 2022). In the following, the focus is on AI and big data in heterogeneous application areas of the transport sector. Transportation industries are an important driver of the IoT. Big data is becoming relevant in all areas of transportations. Big data-driven, AI-powered intramodal and intermodal transportation value chains are enabling an open set of innovative use cases in all areas of transportation industries, supply chains, and logistics. Best use cases of AI—machine learning in the transportation, shipping, and logistics industries—are manifold, resulting in decreasing transportation costs and increasing transportation performance characteristics (punctuality, reliability, safety). ⁸ AI and big data algorithms transform the real-time vehicle tracking data into real-time analysis and provide real-time information to drivers or self-driving autonomous vehicles, generating the best possible routes by adaptive simulations (Gesing et al., 2018).

AI-powered big data value chains act as enablers of vehicle platooning on public roads, self-driving aircraft, and autonomous trains (ISO/IEC, 2021, pp. 87–91). The role of AI in railway transport—such as intelligent train automation and operational intelligence for rail operators and infrastructure managers—enables the forecasting of infrastructure or rolling stock conditions, faster and less resource-intense repairs, reduction of maintenance costs, and the optimization of fleet reserves in case of train breakdowns. AI in maritime shipping and inland navigation, as well as smart ports and logistics, improve the allocation of relevant resources in a manner similar to that of just-in-time operations and offer transparency to optimize the real-time allocation of available dock spaces (European Parliament, 2019).

The transition towards fully automated driving may be considered as an evolutionary process. “It may be decades before the majority of cars on the road are fully autonomous. During this time, the human is likely to remain the critical decision maker either as the driver or as the supervisor of the AI system doing the driving” (Fridman, Brown, Glazer et al., 2017, p. 3). To the extent that data supporting enforcement actions relate directly to unique vehicles, individuals, and their specific behaviors, the most privacy-sensitive data reporting occurs in this context (ITF, 2022, p. 7).

Heterogeneity of AI-Powered Big Data Virtual Networks in Transportation Industries

The analysis of the heterogeneity of AI-powered big data virtual networks in transport leads immediately to the question regarding the heterogeneity of big data value chains and AI algorithms. Heterogeneous AI-powered big data virtual networks for physical applications are based on heterogeneous big datasets and heterogeneous algorithms. A fundamental shift in the transportation sector (physical side) towards smart network industries is based on big data-driven innovations due to the combination of enhanced sensor technologies, the drop in data storage costs, and the availability of new data processing algorithms and innovations in communication technologies (OECD/ITF, 2015, p. 9). Data aggregation is not application-agnostic and depends on the particular use case under consideration. There are multiple ways to aggregate vehicle data (ITF, 2021b, p. 23). The degree of aggregation and the heterogeneity of requirements for the geographical footprint (local, regional, global), as well as the heterogeneous requirements for the time-sensitivity of datasets (static, dynamic real time), are dependent on the different use cases under consideration. Data transmission is time-sensitive according to different requirements (latency sensitivity varying between ultra-low latency of milliseconds, seconds, minutes, hours, etc.).

Appropriate aggregation of data depends on the underlying data value chain problems to be solved: e.g., planning data, operational data, and data supporting enforcement actions and vary in terms of time sensitivity. Planning data may be aggregated according to the requirements of the relevant time horizon and are non-time-sensitive, operational data respond to real-time events and may be aggregated according to the logistic transport requirements, congestion management may be time-sensitive depending on the congestion levels and aggregated on a given infrastructure corridor or slot, autonomous vehicle data require ultra-low latency-dependent traffic management within a localized mobility platform. Heterogeneous footprints can be a local, regional, national, or cross-border focus of big data virtual networks, depending on the underlying space dimension of the big data virtual network. Heterogeneous data protection and safety requirements differentiate between data-related risks, privacy, trade secrets, and commercially sensitive data with concomitant impact on incentives for the collection and sharing of data (OECD/ITF, 2015).

AI in proactive road infrastructure safety management requires the sensing and sharing of safety-relevant data along the entire road network (ITF, 2021b). The focus is on accurate risk prediction and guidance for proactive road network safety management. Data often remain in separate entities instead of being shared; a barrier for AI implementation is the fear of litigation for disclosure of identifiable personal information, which highlights the necessity of decision-relevant aggregation of data. The automotive industry produces high volumes of moving vehicle data including indicators of traffic volume, speed, and information from active safety systems. In this context, it would be valuable to enable market platforms to utilize heterogeneous data from vehicles and smartphone apps. Risk prediction models based on data aggregated over different time periods instead of the real-time date, are not critical for private data protection, resulting in less complex AI algorithms compared to the big data processing of real-time datasets.

Shipping volume prediction turns historical data from a transportation management system into a data-driven tool for forecasting sales numbers and shipping volumes. Forecasts consist of historical data and also external factors, like seasonality or weather. Another AI-powered big data application (use case) is real-time vehicle tracking. Based on IoT sensors, this process involves real-time monitoring of the speed, location, and direction of a vehicle: A local gateway delivers the data to a network server while the truck is still on the road and sends information to a cloud-based application server.

AI-powered big data virtual networks for autonomous driving systems are to be considered as an evolutionary process: Many aspects of driver assistance and vehicle performance are increasingly being automated with data-driven, learning-based approaches. The beneficial role of data sharing can be seen in situations whereby competitors win by collaborating, sharing high-level insights and large-scale, real-world data. Parties involved are car companies, automotive parts suppliers, insurance companies, technology companies, government regulatory agencies, and researchers in applied AI and vehicle safety (OECD, 2019b, p. 25, pp. 48–53; Fridman, Brown, Glazer et al., 2017).

AI systems in driverless cars must be trained to deal with critical situations, which raises the problem of training in computer-simulated environments. Autonomous driving systems and their adaption to ride-sharing services entail potential liability problems in the case of accidents. The importance of sharing real-time data among different development firms in the course of their trials to train AI algorithms has been pointed out (OECD, 2019b, pp. 48–52). The role of large-scale deep learning is based on a massive-scale dataset collected from the instrumental vehicle fleet (Fridman, Brown, Glazer et al., 2017, pp. 6f.).

One aim of this paper is to make more explicit the neglected role of broadband-driven networked AI: e.g., not only a particular model of a car is required for autonomous driving systems, but a more general model of networked autonomous vehicles focusing on interoperability based on a car-to-cloud data standard such that data from different manufacturers can be combined (OECD, 2019b, p. 25; ITF, 2021a).

Building accessible datasets in the context of automated driving for understand driving behavior raises the issue of big data access, either with proprietary data, shared data across firms, or the role of government funding for open data collection. “AI is notoriously dependent on provision of large amounts of quality data. Will drivers, carmakers, commercial transport operators and telematics companies be willing to share the data they produce?” (ITF, 2021b, p. 13). AI creates the necessity for big data sharing. This also creates incentives for competitors to collaborate in an environment where competing firms win by sharing large-scale, real-world data in the search for a better understanding of the complexity of networked autonomous driver behavior (Fridman et al., 2017, p. 9). ⁹