Generative AI technology such as ChatGPT, Grammarly, and DeepL has changed the way linguistic learning occurs. Although such technologies can be of great benefit in terms of linguistic scaffolding, overuse can disrupt learner autonomy by diminishing cognitive activities and metacognitive monitor. The study focuses on the impact of AI-mediated learning on the autonomy of Jordanian English as a Foreign Language (EFL) university students to fill a gap in the existing empirical research on the topic of cognitive offloading in non-Western learning institutions. A mixed-method design was used involving quantitative survey (N = 376) to measure the frequency and metacognitive strategies involved in the use of AI, and semi-structured interviews (n = 22) to understand perceptions held by students. Demographic variables were controlled through hierarchical regression, and thematic code of qualitative data were subjected to inter-rater validation. The results of the quantitative research are that intensive use of AI is a significant predictor of reduced self-regulation, confidence in error-correction, and retention of vocabulary (b = −0.46, p < 0.001). Qualitative data show there is a paradox of dependence, AI will give immediate feedback whilst it tends to discourage any strategic interaction. It is important to note that, students who utilised reflective strategies in addition to moderate use of AI had higher autonomy levels. Pedagogically, the findings suggest that digital literacy should be trained and that AI-based scaffolds should be designed so that they do not interfere with the agency of learners in EFL settings.