Beyond Automatic Fluency: Human Agency, Quality Risk and Professional Practice in AI-Assisted Machine Translation

Artificial intelligence has moved machine translation from a specialist technology into an everyday infrastructure for multilingual communication. Neural machine translation and large language models can generate fluent output at remarkable speed, but fluency can conceal errors of meaning, terminology, register, culture and factual relation. This paper examines how AI-assisted machine translation is reshaping the work, competence and professional agency of human translators. It uses a critical integrative review and documentary policy analysis, supported by a proposed future mixed-methods design for subsequent empirical validation with professional translators. The analysis is organised around five interacting challenges: changing competence requirements; calibrated trust in opaque models; quality risks, including hallucination and domain mismatch; ethical, economic and labour pressures; and confidentiality, data governance and unequal language coverage. The paper argues that neither technological rejection nor uncritical automation offers a defensible professional strategy. Instead, AI should be governed as a risk-sensitive component of a translation service in which accountable human judgement remains decisive. To operationalise this position, the paper proposes the TRACE-MT framework: task and risk classification, responsible data handling, agency-preserving human oversight, contextual quality assurance, and equity in language coverage, expertise and working conditions. The framework connects translator education, post-editing protocols, procurement, model documentation and organisational accountability. Particular attention is given to low-resource African languages, for which expanding system coverage does not automatically guarantee contextual reliability or community legitimacy. The paper concludes that sustainable use of AI in translation depends on transparent workflows, professionally meaningful human control, fair recognition of cognitive labour, and evaluation methods that test adequacy and context rather than surface fluency alone.