Investigating AI-based image generation and prompt engineering in interventions for acquired language disorders

Introduction: The visual materials used in language and speech interventions for acquired language disorders have many features that affect an individual’s performance, including imageability, naming agreement, content, context, and cognitive load. Artificial intelligence (AI)-based image generation has also begun to be used frequently in interventions for language and speech disorders. This study evaluated the suitability of AI-generated images for use in therapy and examined the characteristics of the prompts given to the model during image generation.Methods: For each target word with high imageability in Turkish, prompts of three different types (descriptive, conceptual, and action-based) were given to the Midjourney model, and three alternative images were generated per word. The generated images were rated by 35 speech and language therapists on a 0–5 Likert-type scale. A total of 12,495 ratings were made for 357 images. Image suitability was assessed with a one-sample t-test, and the effect of prompt type on image suitability with the Friedman test. Cronbach’s alpha was calculated for the internal reliability of the scale and the ICC for inter-rater reliability. For statistically significant comparisons, post hoc comparisons were made with the Dunn-Bonferroni test to identify which prompt type accounted for the difference.Results: Across all 357 images, the mean suitability score was 3.71 (SD = 1.40), and 64.7% of the images scored above the threshold value (3.5). The effect of prompt type on image suitability was significant (Cohen’s f = 1.20). The internal consistency of the ratings was excellent (α = .993). For ten words, no suitable image was obtained with any prompt type; these corresponded to general category names, body parts, actions, or geometric shapes (Table 1 and Figure 1).Discussion/Conclusion: Text-to-image models can produce original images free of copyright concerns that can be used in the rehabilitation of acquired language disorders; however, output quality depends strongly on how the prompt is structured. Descriptive prompts, which describe the target concept through its concrete features, and action-based prompts, which place it in the context of an action, yielded more suitable images than conceptual prompts, which rely on a fixed template structure. Model performance remained relatively low for abstract categories, body parts, and action words.Keywords: