TLDR: A new study successfully translates and validates the Attitudes Toward General Large Language Models (AT-GLLM) and Attitudes Toward Primary Large Language Models (AT-PLLM) scales for Arabic-speaking adults. The research confirms a reliable two-factor structure (Acceptance and Fear) and strong measurement invariance across genders. Key findings highlight cultural differences, with perceived societal benefits being a strong driver of acceptance and existential fears being prominent. These validated scales provide crucial tools for understanding and guiding LLM adoption and policy in the Arab region.
As large language models (LLMs) become an integral part of our global digital landscape, understanding how different cultures perceive and interact with these technologies is crucial. A recent study, titled Developing and Validating the Arabic Version of the Attitudes Toward Large Language Models Scale, addresses this need by introducing and validating Arabic versions of two key scales designed to measure attitudes towards LLMs.
The research, conducted by Basad Barajeeh, Ala Yankouskaya, Sameha AlShakhsi, Chun Sing Maxwell Ho, Guandong Xu, and Raian Ali, highlights the rapid growth of LLM adoption in the Arab world, including both global platforms and regional solutions like Fanar and Jais. This underscores the importance of having culturally and linguistically relevant tools to accurately gauge public sentiment.
Building upon existing frameworks for assessing attitudes toward artificial intelligence (AI), specifically the 5-item Attitudes Toward Artificial Intelligence (ATAI) scale, the researchers adapted two English scales: the Attitudes Toward General LLMs (AT-GLLM) and Attitudes Toward Primary LLM (AT-PLLM) scales. The AT-GLLM measures broad evaluations of LLMs as a class, while the AT-PLLM focuses on attitudes towards the specific LLM an individual uses most frequently, acknowledging that direct experience shapes perceptions.
To create the Arabic versions, the scales underwent a rigorous back-translation method and a pilot study with 37 participants to ensure clarity, comprehension, and cultural appropriateness. The final Arabic scales, both AT-GLLM and AT-PLLM, each comprise five items, divided into two dimensions: Acceptance (two items, e.g., “I trust them/it”) and Fear (three items, e.g., “I fear them/it”). Participants rated these items on an 11-point Likert scale.
The study involved a sample of 249 Arabic-speaking adults recruited through the Prolific online platform. The demographic breakdown included a diverse range of educational backgrounds and an average age of 28.31 years, with a near-equal split between male and female participants.
The findings confirm that the translated Arabic scales are reliable and valid tools for the Arab population. Psychometric analyses revealed a consistent two-factor structure (Acceptance and Fear) for both AT-GLLM and AT-PLLM. The scales also demonstrated strong measurement invariance across genders, meaning they measure the same constructs in the same way for both males and females, allowing for valid comparisons.
Interestingly, the study found that participants generally expressed higher acceptance and lower fear towards their primary LLM compared to LLMs in general. While the overall factor structure aligned with previous UK validation, some cultural nuances emerged. For instance, “trust” played a less central role in defining acceptance in the Arab sample, whereas “perceived societal benefits” (e.g., “They will benefit humankind”) were a particularly strong driver of acceptance. Conversely, existential fears, such as LLMs potentially “destroying humankind,” were exceptionally strong components of negative attitudes, while concerns about “job losses” were less central to fear in this context.
External validation showed that general AI acceptance strongly predicted LLM acceptance, and general AI fear strongly predicted LLM fear. However, self-efficacy (confidence in one’s abilities) did not significantly predict LLM attitudes once general AI attitudes were accounted for, a difference from findings in the UK sample. This suggests that while general AI attitudes are a strong foundation, a significant portion of LLM-specific attitudes remains unique and not fully explained by broader AI perceptions.
Also Read:
- Mapping the Landscape of Human–AI Trust in Generative AI
- Unlocking Human-Like Personalities in AI: The Power of Detailed Persona Profiles
The validated AT-GLLM and AT-PLLM scales have significant practical implications. They can be used in ongoing AI initiatives in the Arab region, such as the UAE’s TII-ATRC Falcon Arabic project and Saudi Arabia’s Humain initiative, to monitor public acceptance and concerns. Furthermore, these tools can facilitate localized research and policy-making, inform communication strategies, and guide interventions in sectors like education and healthcare, helping to build a more comprehensive global picture of LLM perceptions.
