ON THE USE OF SMART SYSTEMS: ETHICAL IMPACTS AND LIABILITY OF ARTIFICIAL INTELLIGENCE
By Giorgio Nocera (1), Giulio Francesco Aiello (2); Fabrizio De Vita (3)

(1) Department: Development and Engineering
Company: nabla. Email: gnocera@nabla.studio Orcid: https://orcid.org/0009-0007-9871-0316

(2) Department: Operations Società: nabla Email: giulio@nabla.studio

(3) Department of Engineering. University of Messina. Email: fdevita@unime.it
Orcid: https://orcid.org/0000-0002-6709-8001

Abstract:Artificial Intelligence (AI) represents a technological resource that is revolutionizing the way we interact with the surrounding world. Nowadays, there are numerous services that make use of machine learning algorithms to simplify some of the operations we perform daily. In these systems, data plays a crucial role, especially in the training process of any AI model, to the point that it can influence its functioning. By virtue of this, it is essential to pay particular attention to the collection phase of data, in order to build datasets that are free from biases and that ensure a fair and correct functioning of a system. In fact, since such tools can have consequences – even in the long term – on people’s lives, the use of distorted data could raise important ethical issues. In such a context, the construction and use of regulations and norms that increase the awareness of the end user are fundamental, ensuring the safety and effectiveness of AI without acting as an obstacle to technological progress. This article outlines the importance of analyzing ethical aspects in the design and development phases of AI-powered systems, with particular attention to the social implications that this technology can have.

Keywords: Artificial Intelligence, Ethics, Responsibility

1. Introduction

Artificial Intelligence (AI) is becoming an increasingly important technological resource for humanity. Just in the last decade, thanks to its continuous integration with existing systems, it has revolutionized the world we live in, with the intent to improve people’s quality of life. In fact, despite not always being obvious, many of the applications and services used daily by millions of users rely on AI algorithms.

Search engines, such as Google, use this technology to provide users with a better search experience by personalizing results in real-time; virtual assistants like Siri, Alexa, and Google Assistant use machine learning models to understand human language and make several daily activities more natural, such as finding directions or interacting with home appliances; in the entertainment industry, companies like Netflix and Spotify use AI to identify and suggest suitable content for each of their users. Some car manufacturers, such as Tesla and Volvo, use machine learning techniques to improve the safety of transportation and develop self-driving cars. Moreover, AI has proven to be a valuable ally in the medical field, where machine learning algorithms can speed up the diagnosis of some diseases and provide significant support in the search for new treatments. These examples of application demonstrate how AI permeates our lives, often in imperceptible ways, thus improving the quality of services and technologies.

However, the increasing use of AI in everyday services raises important ethical issues since it has a direct impact on people’s lives. For example, AI-enabled services could be used to create content that should be the result of human labor, such as research activities that lead to the writing of scientific articles or the resolution of school and university exam exercises.

Furthermore, AI could also be used to make decisions regarding personnel hiring, candidate selection for loan programs, or the detention of individuals suspected of crimes, i.e., in decision-making processes that could have long-term effects on human life. In these latter cases, if the programming rules of AI do not take into account diversity, justice, and non-discrimination, ethical problems may arise. If the model is influenced by prejudices and stereotypes that derive from the dataset used during the learning or training phase, this could lead to unfair and discriminatory implementation. Therefore, researchers and professionals in the field are working to ensure that AI is developed and used ethically and responsibly.

The goal of this work is to deepen the relationship between ethics and AI. In particular, Chapter 2 provides a brief overview of some of the most significant events that have marked the history of AI, from its inception to the present day. Chapter 3 outlines the importance of the ethics of systems that make use of AI, highlighting some cases where the use of such intelligent systems has caused unethical behavior. Subsequently,

in chapter 4, the importance of data in all AI-equipped systems is analyzed, highlighting the problems that a poorly representative dataset can cause if used during training phases. Chapter 5 focuses on how the responsibilities of actions involving an AI system are managed, considering the state of the art and what the current and future needs are. Finally, in chapter 6, a summary of the topics covered is presented, and a possible path to pursue for improving the increasingly present coexistence relationship between humans and so-called intelligent machines is highlighted.

2. The Artificial Intelligence

In the 1940s, mathematicians and scientists such as Shannon, Wiener, and McCulloch, delving into the study of information theory and cybernetics, marked the beginning of a period of great scientific ferment that prepared the community for the emergence of AI as a scientific discipline in the Fifties. Since 1950, a series of events related to the development of intelligent machines took place, the main ones of which are given below.

Just in 1950, Claude Shannon described the first computer program capable of playing chess (Shannon 1988) and presented Theseus^[1], a mechanical mouse capable of exploring a maze, finding the way out, and using the acquired experience to solve cases in which the maze had been modified. In the same year, Isaac Asimov proposed his “I, Robot” (Asimov 2013), in which he discussed the possibility of peaceful coexistence between human beings and autonomous robots through the help of his three laws of robotics, providing the basis for future discussions on the ethics of robots and AI. Alan Turing – considered one of the fathers of modern computing – presented, in his famous article “Computing machinery and intelligence” (Turing 1950), a test according to which a machine could be considered intelligent if its behavior was indistinguishable from that of a person when observed by a human being.

In the years following Turing’s test, the first fields of research related to AI and thinking machines began to take shape, and in 1956 McCarthy introduced the term “artificial intelligence” during the Dartmouth Conference (McCarthy et al. 1955). Some of the most eminent scientists, engineers, mathematicians, and psychologists of the time participated in this conference. Among these, Allen Newell, Cliff Shaw, and Herbert Simon presented the “Logic Theorist” (Newell et al. 1956), the first AI program designed to imitate the problem-solving abilities of a human being. Unlike modern approaches to AI, which are based on the so-called learning from experience, the Logic Theorist was based on “symbolic AI,” the dominant branch of AI between the 1950s and 1970s. This approach, based on the manipulation of symbols and rules where knowledge is explicitly encoded, has limitations related to the management of uncertain or imprecise situations and in automatic learning.

Due to the importance of the topics discussed, the Dartmouth seminar is considered the founding event in the history of AI (Crevier 1993). At this conference, in fact, many concepts that are still the subject of research today, such as neural networks and natural language processing, were discussed. In the same period, Arthur Samuel presented the first software capable of playing checkers (Samuel 2000), learning, and adapting its own strategy, thus introducing the concept of reinforcement learning (Sutton 1997). In the early 1960s, the first artificial neural networks were also developed, including Rosenblatt and Pitt’s Perceptron algorithm (McCulloch et al. 1943), capable of recognizing images and learning the difference between geometric shapes.

The exciting successes achieved in the 1950s and 1960s gathered huge investments but, at the same time, caused a wave of excessive confidence and unrealistic expectations. And precisely these exaggerated predictions and the gap with the disappointing results obtained in the 1970s and 1980s led to what was subsequently called the “AI Winter” (Crevier 1993). The AI Winter was a period of disappointment and lack of interest that caused a sharp decline in research activities until the 1990s. During these twenty years, there were alternating periods of excitement and disappointment that led researchers to a paradigm shift that occurred towards the end of the 1980s. In particular, the focus was on the fact that humans also learn through trial and error, i.e., through experience, and based on these considerations, interest in artificial neural networks was rekindled. Significant contributions were made by the development of the backpropagation algorithm by Rumelhart, Hinton, and Williams (Rumelhart et al. 1986), which allowed the optimization of the training of such networks; the presentation of convolutional neural networks by LeCun (LeCun et al. 1989), among whose first implementations it is possible to identify the recognition of characters and numbers^[2]; and the studies on Deep Learning, which began in 2010 and thanks to Bengio, Hinton, and Le Cun (LeCun et al. 2015), led to algorithms capable of solving increasingly complex problems. At the same time, the advance of miniaturization, the development of more energy-efficient electronic devices, and the expansion of the Internet network, ensured the evolution of what is now called the Internet of Things (IoT), configured as a large distribution of devices capable of collecting and processing data.
Starting from the 2000s, AI has experienced exponential growth thanks to advancements in computing power, represented in part by Cloud technology and the availability of large amounts of data in what we call the era of big data. This context has allowed for the development of applications such as virtual assistants, autonomous driving cars, and predictive analysis, enabling AI to be successfully applied in many areas, including medicine, agriculture, finance, and cybersecurity.

For example, in 2011, Google launched a project called Google Brain^[3], which uses a multi-layer artificial neural network to improve the capabilities of the eponymous search engine. That same year, IBM’s AI system Watson participated in the TV quiz show Jeopardy! (Best 2013), competing against some of the best players of the program and winning, demonstrating that AI could be used not only for research but also to effectively answer complex questions. In 2014, Facebook launched an AI research project called DeepFace, producing a facial recognition algorithm that could identify individuals with 97.35%^[4] accuracy. However, current AI is still mainly defined as weak, meaning it is limited to performing specific and well-defined tasks without showing any form of autonomy or generalized intelligence.

At the end of 2015, a group of scientists and entrepreneurs including Elon Musk and Sam Altman founded OpenAI^[5], a non-profit company based in San Francisco. Its goal is to collaborate with various organizations and companies, such as Microsoft and Amazon, to promote AI research and development, paying particular attention to the existential risks arising from so-called general artificial intelligence – also known as strong AI, which should be able to emulate human reasoning to solve general problems – making its patents and research open to the public.

Following OpenAI’s transformation from a non-profit company to a capped-profit company to attract new capital and collaborations, between 2021 and 2022, applications capable of generating images and entertaining conversations were presented to the public. Specifically, research on conversational technologies led OpenAI to create ChatGPT^[6], a tool that uses deep learning to generate responses similarly to how a human would. This model was trained on huge amounts of text data, becoming one of the most advanced conversation-based AI models.

Despite the computing power achieved, many experts believe that strong AI, that is a system capable of reaching and surpassing human intelligence, is still a distant goal. However, as demonstrated by the latest versions of applications like ChatGPT, research and development of new automatic learning techniques are pushing AI closer and closer to this goal.

3. The AI Ethics

The Treccani Encyclopedia defines ethics as the branch of philosophy that studies “the will and actions of man as a free and rational being, capable of giving himself a set of values and norms to respect […] and does not only concern the subjective sphere of personal choices but also affects collective life”^[7]. In other words, ethics evaluates human behavior and decisions based on fundamental values such as justice, truth, freedom, respect for human dignity, the well-being of individuals and communities, distinguishing them as good, just, and lawful, in contrast to behaviors considered unjust, unlawful, inappropriate, or bad. In general, ethics seeks to answer questions such as “What is right or wrong?” “What should I do?” “How should I behave?” providing a framework of reference to help people make more conscious and responsible decisions.

For this reason, and because AI is a technology that can have significant and lasting impacts on society and human rights, it is important to consider the ethical aspects that concern its relationship with humans. In this context, the so-called “AI ethics” is defined, which focuses on how to design, develop, and use AI in a responsible, fair, and transparent manner, also considering issues related to responsibility and management of AI itself, including:

• fairness and justice, as AI could, in some cases and through some decisions, perpetuate and amplify social and economic inequalities;
• transparency of the choices made by machines, as these decisions are often the result of the execution of complex algorithms;
• responsibility for the actions of machines equipped with intelligent algorithms and the definition of liability in the event of damage or errors;
• impact on jobs, as on the one hand, mechanical activities are increasingly automated, on the other hand, new sectors requiring advanced technical and management skills are emerging.

Furthermore, since this technology is becoming more widespread in many sectors, from medicine to finance, to personnel selection activities, it is important to consider the specific circumstance in which AI is used and therefore, the different perspectives and interests at stake. For example, the use of intelligent algorithms in the medical field can raise different ethical issues than those that emerge in finance or recruitment. In fact, it is precisely from the use of AI in everyday decision-making processes that over time, unethical behaviors and, in some cases, ethically incorrect behaviors resulting from intelligent algorithm operations have been identified.

One of the first known cases in which AI showed racial discrimination dates back to 2015 when Google faced criticism for its intelligent image recognition algorithm that “erroneously tagged a black couple as gorillas”^[8]. This problem raised numerous concerns about racism in AI and led Google – the development company that created and provided the service – to publicly apologize as the responsible party, as it was the company that had created and published the software. Specifically, the cause of the problem was related to the model’s training operations, i.e., the lack of diversity in the data set used to train the machine learning algorithm because the dataset used contained a higher quantity of images of light-skinned people compared to those of people of color.

Another news story of discrimination caused by AI algorithms was reported by Reuters, which stated that Amazon decided to stop a pilot project on a pre-selection system for personnel that made use of AI algorithms due to problems related to discrimination against women and minorities (Vincent 2018). The algorithm, trained on historical data of employees – which were mainly composed of men – tended to systematically downgrade the CVs of female candidates and those with typical ethnic minority names, thus showing a bias in the selection process. However, cases of gender or race discrimination due to AI are not limited to the field of personnel selection and recruitment or image classification. In the financial sector, for example, the launch of the Apple Card – a virtual credit card, integrated with the iPhone’s Wallet app, usable for purchases worldwide – using an AI-equipped credit scoring algorithm, resulted in some women receiving lower credit limits than men with the same level of income and credit.

In recent years, moreover, the use of AI algorithms in legal systems is becoming increasingly common. An example is its use in decision-making processes concerning pre-trial detention, which raises numerous concerns about the impartiality of algorithms and the possibility that they may discriminate against certain categories of people. For example, a 2016 study (Mattu et al. 2016) demonstrated how the COMPAS algorithm, adopted to predict the risk of recidivism among detainees, although not using race per se, used many race-related factors and returned higher risk scores for black defendants (Vyas et al. 2020). Furthermore, it is important to note that some AI algorithms work in “black-box” mode, i.e., they produce results without the logic and decision-making process behind the choices being completely clear and transparent. Obviously, this characteristic becomes particularly relevant in the field of justice (Bathaee 2018), where the use of these algorithms has raised concerns about transparency, impartiality, and human rights violations.

In 2016, Microsoft launched Tay, a chatbot – a program that uses AI to simulate a conversation with a user through a chat – on Twitter, whose goal was to learn from user conversations through a technique called continuous learning (Hancell 2023). However, Tay began sending offensive and inappropriate messages due to the influence of some malicious users. This led Microsoft to deactivate the chatbot within 24 hours (Perez 2016). This problem generated concerns about the ability of AI to learn and imitate human behavior, including offensive and discriminatory behavior, highlighting the need for greater regulation and ethical attention in the implementation of AI. In this case, in fact, the model had not been programmed to distinguish acceptable from unacceptable behavior.

Samantha Cole, on the other hand, published an article (Cole 2017) on Motherboard in 2017 highlighting some cases of AI usage for creating well-known fake videos called “Deepfakes”. In this AI technique, machine learning procedures based on neural networks are used to create manipulated videos or images that appear real. Specifically, the author explained how the technology was used by a Reddit user called “Deepfakes”, who selected the faces of celebrities like Gal Gadot, Maisie Williams, and Taylor Swift and inserted them into the bodies of various porn actresses, resulting in several obscene videos that appear to be performed by the celebrities. In a similar context, however, the fact that the user only used his computer and a machine learning algorithm, which anyone can download from the internet, made it even more dangerous (Cerdan et al. 2019). This ease of creation raised considerable concerns, as videos created using the “Deepfake” technique are not only linked to the porn industry, but over time, have also been used to create fake political speeches and spread fake news, which can have serious consequences for those involved.

In the medical field, the use of AI algorithms has raised ethical concerns as well. An example is represented by the research work carried out by Raj et al. (Raj et al. 2019), in which the objective was to create machine learning algorithms capable of predicting the mortality of patients in intensive care after traumatic brain injury. The system created was able to provide results with an accuracy of over 80% in predicting mortality at 30 days. In this case, although the algorithm was developed with the objective of improving patient care, this discovery raised concerns about the use of AI technologies for end-of-life decision-making, as such predictions could influence the choice of doctors and patients regarding the care to be provided.

In the social sphere, the spread of intelligent algorithms has elicited different reactions, ranging from the enthusiasm of those who see it as a potential tool to increase efficiency and innovation, to the concern of those who are worried about the possible impact on job loss. In particular, in 2023, the University of Queensland and KPMG Australia conducted a survey (Gillespie et al. 2023) interviewing over 17,000 people from 17 different countries. From here, although it was found that over two-thirds of those interviewed feel optimistic about the benefits that AI can bring to society, 42% (about two out of five people) are concerned that AI could replace jobs in their sector. Today, the number of platforms powered by AI algorithms available online is growing very quickly, and tools like ChatGPT are able to provide substantial support in numerous activities, such as data collection and analysis, source finding, and support in copywriting activities.

However, even though OpenAI CEO Sam Altman has state (Altman 2022) that relying on ChatGPT for important matters is not advisable, as it represents only an embryonic form of progress and there are still many aspects to be developed to ensure its solidity and truthfulness, several users have shown themselves to be extremely enthusiastic about the results already obtained. Some have described on various social networks the steps they have taken to develop smartphone applications without any experience in specific programming languages (Pigford 2023). At the same time, ResumeBuilder.com conducted a survey^[9] on a sample of 1,000 American companies, in which about half claimed to have replaced some of their employees with ChatGPT.

It was designed with the goal of conversing simply with all human beings, and in fact, one of its strengths lies in its ease of use and interaction, that is, through a chat where the user can ask the model to answer questions, write code or texts, or even solve debugging problems.

The Generative Pretrained Transformer 3 (GPT-3) – the AI model developed by OpenAI, as the basis of ChatGPT – effectiveness in producing results for most of these tasks allows us to understand the enthusiasm of the community. In addition, the ability to remember past conversations, and its way of correcting what is highlighted as an error, increases the user’s temptation to suppose that the machine understands what he or she is writing. However, as pointed out by Floridi – a full professor of philosophy and ethics of information at the University of Oxford – in an interview (Marchetti 2023), obviously the tool does not have real intelligence and above all is not capable of understanding but is “simply” able to produce sentences that are probable given the set of information provided during the training phase^[10].It is precisely its operation that is the reason why its outputs can sometimes be inaccurate, untruthful, and sometimes misleading^[11]. Furthermore, since its training phase ended at the end of 2021, it is not able to know everything that has happened since. In particular, the model, equipped with 175 billion parameters, was trained on large amounts of data from the Internet and written by human beings, and that is why the answers it provides can seem human.

4. The Importance of Data in AI

The training of complex AI models, which use deep learning algorithms – an AI technology based on “deep” neural networks, i.e., with multiple layers, and autonomous learning – takes place through the administration of large amounts of data and the exploitation of significant computing power. An example is GPT-3, which has been trained on over 570GB of data (Brown et al. 2020), after filtering a starting dataset containing over 45TB of data. This dataset is composed of data from various sources, such as books, newspapers, scientific articles, patents, online conversations, and so on.

The purpose of the training phase of an AI model is to use the collected data to structure a mathematical model so that it can predict the correct outputs for new inputs that the algorithm has never seen before. In the specific case of GPT-3, the data was used to proceed with supervised training – a type of machine learning approach in which an algorithm is trained on a set of input data for which the corresponding desired output is also provided – whose goal is to allow the model to predict the next word in a given sentence or to generate text autonomously. During training, the model explored the existing relationships between words within the input data, seeking to identify patterns and regularities in the text, which allowed it to learn a wide range of linguistic and cultural knowledge.

Some other examples are deep learning models based on convolutional neural networks and used for image recognition, such as “Amazon Rekognition,” developed by Amazon, or “Inception-v3,” developed by Google. These algorithms are capable of classifying objects present in photos, detecting human faces or analyzing facial expressions. In this case too, the models’ generalization skills – i.e., the ability to accurately recognize a wide range of objects – derive from the fact that they are trained on a large dataset of images, called ImageNet, which contains over 14 million labeled images in more than 20,000 categories, which subsequently allow the model to recognize patterns or elements in input images that the model had not yet seen.

In general, it is precisely the principle of operation of machine learning that requires deep learning models to use large datasets during the training phase. Whether they are used to recognize plant diseases through photos of leaves (De Vita et al. 2020), evaluate a user’s creditworthiness (Wang e Xiao 2022), or translate between multiple languages (Johnson et al. 2017), in all these models, data plays a crucial role, so much so that Google’s Director of Research, Peter Norvig, wrote in 2009 (Halevy et al. 2009) that “simple machine learning models trained on large amounts of data tend to have higher accuracy than more complex models trained on a smaller amount of data”.

However, the quantity of data alone is not enough to achieve good results. Another critical factor for machine learning models to be able to provide excellent results is the quality of the data. For this reason, for an AI model to have a good chance of learning the associations correctly and making precise and generalized predictions, it is essential that the data be of high quality, representative of the problem to be solved, and correctly labeled. In fact, when a model is trained with low-quality data, for example, containing errors, incomplete information, or not representative of the problem to be solved, it presents limitations in creating links between input and output variables, which can lead to inaccurate or even incorrect predictions.

For example, in cases where the data used to train models are collected in a non-uniform manner, contain factors that may be subject to biases, such as race, or cannot be considered a representative sample of society, there is a risk that the results provided by the model will be distorted. Over time, in fact, it has been shown (Grote et al. 2022) how some machine learning models used for medical diagnoses have lower performance when used to predict outcomes that affect women or ethnic minorities. In the research conducted by Obermeyer et al. (Obermeyer et al. 2019), the authors estimated that a calculation error in the system had reduced the number of African American patients who should have been able to participate in healthcare programs by more than half.

Therefore, in order to reduce – or even better avoid – the risk of bias in producing results or generating incorrect or discriminatory responses, the data used for training AI models must be accurate, complete, sufficient in number, and free of biases. This is precisely why a very important phase during the development of AI algorithms is represented by the data analysis, research, or construction phase. In fact, although the number of datasets available online ^[12] is continuously increasing, depending on the specific implementation to be realized, it may be necessary to build an ad hoc one that takes into account all the peculiarities of the system.

In these cases, an initial analysis of the problem is carried out and the definition of the inputs and outputs of the system is defined (for example, input images and an output number corresponding to a classification, or a group of input numbers and a numerical value output). Subsequently, data collection and selection are carried out that will be provided to the system and, in the event that supervised training is to be carried out, it is essential to “label” the selected data, i.e., associate with these, the corresponding output value that one would like to obtain from the model.

Considering the phases of creation and modeling of the dataset, one of the key aspects of the process is represented by the selection of data sources. In fact, it is appropriate to verify if it is possible to draw on public data already present or if it is necessary to collect new samples. In the latter case, it is necessary to be able to collect the data in reasonable times. In addition, depending on the type of data collected, it is fundamental to emphasize the importance of privacy regarding the information collected, as it is necessary to ensure that such data are used ethically and respecting the rights of users.

An example among all is represented by the healthcare sector. Here, the fact that health data is rich in sensitive information has highlighted, over time, the need to regulate its use to limit the potential risks related to possible “lack of privacy, confidentiality, and data protection for patients and citizens” (Directorate-General for Parliamentary Research Services (European Parliament) et al. 2022). In fact, although anonymization practices allow sensitive information to be removed from health data, compromises often have to be managed: if the data are anonymized to the point of not providing any useful information about patients, there is a risk of losing fundamental information for the model; conversely, when the usefulness of the data – and therefore the information contained within them – is high, the risk of re identification increases (Sepas et al. 2022).

The issue of privacy, however, does not only concern healthcare data. Another case that has raised significant controversy regarding the collection and management of data for AI algorithm training activities has to do with Bard, the AI chatbot created by Google. This model, like ChatGPT, is equipped with generative AI – that is, AI that uses machine learning models to create completely new results from a training dataset – and, just like its competitor developed by OpenAI, although it is able to replicate results in a “realistic” manner, it often makes mistakes in the content of the results provided. The controversy was sparked in 2023 by a tweet (Crawford 2023) from researcher Kate Crawford, who reported the model’s response to the question “Where do Bard’s datasets come from?” The software, in fact, indicated that some of the data sources included Google products, such as Gmail – one of the most popular email clients in the world. The response, which was not well-received by various Gmail users, was immediately debunked by the official Twitter profile of Google Workspace, which also emphasized that Bard is still an experiment and, as such, may make mistakes.

In this context, there are regulations available in various countries to guarantee users’ privacy. In Europe, for example, the General Data Protection Regulation (GDPR) is in force, which regulates the management of personal data of European Union citizens. In the specific case of Italy, there is the Privacy Guarantor, an administrative authority whose responsibilities include ensuring compliance with the regulations on the processing of personal data. This authority oversees platforms, services, and large companies that deal with large amounts of personal data and, if it detects unclear aspects, it can decide to decree the limitation of the processing of Italian users’ data towards one or more companies, as happened with ChatGPT on March 31, 2023^[13]. This decision led OpenAI to temporarily suspend the service for all users connecting from Italy, thus dividing the digital community into two groups: those who agree with the Guarantor’s decision, opposing an opaque and unregulated use of their personal data by OpenAI(Bernieri 2023), and those who are hostile to this solution, who would have preferred to continue using the service and see this block as “a block to Italy’s progress, which is excluded from using this technology” (Cremonesi 2023).

5. The EthicalResponsibility of AI

As previously discussed, AI can be used to improve people’s quality of life, protect the environment, and promote social justice. It can provide the necessary support for the development of new technologies that can be used in fields ranging from entertainment to the automotive industry, passing through applications in the healthcare field. However, a wrong implementation of AI could perpetuate inequalities, cause privacy violations, allow for the manipulation of opinions, cause damage, and in some cases even death.

Indeed, although a study (Luttrell et al. 2015) estimated that if 90% of cars in the United States were autonomous, 25,000 lives could be saved every year, there are several cases of fatal accidents involving such vehicles. Specifically, the authority that regulates transport in the United States, the National Highway Traffic Safety Administration (NHTSA), has published a report^[14] indicating that 392 accidents that occurred between July 1, 2021, and May 15, 2022, involved cars with advanced driver assistance technology, and among them, there were 6 deaths.

The first attested case of a similar incident occurred on March 18, 2018, in Arizona when a self-driving car hit and killed a woman who was crossing a four-lane road (Levin et al. 2018). In this specific circumstance, Rafaela Vasquez, the backup driver – the person in the driver’s seat with the ability to regain control of the vehicle in case of an emergency to ensure safety during the testing phase of a self-driving car – was charged with negligent homicide for her “negligence in the performance of her duties in intervening in a dangerous situation”^[15], as it was documented that the woman was watching the program “The Voice” in the minutes before the accident, according to a police report.

With the aim of standardizing autonomous car solutions and defining responsibilities in various scenarios, the Society of Automotive Engineers (SAE) – a standardization entity in the aerospace, automotive, and vehicular industries – has created a scale^[16] of six levels that specifies and defines the degrees of vehicle autonomy, where the last two correspond to completely autonomous systems, in which human presence is not necessary. In situations where machines can make autonomous choices, it may not be possible to find a human figure directly responsible for the accident. In these cases, if there are problems related to the physical realization of the car, it will be possible to contest the responsibility of the defects directly to the manufacturer. Similarly, responsibility for damage caused by programming errors that did not anticipate what should have been anticipated is attributable to the programmer (Cappellini 2019). But what happens in cases not directly related to product damage?

Indeed, since during its operation, an autonomous car may find itself in unexpected situations, it must be able to act by trying to minimize the damage. To do this, these systems must proceed with choices that are the result of evaluations and settings provided to the machine by designers.

With these issues in mind and to better explore such a “moral dilemma” in the field of autonomous driving, a group of researchers from the Massachusetts Institute of Technology (MIT), Harvard, and other universities published a paper entitled “The Moral Machine Experiment” in the journal Nature (Awad et al. 2018). In this work, Awad et al. described the result of a survey conducted on a platform through which more than 40 million responses were collected from millions of people living in 233 different countries and territories. This survey presented various scenarios in which users were forced to make a moral choice about whom to save in case of an accident involving different groups of people, such as between a group of pedestrians and a car driver, between a group of elderly people and a group of children, or between a group of obese people and a group of thin people, the authors found that people generally preferred to save as many lives as possible, in any context, even at the cost of sacrificing the driver. Furthermore, the study highlighted some cultural and individual differences in ethical preferences, such as age, gender, and education level.

It is, therefore, quite evident that the massive adoption of AI systems puts traditional concepts of responsibility to the test. Consider a scenario where a patient undergoes surgery and suffers harm due to a robot equipped with an AI algorithm working in black-box mode. Over time, various ideas have been proposed to address this issue. Among these is the granting of legal status to the device and reserving a heritage to enable it to compensate for the damages caused, or a partial legal recognition of the AI device so that it takes direct responsibility for harmful actions while continuing to consider programmers and users indirectly responsible based on specific damage.

Another option is to implement a mandatory insurance system similar to liability insurance (Colletti 2021).

In any case, these proposals share the fact that current models of responsibility management are not prepared for this type of interaction, since the norms that govern them were conceived at a time when injuries and errors were caused by humans or machines directly operated by humans. Since the integration of AI systems can foresee scenarios where injury occurs without any human intervention, such models must be adapted. In fact, just as the user is responsible for the improper use of an AI-enabled system, programmers and builders are responsible for problems related to design or implementation. More generally, the goal is to be able to distribute responsibilities correctly and fairly to the point of guaranteeing the safety and effectiveness of AI while not acting as a barrier to innovation and technological progress (Parikh et al. 2022).

In order to ensure ethical use of AI, the European Union has developed and published a series of guidelines in the form of an ethical code that includes recommendations for the development and implementation of AI-enabled systems. Subsequently, in 2021, it proposed a law called “AI Act,” within which it provides a definition of AI, and presents a risk-based approach that divides the uses of AI into four levels^[17]:

1. Minimal risk or no risk: allowed use without restrictions, such as spam filters or AI-enabled video games;
2. Limited risk: allowed use subject to information/transparency obligations, such as chatbots;
3. High risk: allowed use subject to certain requirements being met and obtaining ex-ante conformity assessment, such as autonomous driving vehicles or medical devices;
4. Unacceptable risk: prohibited, such as social scoring systems.

This regulation also contains some rules regarding strong AI, also known as Artificial General Intelligence (AGI), and represents “the first worldwide initiative that provides a legal framework for Artificial Intelligence (AI)”^[18]. The AI Act, therefore, constitutes a solid foundation for ensuring that the development of AI in the EU is ethical, legal, socially equitable, and environmentally sustainable, with the aim of contributing to the improvement of the economy, society, and the environment.

Another important social aspect, previously mentioned, concerns the possibility that some AI systems may replace workers and therefore eliminate numerous jobs. Although this practice has been common since the period of the first industrialization, when machines allowed some processes to be automated, today, thanks to the advent of AI, the process is also extending to the service sector, where, for example, chatbots and virtual assistants can replace the “human colleagues” of customer service (Frank et al. 2019). This scenario makes it quite clear that job loss does not have a purely financial impact on the population. Often, in fact, this has significant consequences on a person’s mental health and well-being, as uncertainty, which could also result from possible difficulty in reintegrating into the workforce, can lead to stress, anxiety, and depression.

In this context, governments have the responsibility to support workers through support and retraining programs, and through the regulation of the use of AI in sectors where job loss could be more severe. Finally, individuals can also take proactive measures to protect their future careers, such as developing highly sought-after skills that can help them reintegrate into the workforce.

In this era of strong AI expansion, over 50,000 people have signed an open letter^[19] calling for a pause in the development of more powerful AI systems than GPT-4 – the next model after GPT-3 and declared by OpenAI to be much more powerful than the previous version – in order to allow the definition of rules that ensure their control. Among the signatories are some of the leading figures in the digital world, such as the Turing Award winner YoshuaBengio, entrepreneur Elon Musk, and Apple co-founder Steve Wozniak. The appeal urges all AI labs to halt the development of such systems for at least six months and calls on governments to impose a moratorium on companies that refuse to comply.

During this period, efforts should be focused on developing AI safety protocols and governance systems to ensure that AI systems are accurate, safe, reliable, and fair.

6. Conclusions

The growing use of AI-enabled systems raises important ethical questions concerning fundamental social values such as privacy, security, responsibility, and equity. Indeed, although the benefits that can be derived from the integration of such technology in various fields – from medicine to education – are manifold, since its applications can have long-term consequences on the population, it is necessary to focus on the development and use of these systems.

In such a context, the awareness of the users of these platforms plays an important role, especially in algorithms that make use of continuous learning, as they are responsible for providing feedback to the system. Moreover, it is essential to carefully manage the data collection processes, in order to reduce or eliminate biases and discrimination of any kind while at the same time ensuring the protection of users and their personal data.

It is therefore important that AI researchers, developers, and users work together to ensure that this technology is used in a responsible and ethical manner. This requires not only greater awareness, but also continuous improvement and compliance with regulations and norms that allow for the safety and effectiveness of AI without acting as an obstacle to technological progress.

Bibliography

Altman Sam [@sama]. (2022). «ChatGPT Is Incredibly Limited, but Good Enough at Some Things to Create a Misleading Impression of Greatness. It’s a Mistake to Be Relying on It for Anything Important Right Now. It’s a Preview of Progress; We Have Lots of Work to Do on Robustness and Truthfulness.» Tweet. Twitter. https://twitter.com/sama/status/1601731295792414720.

Asimov Isaac. (2013). I, Robot. This paperback edition published by HarperVoyager 2013. London: HarperVoyager.

Awad Edmond, Sohan Dsouza, Richard Kim, Jonathan Schulz, Joseph Henrich, Azim Shariff, Jean-François Bonnefon, e Iyad Rahwan. (2018). «The Moral Machine Experiment». Nature 563 (7729): 59–64. https://doi.org/10.1038/s41586-018-0637-6.

BathaeeYavar. (2018). «The Artificial Intelligence Black Box and the Failure of Intent and Causation» 31.

Bernieri Christian – DPO [@prevenzione]. (2023). «Attenzione… qui abbiamo un problema: il @GPDP_IT fa il suo dovere e pone domande, chiede spiegazioni, vietando trattamenti illeciti di dati personali. Il tutto senza inventare nulla ma applicando leggi ben note a tutti. @OpenAI esclude l’Italia dall’accesso ai suoi servizi,…». Tweet. Twitter. https://twitter.com/prevenzione/status/1642034490410246147.

Best Jo. (2013). «IBM Watson: The inside Story of How the Jeopardy-Winning Supercomputer Was Born, and What It Wants to Do Next». TechRepublic. 9 settembre 2013. https://www.techrepublic.com/article/ibm-watson-the-inside-story-of-how-the-jeopardy-winning-supercomputer-was-born-and-what-it-wants-to-do-next/.

Brown Tom B., Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al. (2020). «Language Models are Few-Shot Learners». arXiv. http://arxiv.org/abs/2005.14165.

Cappellini Alberto. (2019). «MACHINA DELINQUERE NON POTEST? * BREVI APPUNTI SU INTELLIGENZA ARTIFICIALE E RESPONSABILITÀ PENALE».

Cerdan Victor, e Graciela Padilla-Castillo. (2019). «Historia del fake audiovisual: deepfake y la mujer en un imaginario falsificado y perverso». Historia y Comunicación Social 24 (settembre). https://doi.org/10.5209/hics.66293.

Cole Samantha. (2017). «AI-Assisted Fake Porn Is Here and We’re All Fucked». Vice (blog). 11 dicembre 2017. https://www.vice.com/en/article/gydydm/gal-gadot-fake-ai-porn.

Colletti Elisa. (2021). «INTELLIGENZA ARTIFICIALE E ATTIVITÀ SANITARIA. PROFILI GIURIDICI DELL’UTILIZZO DELLA ROBOTICA IN MEDICINA».

Crawford Kate [@katecrawford]. (2023). «Umm, Anyone a Little Concerned That Bard Is Saying Its Training Dataset Includes… Gmail? I’m Assuming That’s Flat out Wrong, Otherwise Google Is Crossing Some Serious Legal Boundaries. Https://T.Co/0muhrFeZEA». Tweet. Twitter. https://twitter.com/katecrawford/status/1638238113276346368.

Cremonesi Paolo. (2023). «Il Garante della privacy ha bloccato ChatGPT in Italia.» 2023. https://www.linkedin.com/posts/paolocremonesi_il-garante-della-privacy-ha-bloccato-chatgpt-activity-7047826026084626432-HXLO/?utm_source=share&utm_medium=member_desktop.

Crevier Daniel. (1993). AI: The Tumultuous History of the Search for Artificial Intelligence.

De Vita Fabrizio, Giorgio Nocera, Dario Bruneo, Valeria Tomaselli, Davide Giacalone, e Sajal K. Das. (2020). «Quantitative Analysis of Deep Leaf: a Plant Disease Detector on the Smart Edge». In 2020 IEEE International Conference on Smart Computing (SMARTCOMP), 49–56. https://doi.org/10.1109/SMARTCOMP50058.2020.00027.

Directorate-General for Parliamentary Research Services (European Parliament), Karim Lekadir, Gianluca Quaglio, Anna Tselioudis Garmendia, e Catherine Gallin. (2022). Artificial Intelligence in Healthcare: Applications, Risks, and Ethical and Societal Impacts. LU: Publications Office of the European Union. https://data.europa.eu/doi/10.2861/568473.

Frank Morgan R., David Autor, James E. Bessen, Erik Brynjolfsson, Manuel Cebrian, David J. Deming, Maryann Feldman, et al. (2019). «Toward understanding the impact of artificial intelligence on labor». Proceedings of the National Academy of Sciences 116 (14): 6531–39. https://doi.org/10.1073/pnas.1900949116.

Gillespie Nicole, Steven Lockey, Caitlin Curtis, Javad Pool, e Ali Akbari. (2023). «Trust in Artificial Intelligence: A Global Study». Brisbane, Australia: The University of Queensland; KPMG Australia. https://doi.org/10.14264/00d3c94.

Grote Thomas, Geoff Keeling. (2022). «Enabling Fairness in Healthcare Through Machine Learning». Ethics and Information Technology 24 (3): 39. https://doi.org/10.1007/s10676-022-09658-7.

Halevy Alon, Peter Norvig, e Fernando Pereira. (2009). «The Unreasonable Effectiveness of Data». IEEE Intelligent Systems 24 (2): 8–12. https://doi.org/10.1109/MIS.2009.36.

Hancell Greg. (2023). «(14) Continuous Learning is a must… | LinkedIn». 2023. https://www.linkedin.com/pulse/continuous-learning-must-greg-hancell/.

Johnson Melvin, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, et al. (2017). «Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation». arXiv. http://arxiv.org/abs/1611.04558.

LeCun Y., B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, e L. D. Jackel. (1989). «Backpropagation Applied to Handwritten Zip Code Recognition». Neural Computation 1 (4): 541–51. https://doi.org/10.1162/neco.1989.1.4.541.

LeCun Yann, Yoshua Bengio, e Geoffrey Hinton. (2015). «Deep Learning». Nature 521 (7553): 436–44. https://doi.org/10.1038/nature14539.

Levin Sam, e Julia Carrie Wong. (2018). «Self-Driving Uber Kills Arizona Woman in First Fatal Crash Involving Pedestrian». The Guardian, 19 marzo 2018, sez. Technology. https://www.theguardian.com/technology/2018/mar/19/uber-self-driving-car-kills-woman-arizona-tempe.

Luttrell Kevin, Michael Weaver, e Mitchel Harris. (2015). «The Effect of Autonomous Vehicles on Trauma and Health Care». Journal of Trauma and Acute Care Surgery 79 (4): 678–82. https://doi.org/10.1097/TA.0000000000000816.

Marchetti Dario, reg. (2023). Il miracolo di ChatGPT è scrivere testi che per un umano richiederebbero intelligenza senza averne. https://www.rainews.it/video/2023/02/il-miracolo-di-chatgpt–scrivere-testi-che-per-un-umano-richiederebbero-intelligenza-senza-averne-b4f858a2-8667-40ef-86ce-8cb4cabb0534.html.

Mattu Julia Angwin, Jeff Larson,Lauren Kirchner,Surya. (2016). «Machine Bias». ProPublica. 2016. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing.

McCarthy J, M L Minsky, N Rochester, e C E Shannon. (1955). «A PROPOSAL FOR THE DARTMOUTH SUMMER RESEARCH PROJECT ON ARTIFICIAL INTELLIGENCE», agosto.

McCulloch Warren S., e Walter Pitts. (1943). «A Logical Calculus of the Ideas Immanent in Nervous Activity». The Bulletin of Mathematical Biophysics 5 (4): 115–33. https://doi.org/10.1007/BF02478259.

Newell A., e H. Simon. (1956). «The logic theory machine–A complex information processing system». IRE Transactions on Information Theory 2 (3): 61–79. https://doi.org/10.1109/TIT.1956.1056797.

Obermeyer Ziad, Brian Powers, Christine Vogeli, e Sendhil Mullainathan. (2019). «Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations». Science 366 (6464): 447–53. https://doi.org/10.1126/science.aax2342.

Parikh George Maliha, Ravi B. (2022). «Who Is Liable When AI Kills?» Scientific American. 2022. https://www.scientificamerican.com/article/who-is-liable-when-ai-kills/.

Perez Sarah. (2016). «Microsoft Silences Its New A.I. Bot Tay, after Twitter Users Teach It Racism [Updated]». TechCrunch (blog). 24 marzo 2016. https://techcrunch.com/2016/03/24/microsoft-silences-its-new-a-i-bot-tay-after-twitter-users-teach-it-racism/.

Pigford Josh [@Shpigford]. (2023). «✨ AI Challenge: I’m Going to Build an IOS Game Using Only GPT-4 (and Probably Midjourney for Some Design Elements). I Have Basically Zero Experience Building IOS Apps. I’ve Tried to Learn the Traditional Way and Failed Every Time. May the Odds Be Ever in My Favor.» Tweet. Twitter. https://twitter.com/Shpigford/status/1637303300671275008.

Raj Rahul, Teemu Luostarinen, Eetu Pursiainen, Jussi P. Posti, Riikka S. K. Takala, Stepani Bendel, Teijo Konttila, e Miikka Korja. (2019). «Machine Learning-Based Dynamic Mortality Prediction after Traumatic Brain Injury». Scientific Reports 9 (1): 17672. https://doi.org/10.1038/s41598-019-53889-6.

Rumelhart David E., Geoffrey E. Hinton, e Ronald J. Williams. (1986). «Learning Representations by Back-Propagating Errors». Nature 323 (6088): 533–36. https://doi.org/10.1038/323533a0.

Samuel A. L. (2000). «Some studies in machine learning using the game of checkers». IBM Journal of Research and Development 44 (1.2): 206–26. https://doi.org/10.1147/rd.441.0206.

Sepas Ali, Ali Haider Bangash, Omar Alraoui, Khaled El Emam, e Alaa El-Hussuna. (2022). «Algorithms to anonymize structured medical and healthcare data: A systematic review». Frontiers in Bioinformatics 2. https://www.frontiersin.org/articles/10.3389/fbinf.2022.984807.

Shannon Claude E. (1988). «Programming a Computer for Playing Chess». In Computer Chess Compendium, a cura di David Levy, 2–13. New York, NY: Springer New York. https://doi.org/10.1007/978-1-4757-1968-0_1.

Sutton Richard. (1997). «11.2 Samuel’s Checkers Player». 30 maggio 1997. http://incompleteideas.net/book/11/node3.html.

Turing A. M. (1950). «I.—COMPUTING MACHINERY AND INTELLIGENCE». Mind LIX (236): 433–60. https://doi.org/10.1093/mind/LIX.236.433.

Vincent James. (2018). «Amazon Reportedly Scraps Internal AI Recruiting Tool That Was Biased against Women». The Verge. 10 ottobre 2018. https://www.theverge.com/2018/10/10/17958784/ai-recruiting-tool-bias-amazon-report.

Vyas Darshali A., Leo G. Eisenstein, e David S. Jones. (2020). «Hidden in Plain Sight — Reconsidering the Use of Race Correction in Clinical Algorithms». A cura di Debra Malina. New England Journal of Medicine 383 (9): 874–82. https://doi.org/10.1056/NEJMms2004740.

Wang Chongren, e Zhuoyi Xiao. (2022). «A Deep Learning Approach for Credit Scoring Using Feature Embedded Transformer». Applied Sciences 12 (21): 10995. https://doi.org/10.3390/app122110995.

1. https://www.youtube.com/watch?v=_9_AEVQ_p74 ↑
2. https://www.thats-ai.org/en-GB/units/a-brief-history-of-ai ↑
3. https://www.geeksforgeeks.org/what-is-google-brain/ ↑
4. https://research.facebook.com/publications/deepface-closing-the-gap-to-human-level-performance-in-face-verification/ ↑
5. https://openai.com/ ↑
6. https://openai.com/blog/chatgpt ↑
7. https://www.treccani.it/enciclopedia/etica_(Enciclopedia-dei-ragazzi) ↑
8. https://www.bbc.com/news/technology-33347866 ↑
9. https://www.resumebuilder.com/1-in-4-companies-have-already-replaced-workers-with-chatgpt/ ↑
10. https://www.alphagalileo.org/en-gb/Item-Display/ItemId/230890 ↑
11. https://help.openai.com/en/articles/6783457-chatgpt-general-faq ↑
12. https://en.wikipedia.org/wiki/List_of_datasets_for_machine-learning_research ↑
13. https://www.agi.it/innovazione/news/2023-04-01/chatgpt_stop_in_italia_openai_garante_privacy-20770966/ ↑
14. https://www.nhtsa.gov/sites/nhtsa.gov/files/2022-06/ADAS-L2-SGO-Report-June-2022.pdf ↑
15. https://www.azcentral.com/story/news/local/tempe-breaking/2018/06/21/uber-self-driving-car-crash-tempe-police-elaine-herzberg/724344002/ ↑
16. https://www.sae.org/site/blog/sae-j3016-update ↑
17. https://www.adalovelaceinstitute.org/resource/eu-ai-act-explainer/ ↑
18. https://edps.europa.eu/press-publications/press-news/press-releases/2021/artificial-intelligence-act-welcomed-initiative ↑
19. https://futureoflife.org/open-letter/pause-giant-ai-experiments/ ↑