State-of-the-art technologies, such as AI and robotics, can dramatically enhance educational efforts by making learning events more interactive and engaging for children, especially the younger ones. Social Assistive Robots (SARs), such as the humanoid NAO, have been proven beneficial in educational scenarios, bringing added pedagogical value, especially for students with Autism Spectrum Disorder. However, the NAO humanoid is a robot developed two decades ago, with technological limitations that predate the AI era. In the current paper, we present our integration of Large Language Models (LLM), and specifically GPT-4, into Humanoid ( \(\hbox {NAO}^{6}\) ), to build a robotic AI dietitian, capable of engaging in open dialogues with children (4–6 years old) about a healthy diet. We tested our solution in real school settings (Kindergarten) as a proof-of-concept. Future steps require more extensive experimentation and integration of computer vision to achieve more ’human-like’ learning experiences.