The advent of LLMs in higher education has profoundly transformed not only the production of knowledge, but also the ways of teaching, assessing and learning. In this context of generative uncertainty, we present the ZHIP framework (Heuristic Zone of Pedagogical Intervention), a proposal aimed at reconfiguring teacher assessment in AI-mediated environments. It articulates two dimensions: the Serendipity Yield (SY), which measures the heuristic value of unexpected but fertile ideas generated by AI, and the Assisted Pedagogical Materiality, which delimits the threshold of teacher intervention according to the epistemic risk detected in the assisted production. Through a two-dimensional matrix, the model defines four intervention zones—consolidation, canonical risk, fruitful exploration and creative distortion—that allow for the adjustment of teaching strategies according to the type of output generated. The proposal is based on a critical and interdisciplinary review that combines critical pedagogy, digital epistemology, algorithmic creativity and AIED systems theory. ZHIP is not only an analytical framework, but also a practical tool for redesigning the curriculum and guiding educational intervention in times of discursive automation. This paper invites us to rethink the role of the educator as epistemic mediator and to value the “plausible hallucinations” of LLMs as formative opportunities, beyond the logic of control. Thus, ZHIP is proposed as a strategic tool for managing the formative ambiguity of generative artificial intelligence in higher education.