In recent years, artificial intelligence has achieved remarkable progress thanks to large‑scale models that can generalize and generate complex outputs. Yet transferring this potential to embodied agents reveals a critical bottleneck: even the most advanced systems still rely on pre‑existing datasets and human feedback, which are often insufficient in dynamic or unknown contexts.
To become adaptable, an agent must acquire knowledge through direct interaction with its environment. One viable strategy is to introduce intrinsic motivation mechanisms—such as curiosity and competence—to drive exploration and learning in complex settings.
While this autonomy enhances flexibility, it also makes it harder to keep the agent aligned with human goals. Alignment is already a challenge for artificial systems, and in unstructured, dynamic contexts predefined rules become inadequate.
Consequently, norms should emerge through an epistemological process grounded in experience: starting from simple, situated principles, agents can gradually construct more sophisticated rules by means of autonomous learning and cooperation with other moral agents.
Much like children learn social norms by exploring their environment and participating in collective practices, artificial agents need an educational process to achieve alignment. Following Dennett, the status of a moral agent is not innate but is progressively attributed as the ability to responsibly manage increasing degrees of freedom develops.
From this perspective, regulatory sandboxes can be viewed as pedagogical environments for AI—dynamic spaces where agents, through interaction and cooperation, shape aligned behavior as scenarios grow in complexity.
Review: The paper reframes alignment as a developmental process inspired by child socialization, offering a promising direction for building safe, autonomous agents capable of operating in unknown environments.