NotebookLM AI for Omani EFL Vocabulary: Experimental Group Kept Gaining on the Delayed Posttest While the Control Group Declined

On vocabulary learning and retention, the experimental group significantly outperformed the control group on both the immediate and delayed posttests, and on the delayed posttest the experimental group was still improving while the control group declined. Prior empirical work on generative AI and vocabulary focused largely on ChatGPT, Copilot, and Memrise; this study brings NotebookLM AI into an Omani EFL classroom and measures both immediate and delayed time points. Seventy pre-intermediate Omani learners were randomly divided into two groups of 35; the intervention lasted 6 weeks and covered 80 target words; vocabulary tests were researcher-made 20-item tests with Cronbach's alpha of .850, .790, and .820 for pretest, posttest, and delayed posttest; groups did not differ at pretest (p = .153), differed at posttest (p < .001) and delayed posttest (p < .001), with eta-squared rising from .033 at pretest to .488 at posttest and .763 at delayed posttest.

On self-directed learning (SDL), the experimental group's SRSSDL score rose from 121.00 to 222.63 while the control group barely changed (119.89 to 120.89), with a significant Time x Group interaction. Existing studies mostly examined ChatGPT or chatbots; this study provides quantitative evidence on SDL for NotebookLM AI, a multimodal tool built on learner-uploaded materials. The 60-item SRSSDL (Williamson, 2007) was used; mixed-design ANOVA showed a main effect of time, F(1, 68) = 3730.114, p < .001, partial eta-squared = .982, and a Time x Group interaction, F(1, 68) = 3586.147, p < .001, partial eta-squared = .981; experimental-group subdomain means ranged from 41.06 to 47.17.

The study combined NotebookLM AI's video generation, infographics, quizzes, and flashcards with small-group collaboration in a weekly cycle of 20 words presented over four days plus a teacher-activity day. The authors state the originality lies in empirically applying NotebookLM AI in an EFL context and emphasize that the tool offers personalized, context-aware explanations and practice based on learner-uploaded materials. The experimental group was split into seven subgroups of five; 20 weekly target words were sent as PDFs via Microsoft Teams; each session included a 30-minute video, infographic, clarification, and quiz sequence, with the final 10 minutes using the flashcard feature for whole-class discussion; the control group learned the same words without AI support.

In the discussion, the authors attribute part of the experimental group's advantage to AI multimodality and adaptive feedback while explicitly noting that group discussion may also have contributed, and they caution that the small sample and short intervention require careful interpretation. Compared with studies reporting only AI effects, this study lists peer collaboration as a possible co-explanation and sets out scope limits including sample size, a single skill, and no analysis of gender or age. The discussion states the interpretation "needs to be made cautiously as the study's sample size was limited and the intervention period was short," and the conclusion notes that collaborative group activities "might be responsible for students' vocabulary development and SDL growth."

The results speak to practitioners and learners in self-access language learning settings: teachers can use NotebookLM AI's video generation, infographic generation, interactive chat, quizzes, and feedback to arrange individualized vocabulary practice beyond teacher-led instruction, and learners can use it outside class to review vocabulary, define new words, get instant help, take quizzes, and study at their own pace. The study was set in a foundation program at a university in Oman with 70 pre-intermediate, native Arabic-speaking students aged 19 to 21, over 6 weeks and 80 target words, so its conclusions apply to vocabulary learning and SDL development at this level and in this context.

Readers should note that the experimental group's posttest and delayed posttest data did not meet normality (p = .018 and p = .016), so non-parametric methods were used for between-group comparisons; the SRSSDL was analyzed only at the overall level rather than item by item; gender and age were not examined; and the authors themselves suggest that group collaboration may be a co-source of the vocabulary and SDL gains, so the independent contribution of AI remains to be isolated. In addition, the available text is an incomplete version in which tables appear only as prose descriptions, so specific numerical details should be checked against the original.