Artificial Intelligence in EFL and ESP Academic Writing: A Bibliometric and Critical Systematic Review of Pedagogical Gains and Assessment Risks

Abstract Background Artificial intelligence (AI) writing tools, from automated writing evaluation (AWE) and paraphrasers to generative chatbots, are now routine in the academic writing of learners of English as a foreign language (EFL) and English for specific purposes (ESP). The literature has expanded faster than it has been integrated: studies of learning gains, feedback, teacher attitudes and academic integrity have developed largely in parallel. Aim This review maps the structure and evolution of research on AI in EFL/ESP academic writing and critically evaluates what the evidence does and does not show, with particular attention to the relation between pedagogical benefit and assessment risk. Method A hybrid bibliometric–systematic design informed by PRISMA 2020 was applied to OpenAlex. A three-block title/abstract search (2019–2026) and a journal-scope supplementary search yielded 4,846 records; transparent rule-based screening retained 2,612 journal documents. VOSviewer 1.6.21 was used for term co-occurrence (130 terms), overlay, reference co-citation (200 references) and country co-authorship (36 countries) analyses. Bibliometric results were integrated with abstract-level coding of 2,448 records and close reading of the most influential and gap-relevant studies. Results Output grew from 6 documents in 2019 to 1,261 in 2026 (to 4 November); 79.5% appeared in 2025–2026. Six thematic clusters emerged: feedback efficacy, integrity and governance, assessment and measurement, agency and authorship, scholarly publishing by non-native writers, and technology acceptance. Only 16.1% of term-link strength connected the pedagogical clusters with the integrity/authorship clusters. The intellectual base comprised an AWE and corrective-feedback foundation, a 2023 ChatGPT opportunity–risk wave and an emerging GenAI-feedback and assessment front. Of 492 experimental or quasi-experimental studies, 9.8% reported delayed or longitudinal measures, 5.7% examined transfer to unaided writing and 3.7% were situated in ESP; only 61 records (2.5%) addressed AI-text detection. Conclusions The field documents robust, short-term text-level gains but has rarely tested durable learning, and it has not connected those gains with the assessment regimes that judge AI-assisted texts. We conceptualise this disconnection as a mediated-accuracy paradox: the linguistic polish that AI makes available to L2 writers is also a feature on which style-sensitive detectors may act. An integrative framework, five propositions and a research agenda are proposed.