Safeguarding in the Age of Artificial Intelligence: Why AI Governance Must Begin with Human Vulnerability

Tehilla Shwartz Altshuler and Michael Sierra

Artificial intelligence is no longer a futuristic technology discussed only in laboratories or among engineers. AI systems already mediate our access to information, shape our interactions with institutions, recommend medical treatments, influence financial decisions, and increasingly participate in the emotional and cognitive dimensions of daily life. The rapid diffusion of generative AI, predictive systems, conversational agents, and autonomous technologies has generated understandable excitement regarding efficiency, innovation, and economic growth. Yet it has also exposed a deeper question: what does it mean to protect human beings in a society increasingly structured around intelligent systems?

In our new book, Safeguarding in the Age of Artificial Intelligence, published by the Israel Democracy Institute in Hebrew, we attempt to address this question through the concept of “safeguarding” (“muganut” in Hebrew). Rather than focusing exclusively on abstract existential risks or purely technical debates about algorithms, the book shifts the conversation toward the concrete human experience of vulnerability in AI-mediated environments.

Our central argument is straightforward but ambitious: the key challenge of AI governance is not merely technological risk, but human vulnerability.

From “AI Risks” to Human Vulnerabilities

Much of the global discussion around AI governance concentrates on categories such as “systemic risk,” “alignment,” “bias,” or “safety.” While these debates are important, they often overlook the lived realities of individuals interacting with AI systems on a daily basis. The book, therefore, proposes reversing the analytical perspective. Instead of asking only “What risks does AI create?”, we ask: “What kinds of vulnerabilities does AI enable or intensify?”

This shift is more than semantic. It changes the focus of governance from abstract technological capabilities to the protection of human resources and capacities: bodily integrity, emotional well-being, cognitive autonomy, financial stability, and social belonging.

We define safeguarding as a condition in which essential human resources are protected from harm arising either from the design of technologies themselves or from malicious uses of those technologies by individuals and organizations. Safeguarding, in this sense, is not merely cybersecurity, privacy, or platform moderation. It is a multidimensional ecosystem of interventions involving law, education, technological design, institutional accountability, and social support.

A New Typology of Vulnerability

One of the book’s central contributions is the mapping of five major categories of vulnerability in AI environments: Physical vulnerability, Emotional and psychological vulnerability, Cognitive vulnerability, Financial vulnerability, and Social and group-based vulnerability. Some of these categories are already familiar. AI systems can facilitate fraud, identity theft, harassment, discrimination, or invasive surveillance. But we also focus on emerging forms of vulnerability that remain underexplored in mainstream policy discussions.

Particularly significant is the discussion of cognitive vulnerability. We identify two related dangers: first, the erosion of people’s ability to independently assess reality in an environment saturated with synthetic content, misinformation, and algorithmically amplified narratives; and second, the gradual weakening of higher-order cognitive skills due to overreliance on AI systems for reasoning, creativity, and decision-making.

This concern resonates strongly with ongoing debates in cybersecurity and information integrity. The challenge is no longer only whether systems can deceive us, but whether continuous interaction with intelligent systems may transform the very conditions of human judgment and autonomy.

The book also addresses emotional vulnerability. AI companions, emotionally responsive chatbots, recommendation systems, and immersive interfaces may generate dependency, false intimacy, social isolation, or manipulative behavioral dynamics. In this sense, the challenge of AI governance extends far beyond technical robustness or data security. It reaches into the architecture of human relationships and emotional life.

Differential Vulnerability and Social Inequality

Another core argument of the book is that vulnerabilities are not distributed equally across society. We therefore focus on four groups within Israeli society: children and youth, Arab citizens, the ultra-Orthodox community, and older adults. Each group experiences AI systems through different combinations of institutional barriers, cultural contexts, cognitive conditions, and levels of digital literacy.

For example, children may be especially exposed to manipulative recommendation systems, synthetic intimacy, and developmental cognitive harms. Older adults may face financial scams and accessibility barriers. Minority communities may encounter algorithmic discrimination or institutional mistrust.

This approach challenges the assumption, common in many policy frameworks, that technological harms affect everyone in the same way. Instead, we advocate what we describe as “differential universality”: common principles of protection adapted to diverse social realities.

For cybersecurity scholars and practitioners, this insight is highly significant. Security is often conceptualized as a universal technical problem. Yet AI governance increasingly requires contextual and sociological understandings of risk, resilience, and institutional trust.

Beyond Regulation Alone

Importantly, the book does not present regulation as a magic solution. We repeatedly emphasize that safeguarding cannot be achieved through a single intervention. Instead, we propose a layered governance model involving: regulation and enforcement, educational and literacy programs, responsible technological design, institutional support systems, auditing and oversight mechanisms, and even the possibility that the control may be performed by the machine over the human, rather than the opposite.

The framework is deliberately pragmatic. We recognize that technological acceleration often outpaces lawmaking and empirical certainty. Yet we reject the idea that uncertainty justifies passivity. In conditions of rapid technological change, inaction itself becomes a risky political choice. This position has important implications for cybersecurity governance. Waiting for perfect evidence before acting may leave societies structurally unprepared for cumulative harms that become visible only after they are deeply embedded.

Why This Discussion Matters Now

AI governance debates are often polarized between techno-optimism and catastrophic speculation. Our goal in writing this book was to avoid both extremes. Instead, we sought to reconnect discussions about artificial intelligence with concrete human experiences of vulnerability, inequality, and institutional responsibility. The framework we propose combines law, sociology, cybersecurity thinking, ethics, and public policy in an attempt to build a more human-centered vocabulary for AI governance.

For the cybersecurity community in particular, the book offers an important reminder: safeguarding people in AI environments cannot be reduced to protecting systems alone. Cybersecurity increasingly concerns the protection of cognitive autonomy, emotional resilience, social trust, and democratic integrity. In this sense, one of the defining governance challenges of the coming decades may not simply be how to build more powerful AI systems, but how to ensure that human beings remain protected, autonomous, and resilient in a world increasingly shaped by them.