After the public launch of ChatGPT, the Future of Life Institute published an open letter calling for a moratorium on the development of frontier artificial intelligence (AI) systems. It was signed by over thirty thousand scholars, industry leaders, and public intellectuals. Judging by media headlines at the time, the signatories were not alone. Open AI’s launch created widespread concern regarding the rapid pace of AI’s development, and the risks it posed for humanity. Luciano Floridi seems to disagree with the reasoning behind this initiative. In his book The Ethics of Artificial Intelligence, he states: “It is because we lack a clear sense of socio-political direction that we are worried by the speed of our technological travel. We should be. Yet the solution is not to slow down, but to decide together where we want to go” (79). With adequate normative frameworks for AI, he says, we can simultaneously mitigate its worst risks and reap its greatest benefits. The book sets out to provide systematic and comprehensive guidelines for what he calls a “good AI society.”

The book is as remarkable for its breadth as it is for its depth. It does not shy away from acknowledging the enormous complexity of AI systems and the wide variety of technologies encompassed by this, somewhat amorphous, concept. Based on the recognition that AI is immersed in complex social environments, the book offers a systematic framework for ethical analysis, which takes into consideration the roles and responsibilities of a wide variety of actors. Its ambition and scope are matched by the level of detail with which he discusses the many nuanced domains and subdomains in the field of AI ethics.

The book is divided in two parts. Part I presents a historical and analytical approach to the concept of artificial intelligence. The first three chapters examine important conceptual developments in AI’s past and future. These chapters adopt an “engineering approach” to AI. Unlike “cognitive approaches,” his proposal circumvents contested theories of the nature of artificial intelligence and its—real or apparent—similarity to human intelligence. This approach focuses instead on a counterfactual conception of AI, as a technology that “behaves in a way that would require intelligence if a human were so behaving” (16).

Part II of the book develops a multilevel approach to the normative questions raised by AI. Drawing from the field of bioethics, he adopts five principles for ethical AI, which aim to streamline our fundamental normative commitments—an important task given the global proliferation of principles and frameworks for AI that occurred over the past decade. The following chapters present an array of taxonomies in the areas of ethics-related risks, categories of governance, ethical considerations for algorithms, uses of AI for social evil, uses of AI for social good, policy guidelines, and AI’s impact on climate action and the UN’s Sustainable Development Goals. Each of these chapters provides a conceptual framework to think through the different levels at which AI presents ethical challenges. Teasing these apart, while providing a clear vocabulary and taxonomy for each, constitutes a substantial contribution to both scholarly and practical debates. Most importantly, the book attempts a sober and balanced assessment of the enormous benefits that could be gained from AI alongside the major risks that we incur by allowing it to occupy increasingly many spaces in human lives and societies.

In what follows, I point to three overlooked areas which constitute ripe terrain for future research and debate.

In the first place, the book aims at taking a global perspective and advancing recommendations that are applicable around the world. However, it tends towards Eurocentrism, especially in its policy recommendations. Most of them are focused on creating bureaucratic bodies to regulate and enforce compliance with ethical standards. While this approach may prove fruitful in some contexts, there are political and practical challenges associated with the creation of such bureaucracies. These are especially pressing for countries with less technical and financial resources, and which will face greater difficulties in funding bureaucracies and endowing them with reasonable technical and democratic legitimacy. Realizing the book’s global aspirations will require further expanding on the themes of international cooperation, industry-wide standards and accountability, and more scalable measures for educating the public, government workers, and technologists.

Secondly, there is a general oversight of the problem of tradeoffs. The book is based on the recognition that developers and regulators need to balance a plurality of societal values, which are embedded in AI systems and, oftentimes, in conflict with one another. It seems to assume that an optimal or acceptable balance between these can be attained. However, tragic choices may be forced upon us. For example, the question whether AI models should be open source creates an inevitable value conflict. While open sourcing powerful models can democratize access, enable greater transparency and promote detection of dangerous flaws in the models, it also opens the door for misuse and creates further challenges for accountability. Another recognized conflict is between privacy and fairness in the context of facial recognition systems. Improving facial recognition accuracy across demographic groups to address bias enables greater surveillance of the marginalized communities that the technology already disproportionately misidentifies (see Buolamwini Reference Buolamwini2023). Value conflicts like these may be the rule rather than the exception and navigating them will be inevitable for both developers and regulators. The book leaves open the question of what it means to address these tradeoffs responsibly. While this may be a difficult task, identifying reasonable procedural constraints (such as participatory frameworks) may provide some guidance for reaching acceptable (if not always optimal) compromises.

Relatedly, the book leaves an open area for investigating the sociotechnical systems in which AI is developed, adopted, and deployed. Investigating how ethical considerations intersect with organizational factors throughout AI’s life cycle could help us ground normative discussions in the complex social environments in which these systems are embedded. Importantly, meso-level institutions like research labs, corporations, and professional associations, among others, will be responsible for wading through some of the value conflicts and tradeoffs that AI poses. Examining the conditions for legitimate decision making inside these organizations, their operational constraints, their relationship with governments, and the particularities of their market contexts could enhance our understanding of how ethical considerations are integrated into AI systems.

These are all areas where business ethicists could greatly contribute to the discussion, by providing frameworks for understanding the responsibilities of the different private actors in this ecosystem and offering normative parameters to evaluate their actions.

Some of these normative questions become more pressing given the ambiguities left open in the early chapters of the book. Floridi’s “engineering approach” explicitly avoids the challenge of formulating a precise definition of AI. In Floridi’s view, the term “AI” functions similarly to “pornography” in Justice Stewart’s concurring opinion in Jacobellis v. Ohio (378 U.S. 184). Justice Stewart held then that, while pornography cannot be intelligibly defined, he “knows it when he sees it.” Floridi adopts a comparable stance, arguing that AI is not a scientific term for which necessary and sufficient conditions can be established. This conceptual flexibility is practical and creates space for other, perhaps more productive, normative discussions. However, Stewart’s ambiguity is often criticized for casting conceptual uncertainty in a critical area, leaving free speech vulnerable to overly broad regulation, failing to shield individuals from potential harms caused by the pornographic industry, and enabling abuses by powerful actors. Similarly, Floridi’s refusal to define AI invites ambiguities that powerful actors could exploit, tailoring definitions of AI to suit their interests.

The book holds that AI should be understood as a technology that enables the unintelligent completion of tasks that would normally require human intelligence. Floridi suggests that AI reveals a distinction between “smart” and “intelligent” technologies. Smart technologies, he says, constitute a successful continuation of human intelligence by other means. They operate by breaking tasks into simple steps that don’t require intelligence (the definition of an algorithm). If AI can perform a task successfully, it means that it didn’t require intelligence in the first place. Smart technologies are, in his view, fundamentally different from what he calls Productive AI—that is, AI that produces truly intelligent behavior. Citing the success of IBM’s Watson in Jeopardy as the key example, he holds that “productive AI does not merely underperform with respect to human intelligence. It has not yet even joined the competition” (21).

Floridi’s dismissal of productive AI’s capabilities raises questions. Watson’s success in Jeopardy occurred over ten years before the book’s publication, during which major breakthroughs were already public knowledge—including the 2017 publication that introduced the transformer architecture, which became the foundation on which Large Language Models were subsequently built (Vaswani et al. Reference Vaswani, Noam and Niki2017). While it would be a stretch to claim that recent advances amount to anything like human capabilities or artificial general intelligence (AGI), denying that AI has “joined the competition” based on a decade-old example calls for an explanation of why recent developments don’t represent meaningful progress in Floridi’s eyes.

Two more recent developments could directly challenge his perspective: “agentic systems” that can pursue complex, multistep goals with minimal human oversight, and “reasoning models” that employ reinforcement learning to enhance their problem-solving capabilities. These advances have made it possible for AI systems to write functional code for nontrivial programming tasks. According to a highly respected benchmark, the duration (and therefore the complexity) of tasks they can complete autonomously has been doubling approximately every seven months (Kwa et al. Reference Kwa, Ben and Joel2025). These capabilities are, of course, the result of an extensive and resource-consuming training, all of which can be broken down to small mindless tasks—to support Floridi’s point. But as AI develops behaviors that increasingly resemble human intelligence, we must ask how much our own cognitive capabilities differ from these. Is human intelligence not also a property that emerges from the unintelligent behavior of billions of neurons?

The main difficulty posed by Floridi’s theory is that it makes no commitment regarding the distinctive features of human intelligence—an unstable position, given the book’s ambitions. While he refuses to define intelligent behavior, he also rejects the idea that AI behaves intelligently. The closest he comes to a substantiation of this statement is the claim that “intelligence will be less about solving problems than about identifying what problems are worth solving, why, for what purpose, and acceptable costs, tradeoffs and consequences” (24). But frontier systems are already developing some capabilities in this domain, like setting intermediate goals and navigating practical tradeoffs.

If AI systems continue this trajectory, it could lead to a reductio: Floridi’s theory would entail that those tasks, which he describes as the last refuge of uniquely human intelligence, never required intelligence in the first place. “With better technology we will demonstrate our total incompetence”—my father used to say, with his characteristically sardonic sense of humor. Reading Floridi’s theory in light of these developments reminded me of my father’s humorous pessimism. If taken at face value, his argument may imply more about his views on human intelligence than about artificial intelligence. And though the book shouldn’t be judged by post-publication advances, these developments show that such sweeping claims may be quickly outpaced by a dynamic technological frontier.

If we want to have a better sense of what sets humanity apart from AI, we need to develop a fully-fledged theory of agency and intelligence. Otherwise, we risk enabling powerful actors in the ecosystem to tailor regulation and safeguards to their convenience based on vague philosophical assertions about the definition of AI, and assessments of its capabilities relative to humans. Vagueness in this analysis creates risks of its own. As the technology approximates human-level performance in a growing number of tasks, views regarding our relative capabilities may shape legislation and regulation. Philosophical contributions should provide conceptual clarity to ground legislation that protects human beings from being devalued.

Alternatively, we could avoid these questions altogether and limit the discussion to substantive ethical considerations. In this respect, Floridi’s book makes a significant contribution to the field by providing thorough frameworks and taxonomies for the ethics of AI.

Given the breadth of the field, and the technical and philosophical depth of each domain, these tasks grow more difficult every year. This is a challenge that the entire discipline must confront and to which we will, I hope, learn to adapt. Just as we should be responsible for counteracting hype and unsubstantiated claims about the technology’s capabilities, we should also seek to provide solid conceptual foundations that do not rely on ambiguity and contested assumptions about the uniqueness of human intelligence. If we can contribute to this endeavor, Floridi’s goal of outlining the conceptual basis for a Good AI Society will be greatly advanced.

Floridi’s normative frameworks will undoubtedly offer compelling resources and insights to motivate deeper philosophical engagement and to guide ethically informed practical discussions. The book will provide rich input for any researchers in business ethics who are interested in the responsibilities of organizations that develop, adopt, deploy, or regulate AI, and the role of frontier technologies in channeling large-scale societal transformation. There is space for rich interdisciplinary engagement in these areas and business ethicists are uniquely situated to contribute to them.