The increasing use of artificial intelligence (AI) in academic translation has raised important questions about translation quality beyond grammatical accuracy and lexical fluency. In particular, the pragmatic dimension of translation, which involves the preservation of context-dependent meaning, authorial stance, and discourse conventions, remains underexplored in comparative human and AI translation research. This study investigates pragmatic errors and their impact on translational adequacy in human and AI-generated translations of academic texts. Adopting a qualitative linguopragmatic approach, the study analyses a corpus of academic research article abstracts translated by human translators and an AI-based translation system. The analysis focuses on key pragmatic features, including implicature, hedging and stance, deixis, and register and discourse organisation. The findings reveal systematic differences between human and AI-generated translations. While AI-generated translations demonstrate high levels of formal fluency, they exhibit recurrent pragmatic weaknesses, such as over-explicitation, inappropriate stance calibration, and discourse-level misalignment, which cumulatively reduce translational adequacy in academic contexts. Human translations, although not free from pragmatic deviation, show greater sensitivity to communicative intent and academic discourse norms through context-aware and strategic decision-making. The study contributes to translation quality assessment by highlighting pragmatic errors as a crucial indicator of adequacy and underscores the continued importance of pragmatic competence in AI-assisted academic translation and translator education.