Technical fixes alone can't solve AI alignment and ethics challenges

With global attention intensifying around the risks and governance of AI systems, a new study argues that current approaches remain too narrowly focused on technical solutions, overlooking deeper questions about human agency, responsibility, and institutional norms.

Published in AI & Society, the study titled "Agency and alignment: toward a normative architecture for human–AI interaction" proposes a fundamental shift in how alignment is conceptualized, moving away from the idea that machines must learn or internalize human values and toward a framework where AI is structurally embedded within human systems of reasoning, law, and accountability.

Rethinking AI as an extension of human agency, not an independent actor

The study challenges a widely held assumption in artificial intelligence discourse: that AI systems should be treated as autonomous agents capable of learning, interpreting, and applying human values. Instead, it proposes a radically different view, positioning AI as a "teleological extension" of human agency.

Under this framework, AI does not act independently but functions as part of a broader system of human action. Its role is to mediate, support, and extend human decision-making rather than replace it. This perspective shifts the focus from what AI is to what it does within real-world contexts, such as assisting judges, supporting medical diagnosis, or optimizing administrative workflows.

The authors argue that treating AI as a separate moral or decision-making entity risks misunderstanding its actual function and inflating expectations about its capabilities. In reality, most AI systems lack consciousness, intentionality, and the ability to reason autonomously. Instead, they operate as complex statistical tools designed and guided by human objectives.

This reconceptualization has major implications for the alignment problem. Rather than asking how machines can learn human values, the study suggests that the real challenge lies in ensuring that AI systems are properly integrated into human environments where goals, responsibilities, and norms are already defined.

The concept of "extended human agency" highlights that modern decision-making is increasingly distributed across humans, technologies, and institutions. AI systems shape how decisions are made by structuring information, ranking options, and influencing outcomes. However, they do so within frameworks created and controlled by humans.

This means that alignment should not be understood as a property of the machine itself but as a characteristic of the broader system in which the machine operates. By embedding AI within human-centered processes, the focus shifts toward maintaining human control, oversight, and accountability.

Practical autonomy and the limits of technical alignment models

Existing approaches to AI alignment fail because they treat values as data that can be learned, optimized, or encoded. While techniques such as reinforcement learning, preference modeling, and constitutional AI attempt to align systems with human goals, they rely on the assumption that these goals can be clearly defined and quantified.

The authors argue that this assumption is fundamentally flawed. Human values are not fixed or universally agreed upon; they are shaped by cultural, historical, and institutional contexts. As a result, attempts to encode them directly into AI systems risk oversimplification and bias.

To address this challenge, the study introduces the concept of "practical autonomy," a philosophical framework rooted in the idea that human action is guided by reasons, norms, and the capacity for reflection. Practical autonomy is not about freedom from constraints but about the ability to justify decisions within a shared system of reasoning.

This concept reframes alignment as a normative problem rather than a purely technical one. AI systems must operate within environments where decisions can be explained, challenged, and justified. This requires more than accurate predictions or optimized outcomes; it demands integration into structures that support accountability and critical evaluation.

The study emphasizes that AI systems themselves do not need to possess autonomy in this sense. They do not need to understand or endorse ethical principles. Instead, their outputs must be interpretable within human frameworks of reasoning, allowing human agents to take responsibility for decisions made with AI support.

This distinction is crucial in high-stakes domains such as law and healthcare, where decisions carry significant ethical and social consequences. In these contexts, alignment cannot be reduced to performance metrics or statistical accuracy. It must ensure that actions remain meaningful within systems of justification and responsibility.

The research also highlights the risks of relying solely on technical solutions. Data-driven models may reproduce existing biases, reinforce dominant perspectives, and fail to capture the complexity of human values. Without a broader normative framework, such systems risk producing outcomes that are efficient but unjust or socially harmful.

The case for a 'Normative Interface' in AI system design

To bridge the gap between technical systems and human norms, the study introduces the concept of a "normative interface." This idea represents a structural approach to alignment, focusing on how AI systems are embedded within institutional and social contexts.

A normative interface is described as a design-level architecture that connects machine outputs with human systems of reasoning, ensuring that AI-generated actions remain intelligible, contestable, and accountable. It serves as the point where technological processes intersect with legal, ethical, and procedural frameworks.

The study identifies three core components of this interface. First is teleological coherence, meaning that AI systems must operate within clearly defined purposes that align with human goals. Second is normative intelligibility, requiring that outputs can be interpreted and evaluated within shared standards. Third is accountability, ensuring that responsibility for decisions remains with human actors.

This framework moves beyond the idea of transparency as a purely technical feature. While explainability remains important, the study argues that it must be understood in a broader sense, encompassing the ability to situate AI outputs within systems of justification and institutional norms.

The legal domain is presented as a key example of how normative interfaces can function in practice. Legal systems already operate as structured environments where decisions are guided by rules, interpreted through reasoning, and subject to accountability. Integrating AI into such systems requires ensuring that these fundamental characteristics are preserved.

The study warns that without such structures, the use of AI could undermine the very foundations of accountability. Decisions influenced by opaque systems may become difficult to attribute, raising questions about who is responsible for errors or harm. This concern is particularly relevant in contexts where AI systems are used to support or automate decision-making processes. If responsibility becomes diffuse or unclear, it could erode trust in institutions and weaken the legitimacy of outcomes.

Alignment as an institutional and ethical challenge, not just a technical one

The findings suggest that alignment should be understood as an ongoing process rather than a one-time solution. As AI systems evolve and are applied in new contexts, their integration must be continuously evaluated and adjusted to ensure that they remain consistent with human values and societal goals.

The study also highlights the importance of interdisciplinary collaboration. Addressing the challenges of AI alignment requires input from fields such as philosophy, law, social sciences, and ethics, in addition to technical expertise. This reflects the complex nature of the problem, which spans both technological and human dimensions.

The research warns against narratives that portray AI as either an all-powerful solution or an uncontrollable threat. Such perspectives can obscure the role of human agency and responsibility, leading to misplaced trust or unnecessary fear. It advocates for a balanced approach that recognizes both the potential and limitations of AI. By embedding systems within robust normative frameworks, it is possible to harness their capabilities while maintaining human control and accountability.