1.1. Artificial Intelligence: stages of development and transformation

What is AI? In concise terms, it may be understood as the attempt to simulate, through computational systems, capacities that correspond to, or may eventually surpass, human intelligence. Although multiple definitions coexist, a widely cited formulation is that of the American mathematician Richard ^{Bellman (1978}), who characterises AI as the automation of activities associated with human thinking, such as decision-making, problem solving, and learning.

This definition raises a second question: what is intelligence? Intelligence is a multifaceted concept and, much like AI itself, remains subject to diverse interpretations (^{Russell & Norvig, 2021}). According to the American Psychological Association (APA), intelligence can be described as the capacity to infer information, learn from experience, adapt to the environment, and understand and appropriately employ thought and reason (^{APA 2018}). Yet what, then, still distinguishes human intelligence from AI? Even though AI is capable of performing highly complex tasks, there remains something singular about the human mind.

A third question thus emerges: that of consciousness. Consciousness may be defined as the capacity to experience, in an aware manner, the self and the surrounding environment (^{Stanford Encyclopedia of Philosophy, 2014}). Consciousness-understood as the subjective experience of the world-constitutes a key feature separating human intelligence from AI. This capacity, which combines an understanding of the world with self-awareness, has not yet been attributed to AI systems and remains without a definitive explanation, generating both concern and enthusiasm (^{Baerren, 2025}). For ^{Damásio (2000}), consciousness is an entirely private, first-person phenomenon that unfolds within the mind, and this understanding makes it possible, at least from a current philosophical and scientific standpoint, to draw a line between AI and biological intelligence.

Although the roots of AI can be traced back to Ancient Greece, particularly to Aristotle (384-322 BCE) and the introduction of deductive reasoning, which provides conceptual foundations for this technology (^{Rai, 2024}), the analysis developed here begins with the contribution of Alan Turing, mathematician and computer scientist, as illustrated in Figure 1.

In 1957, Bellman formalised a class of sequential decision problems known as Markov Decision Processes (MDPs), which are fundamental for understanding decision-making in uncertain and dynamic environments and are now widely applied in robotics and automation.

Source: Adapted from ^{Daylight (2015}), ^{Smith et al. (2006}), ^and Rai (2024)

Figure 1:  Diachronic trajectory of AI evolution 

1.2. Classification of AI: Functionality and Capacity

According to ^{Hassani et al. (2020}), AI may be defined as a computational system capable of reproducing human cognitive processes, drawing on data from multiple sources in order to make decisions and learn from patterns. Like all systems and machines built upon it, AI can be classified in several ways, including (but not limited to) functionality, domains of application, and level of autonomy. In this subsection, two distinct classification dimensions are examined, functionality and capacity, as proposed by Hassani.

The first refers to the “mind” of AI, that is, its ability to think with greater or lesser autonomy and the extent to which it resembles human thought. Within this dimension, four subcategories can be identified: Reactive Machines, Limited Memory Machines, Theory of Mind, and self-aware AI, represented in Figure 2.

Source: Adapted from ^{Hassani et al. (2020)}.

Figure 2:  Classification of AI by functionality 

The second classification category is grounded in technological capacity and distinguishes three subtypes: Artificial Narrow Intelligence (ANI), Artificial General Intelligence (AGI), and Artificial Superintelligence (ASI), shown in Figure 3.

Source: Adapted from ^{Hassani et al. (2020)}.

Figure 3:  Classification of AI by capacity 

1.3. Technological advances and language processing models

According to Jeff ^{Crume (2024}), before the consolidation of Machine Learning (ML), AI relied predominantly on traditional programming approaches. ML, which gained prominence from the 2010s onwards, marked a turning point by enabling systems to learn from large datasets and prove particularly effective in prediction and pattern-detection tasks, with relevant applications in cybersecurity and information analysis. Deep Learning (DL), developed in the same period and based on deep neural networks inspired by the functioning of the human brain, exponentially expanded processing and recognition capabilities, supporting substantial progress in the simulation and mimicry of cognitive functions.

The 2020s witnessed the consolidation of Generative AI (GenAI), now embedded in widely deployed applications such as recommendation systems (e.g. Netflix, Spotify) and virtual assistants (e.g. Alexa, Siri). In 2022, the public release of ChatGPT - built on OpenAI’s GPT model family, initiated in 2015 and iteratively developed since 2018 - represented a decisive milestone in the popularisation of these systems. Grounded in Foundation Models (FM) trained on massive volumes of data and still situated within the scope of Artificial Narrow Intelligence (ANI), these systems inaugurated a new stage in AI by placing Large Language Models (LLM) at the centre of the debate, distinguished by their capacity to engage in dialogue, translate, summarise and automatically generate complex texts. Their diffusion not only democratised access to advanced tools and turned them into a global cultural phenomenon, but also fuelled the emergence of multimodal applications capable of generating images, audio and video, giving rise to phenomena such as deepfakes.

The reach of these technologies goes far beyond strictly academic domains, as they are now embedded in writing tools, search engines, educational platforms, entertainment applications and decision-support systems (^{Rashid and Kausik, 2024}). This ubiquity helps explain why LLMs are often considered emblematic of the current phase of AI, while simultaneously fuelling debates on creativity, authenticity and accountability. Although these developments confirm the transformative nature of the technology, they also raise a series of ethical, political, legal and social challenges, which will be examined in the following section.

2.1. Beyond the ethical void: from uncertainty to trust

As the replacement of human labour by machines in the name of resource optimisation becomes a reality, questions arise regarding the manner in which human-technology interactions are being configured (^{Rashid & Kausik, 2024}). As Rashid and Kausik emphasise, it will be necessary to find a balance between innovation and ethical principles in order to ensure that this relationship maximises benefits while minimising risks.

In July 2020, the High-Level Expert Group on AI (AI HLEG) presented to the European Commission the Assessment List for Trustworthy AI (ALTAI), the final version of a set of recommendations for AI development. According to this document, in addition to complying with existing legal frameworks, an AI system can be considered trustworthy if it adheres to seven ethical principles, represented in Figure 4.

Source: Adapted from ^{Floridi (2019)} and the ^{European Commission (2019)}

Figure 4:  Seven ethical principles for trustworthy AI 

According to Luciano Floridi, a member of the HLEG, ethical uncertainty generates both unnecessary risks and excessive caution, preventing society from fully benefiting from the technology, whether through underuse, disproportionate use or malicious appropriation (^{Floridi, 2019}). The guidelines seek to steer AI development on the basis of political and social consensus, involving broad deliberation and a long-term view of its purposes and priorities. Despite their specifically European origin and their integration within a wider landscape of ethical frameworks, this value system aspires to global reach. To succeed, however, it will have to overcome a range of normative, geopolitical and social challenges (^{Floridi, 2019} and ²⁰²²). The following pages address these challenges in detail.

2.2. Algorithmic decision-making: autonomy and the crisis of responsibility

As decision-making processes are increasingly delegated to AI systems, new debates emerge regarding where responsibility ultimately lies. From autonomous vehicles - capable of taking decisions within fractions of a second, with direct, immediate and potentially irreversible consequences for human life - to algorithms used in medical diagnosis, judicial decision-making or recruitment processes, it becomes ever more urgent to determine who is accountable for the errors, harms or injustices resulting from such choices (^{Han, 2015}).

Situated within the so-called Fourth Industrial Revolution, characterised by the centrality of automation and the replacement of human tasks with more efficient processes, AI systems herald profound transformations across multiple sectors (^{Rashid & Kausik, 2024}). From healthcare to the military, and including telecommunications, finance, education, manufacturing, agriculture, energy and the automotive industry, AI has driven an exponential increase in productive capacity and accelerated decision-making processes, including those in the judicial domain (^{Vasdani, 2020}). However, this progress is accompanied by a critical problem of interpretability: systems make decisions but cannot explain in a transparent manner the reasons underlying those decisions, which should in itself - and especially from a legal standpoint - render such decisions unacceptable (^{Benjamins & Salazar García, 2019}).

Moreover, AI currently operates only in an apparently autonomous fashion: “it depends on human input for data collection, classification and maintenance and, although it can identify patterns, it remains unable to grasp their meaning, as in the case of cancer cells” (^{Dignum 2023}, p. 199). This reinforces the argument for the need to assign responsibility for the decisions produced. Regardless of the ongoing debate on whether a “robot legal personality” should be created (^{Benjamins & Salazar García, 2019}, p. 92), Virginia Dignum (2023) emphasises that AI systems are artefacts created by human beings and that ultimate responsibility for their use must rest with those who design and deploy them.

2.3. Three Key Challenges in AI Regulation

Recognizing that the debate surrounding AI regulation spans multiple dimensions that cannot be exhaustively covered here, this study prioritizes a synthetic overview of three structural legal and political challenges that constrain its implementation.

The first challenge is the so-called pacing problem (^{Yordanova & Bertels, 2024}). The interval between the proposal, voting, and enactment of legislation often allows technology to outpace the law, creating a near-perpetual state of legal vacuum conducive to unregulated development and inadequately monitored application and use. This asymmetry highlights not only the disparity between the lifecycle of AI technologies and political or human cycles, but also underscores the necessity for political actors to ensure regulatory continuity and oversight regardless of shifts in political power.

The second challenge concerns the problem of taxonomy, emphasized by the HLEG (^{Yordanova & Bertels, 2024}): the absence of a single, consensual definition of AI. The plurality of concepts and definitions generates ambiguity regarding the object of regulation, compromising normative clarity and effectiveness. As ^{Krafft et al. (2020}) point out, this conceptual fragmentation complicates policy formulation, since researchers tend to adopt functional definitions focused on the technical capabilities of current technologies, whereas policymakers favour comparisons with human cognition and behaviour projected onto hypothetical scenarios. This divergence can lead to an overemphasis on future risks at the expense of addressing current challenges (^{Krafft et al., 2020}).

The third challenge relates to the transnational nature and scale of AI, which contrasts sharply with the absence of centralized global governance (^{Tallberg et al., 2023}). Despite legal advances promoted by the EU (AI Act, GDPR), the US (National AI Initiative Act), and China (AI Ethics Code) - among others - and private sector efforts towards self-regulation (Meta’s GenAI Content Risk Standards), the current regulatory framework proves insufficient to manage the cross-border externalities of AI (^{Tallberg et al., 2023}, p. 9). Moreover, this multiplicity of initiatives contributes to an increasing “mushrooming of declarations” (^{Floridi, 2019}), producing inconsistencies and confusion among stakeholders. This process results in the creation of a “supermarket of principles” (^{Floridi, 2019}, p. 2), a system of selective preferences and pick-and-choose that hampers adaptation to a common ethical-normative framework (^{Floridi, 2019}).

3.1. (Dis)Information, Filter Bubbles, and Opinion Manipulation

Algorithmic communication - the management of information through systems that select, prioritize, and personalize content - presents serious challenges to democracy. By facilitating the diffusion of (dis)information, it has immense potential to corrode the shared knowledge base that underpins consensus and democratic institutions (^{Christiano, 2022}). As Harari (²⁰²⁴) observes, the core issue at hand is primarily a network problem, more specifically an information problem. The distortion of information on which human societies depend inevitably compromises the quality of our decisions, regardless of intelligence levels.

In 2011, Eli Pariser warned of the algorithmic logic of platforms like Google and Facebook, now common across major platforms, where content personalization delivers the information users want rather than need, contributing to unprecedented intellectual isolation. This leads to the formation of what Pariser terms filter bubbles: information spheres that restrict access to diverse information, reinforce pre-existing tendencies, and isolate individuals from the community (^{Pariser, 2011}). This process progressively erodes the common cognitive heritage that supports social consensus and democratic institutions.

While algorithmic manipulation is complex, Thomas ^{Christiano (2022}) identifies three processes manifesting in the public sphere: filtering, hypernudging, and microtargeting. Filtering involves the selection and presentation of information to users, limiting access to alternative content and reinforcing filter bubbles. Hypernudging refers to the continuous reconfiguration of accessible content, influencing choice patterns without users’ awareness. Microtargeting involves directing messages to specific segments by exploiting personal data and vulnerabilities to enhance persuasive effects.

As the digital environment matures, social networks and digital platforms have consolidated not only as channels for socialization, commerce, and entertainment but also as tools for ideological dissemination and political mobilization (e.g., the Cambridge Analytica case). Although these processes need not be manipulative or harmful, the algorithmic communication underlying these platforms amplifies manipulation potential, given the massive scale of processed data, precision, and speed of adaptation (^{Christiano, 2022}). The result is the control of information flow, conditioning of individual autonomy, and erosion of full civic participation.

3.2. Erosion of Public Debate and Democratic Plurality

When analysing the challenges facing public debate and democratic plurality, it is crucial to consider how epistemic vulnerabilities in certain contexts, such as disadvantaged communities, areas with limited educational access, or environments saturated with (dis)information, can impair individuals' capacity for critical thinking and advocacy of their interests (^{Christiano, 2022}). When some citizens lack essential cognitive tools, like access to quality information, critical thinking skills, or reliable communication platforms, while others benefit from contexts that encourage learning, engagement, and access to diverse perspectives, inequality emerges in political sophistication and power dynamics that undermines democracy’s foundational principle of equality. This disparity is particularly evident in algorithmic communications, where individuals with more resources dominate discourse, exacerbating pre-existing asymmetries. Consequently, the interplay of conflicting interests and varying political sophistication fosters conditions ripe for manipulation, posing significant threats to vibrant public debate and the consolidation of genuinely pluralist democracy (^{Christiano, 2022}).

However, challenges to public debate and democratic plurality extend beyond information access and political sophistication asymmetries (^{Couldry & Mejías, 2019}). Central to this dynamic is a rising class that, controlling strategic resources such as data and technology, is shaping new power relations.

3.3. Algocracies and the Eclipse of Democratic Judgment

The current concentration of AI dominance lies mostly with the GAFAM (Google, Amazon, Facebook/Meta, Apple, and Microsoft) and China’s BAT (Baidu, Alibaba, & Tencent), whose competitive advantage is based on near-exclusive access to massive proprietary data volumes. This setting fosters what some authors term a technological oligopoly characterized not only by limited competition but also by the risk of AI innovation remaining locked in and accessible to society only on terms defined by these companies’ commercial interests (^{Benjamins & Salazar García, 2019}). Such concentration of economic and technological power may result in a public sphere subordinated to algorithmic logics geared toward profit maximization at the expense of collective interests (^{Han 2022}). ^{Van Dijck et al. (2021}) criticise what they call the platform society, the role of digital platforms in the organisation of contemporary societies, affecting labour relations, social and civic practices, and affecting democratic processes.

Shoshana ^{Zuboff (2019}) describes the emergence of “instrumentarian power” (p. 37), a logic that organizes, aggregates, and standardizes society, replacing politics and democracy with the certainty of computers and social conformity. Zuboff identifies Google and Facebook as pioneers of this model, resisting corporate governance and rejecting democratic demands, securing their claims through economic and political strength. The outcome is a brutally efficient system that reduces space for democratic deliberation, public debate, and individual self-determination, subverting democracy from within and displacing popular sovereignty (^{Zuboff, 2019}).

The term algocracy - government by algorithms - emerges in this context. It denotes a form of governance mediated by algorithms that critically challenges ideals of pluralism, democratic participation, public debate, transparency, and institutional accountability. John ^{Danaher (2016}) highlights a problem deeper than mere concealment: opacity. Beyond concealment, which impedes accountability, the opacity of algorithmic systems diminishes society’s ability to exercise scrutiny due to the potential incomprehensibility of these processes for human reasoning. Herein lies the problem of legitimacy: growing reliance on algorithmic decisions threatens public governance not only by limiting accountability and transparency but primarily by making deliberation and scrutiny dependent on processes beyond human understanding (^{Danaher, 2016}).

If indeed we are moving toward a world governed by algorithms, digital literacy ceases to be optional and becomes essential: emerging as a crucial instrument for conscious and effective civic participation. It is in this spirit that the final chapter of this study is approached, where it will be examined how, faced with technological forces potentially greater than ourselves, we can learn to coexist, reflect, and act to ensure that our interactions with AI are grounded in ethics, responsibility, and full participation.

4.1. Digital Literacy as a Critical Competence in the 21st Century

Accepting that the system described by ^{Zuboff (2019}) constitutes “a threat to human nature in the 21st century akin to industrial capitalism’s threat to the natural world in the 19th and 20th centuries” (p. 15), we have an individual and collective moral and ethical obligation to acquire the know-how to navigate this ecosystem safely. Autonomy of choice is not a future projection: today, it inherently depends on our degree of digital literacy.

As the human and digital increasingly merge to become almost inseparable, it is not an exaggeration to state that AI constitutes “a social revolution occurring in real time” (^{Beam, 2025}). From candidate selection to credit eligibility, from the content we consume to judicial decisions, AI’s incorporation into modern life architecture transcends the individual or sectoral sphere. The intellectual fabric of human societies is undergoing profound change as our modes of interaction and thought begin to reflect mechanized logics - with deep effects on empathy (^{Matzner, 2022}), intersubjectivity, and the social bonds that structure community life (^{Beam, 2025}).

This passage captures the essential understanding of digital literacy as a critical, ethical, and educational imperative in the AI era, emphasizing its role in enabling responsible participation and safeguarding autonomy.

It is also important to note that the digital environment operates as a network: unlike traditional television, which flowed only from producer to viewer, digital flows are bidirectional. Every interaction leaves traces; each click, like, or share communicates personal information to content producers; often without the user’s awareness. Understanding these behind-the-screen dynamics is crucial, and users must be fully aware of what they are giving up, notably their privacy, when they are online.

Our model of interaction with AI must evolve from immediate and indiscriminate access to critical, informed, and responsible participation. Digital literacy should be cultivated daily: knowing how to identify risks, investigate recommendations, verify information accuracy, diversify the AI tools accessed, and reflect on digital habits, beginning to take an active role before technology. Such practices can have a profound impact by transforming citizens into conscious and reflective participants, abandoning the role of mere receivers, consumers, or products. In light of these challenges, the focus shifts to reflecting on how to conceptualize the human-machine relationship.

4.2. Teaching, Ethics, and Democratic Sustainability

The concept of literacy must expand. Learning to read and write is no longer sufficient: integrating the digital component into school curricula from basic to higher education is imperative, establishing it as a permanent teaching element. As ^{Dignum (2023}) observes, digital literacy should not remain confined to engineering fields; it must become a transversal element in education, capable of cultivating critical citizens aware of technological impacts and capable of responsible and ethical participation in the digital society.

A lack of proper guidance opens space for distorted learning, exposing individuals to significant risks due to their susceptibility within an ecosystem overloaded with sometimes antagonistic, inaccurate, and fallacious information. Although young people are familiar with technology, they need to develop competencies to critically evaluate available information, identify risks, and act responsibly in the digital environment. Furthermore, invoking the principle of intergenerational solidarity, digital literacy must be accessible to all: unlike digital natives, a significant portion of the global population was not born immersed in technology and requires greater attention to the risks of interconnectedness.

For AI not to become an obstacle to critical thinking, education must go beyond technical transmission and promote literacy that preserves human intellectual autonomy. Initiatives such as AI4People - An Ethical Framework for a Good AI Society (^{Floridi et al., 2018}), the Elements of AI program launched in Finland to democratize technological knowledge (^{Dignum, 2023}), and the work by the Center for Humane Technology reflect efforts toward empowerment that transcend technique, fostering critical reflection and discernment essential for conscious citizenship in the AI era.

What is required, therefore, is not the prohibition of AI use - especially in academic contexts, where restriction may be counterproductive - but rather its integration into education and society, recognizing its inevitability. Failure to do so subjects users to unsupervised learning processes that tend to foster the illusion of choice and the erosion of subjectivity, with the added risk of a significant decline in the intellectual and creative capacities of new generations (^{Stezhko & Khmil, 2023}). As ^{Dignum (2023}) notes, this challenge demands diversification of knowledge, incorporating contributions from fields such as philosophy, law, and economics to train citizens and professionals capable of articulating technical skill, ethics, and social responsibility. ^{Unwin (2021}) argues that AI should be used ethically in development contexts, emphasising that the focus should not only be on technological innovation, but also on equity, human rights and the empowerment of vulnerable populations. AI can contribute to inclusive and sustainable development.

The key lies in enhancing human intelligence rather than replacing it, following an Intelligence Augmentation logic (^{Hassani, 2020}) whereby AI acts as a partner to our cognitive faculties: technology not as a suppressor of human thought autonomy, but as an amplifier that opens new possibilities for human cognitive performance. From this premise, we advance to reflecting on one of the structuring dilemmas in the AI era, the balance between security and liberty, which constitutes the final topic addressed.

4.3. Balance Between Security and Freedom

This balance is not easily achieved; rather, it is a double-edged sword. The classic tension between security and freedom intensifies as societies digitalize and deepen their integration in an increasingly interconnected world. Technology use for monitoring and surveillance can enhance collective security - through crime prevention, intelligent management of urban systems, or rapid response to emergencies - but simultaneously threatens fundamental freedoms, particularly privacy and individual autonomy, subjecting citizens to forms of permanent control and risks of algorithmic discrimination, whose effects, as noted earlier, produce asymmetrical impacts across social groups.

Digital monitoring and data analysis technologies undoubtedly offer tangible and immediate public safety benefits. Algorithmic surveillance tools contribute to crime prevention, rapid threat detection, and efficient crisis management (^{Rashid & Kausik, 2024}). This promise of greater efficiency and predictability partially explains societal acceptance, especially when associated with protecting essential goods such as life, public safety, or the environment. Nonetheless, this promise carries a structural danger: the normalization of permanent surveillance and the redrawing of boundaries between public and private spheres (^{Zuboff, 2019}). The dilemma is that in the name of security, a control architecture is legitimized that undermines the democratic principles it ostensibly protects.

Today, Nick Bostrom’s caution that humanity behaves like children playing with a bomb (2016) resonates with renewed urgency. This vulnerability is reinforced by ^{Harari (2024, 2025}), who reminds us that the greatest threat lies not only in AI’s potential to create agents more intelligent and creative than ourselves but also in this process reflecting our own nature - competitive, impatient, and often inclined to prioritize progress over safety. ^{Floridi (2019}) also reminds us that this trend is not unprecedented: both the climate crisis and the interference of digital platforms in democracy have already exposed risks associated with innovation lacking prior planning.

^{Benjamins and Salazar García (2019}) push this reflection to its hypothetical limit: if there is a deadline to confront the major challenges posed by AI, it may coincide with the so-called Singularity, the moment when machines surpass human intelligence. Even if such a horizon never materializes, it is imperative that human societies develop mechanisms for collective deliberation, with clear concepts and solid frameworks, lest critical decisions affecting everyone are made too late or left to actors indifferent to the common good. For if human power relies on our capacity for large-scale cooperation, it is also this structure that predisposes us to misuse it recklessly (^{Harari, 2024}).