Abstract
Background
Teaching Turkish to A1–A2 learners relies heavily on word selection and multimodal presentation. While traditional curricula prioritize frequency-based lists, they often underestimate the visual-semantic dimension critical for novice learners. Computational tools such as Word2Vec and CLIP offer a data-driven alternative by quantifying word–image alignment, yet their pedagogical efficacy in authentic classrooms remains empirically underexamined. Furthermore, the comparative effectiveness of AI-generated versus authentic photographs in vocabulary instruction constitutes an unresolved issue in computer-assisted language learning research.
Aim
This pilot study investigated whether high visual-semantic coherence, computed using CLIP and Word2Vec, enhances immediate vocabulary gains compared with low-coherence lists. It further examined the non-inferiority of AI-generated images relative to real photographs and explored the educational-psychological mechanisms—motivation, self-regulation, cognitive load, and achievement goals—that underlie stakeholders’ experiences. Phase 2 was designed to explain and contextualize Phase 1 results by identifying the psychological mechanisms underlying observed learning outcomes, thereby achieving integration through the connecting strategy.
Materials and methods
An explanatory sequential mixed-methods design was employed. Phase one assigned 60 A1–A2 learners to three quasi-experimental conditions: (a) high-CLIP words with AI-generated images (
n
= 15), (b) high-CLIP words with real photographs (
n
= 15), and (c) low-CLIP words with textbook images (
n
= 30). The use of non-equivalent word sets across conditions constrains causal interpretation. Vocabulary knowledge was assessed via the Vocabulary Knowledge Scale at pretest, post-test, and week four, though the absence of delayed testing limits conclusions to immediate learning. Phase two administered validated instruments—TAM, MSLQ, AGQ, Paas cognitive load scale, and SDT basic needs scale—to 60 students and 15 instructors.
Results
Both experimental conditions significantly outperformed the control (
p
d
= 1.42). No significant difference emerged between AI and real images (
p
= 0.678), though the small per-group sample (
n
= 15) provided adequate power only for large effects (
d
≥ 0.75). Students reported moderate motivation (
M
= 3.52) and perceived fairness (
M
= 3.61), whereas instructors exhibited lower motivation (
M
= 2.89,
p
= 0.005) and higher perceived effort (
M
= 3.67 vs. 2.94,
p
= 0.003). Mastery-approach goals correlated positively with vocabulary gains (
r
= 0.48,
p
M
= 4.2/9) and uniform across conditions. Basic-needs satisfaction significantly predicted intrinsic motivation (
β
= 0.62,
p
Conclusion
AI-driven visual-semantic word selection demonstrates preliminary promise for concrete vocabulary instruction. Nevertheless, the quasi-experimental design, confounding word sets, absence of delayed post-tests, and limited statistical power necessitate cautious interpretation. Replication with standardized word pools, delayed assessments, and objective proficiency measures is essential to establish causal efficacy and to determine whether AI-generated images offer genuine equivalence—or merely undetected non-inferiority—relative to authentic photographs. The non-significant difference between AI-generated and authentic photographs should be interpreted as the absence of evidence for a difference, not as evidence of absence of difference.