Abstract:

Artificial intelligence chatbots have been widely adopted across higher education sectors as pedagogical support tools . However, their efficacy is not uniform across academic disciplines due to the systemic generation of hallucinations, defined as coherent and plausible text that lacks factual basis . This article presents a critical review of the integration of these tools, specifically contrasting their impact on the Sciences and the Humanities . It further examines recent technological developments, including models such as GPT-5 and Gemini Advanced, to understand how disciplinary epistemologies shape the severity and pedagogical implications of these errors.

Adopting an interdisciplinary perspective, this study synthesizes recent empirical literature and technical reports published through 2025 . The research employs a comparative epistemological framework to analyze the structural differences between scientific knowledge, characterized as cumulative and verifiable, and humanistic inquiry, which operates under a hermeneutic and interpretive paradigm . The analysis focuses on hallucination typologies, the limitations of mitigation techniques such as Retrieval-Augmented Generation, and evaluation benchmarks that incentivize models to guess rather than express uncertainty .

The analysis reveals that in STEM disciplines, where knowledge is structured and verifiable, hallucinations typically manifest as incorrect formulas, misapplied procedures, or inaccurate numerical outputs . Because scientific knowledge allows for objective validation, these errors can be systematically corrected through standard protocols, enabling chatbots to reduce cognitive load and enhance learning outcomes when properly supervised . Conversely, in the Humanities, hallucinations often involve fabricated citations, misattributed ideas, or the generation of an artificial consensus that oversimplifies complex debates . These distortions are epistemically corrupting as they compromise hermeneutic integrity and are difficult for novices to detect without specialized disciplinary expertise . While technical advances such as Semantic Entropy Probes demonstrate improvements in distinguishing plausible but incorrect statements, they do not fully replicate the critical dialogic capacity required for deep interpretive analysis .

The study concludes that technological solutions alone are insufficient and proposes a dual strategy framework . First, technological alignment requires discipline-specific systems that incorporate uncertainty indicators and traceable reasoning rather than relying on generic accuracy metrics . Second, pedagogical adaptation necessitates that educators transition from viewing chatbots as autonomous tutors to treating them as flawed dialogic partners . Specifically, in the Humanities, this involves transforming hallucinations into pedagogical opportunities by designing exercises where students explicitly identify and correct AI-generated fabrications, thereby cultivating essential skills in source evaluation and critical judgment .

Keywords:

Hallucinations, generative artificial intelligence, higher education.