💬 NLP

Developing and Validating the Spanish Version of the Large Language Models Dependency Scale (LLM-D12-SP)

This study presents the first psychometric validation of the Spanish version of the Large Language Model Dependency Scale (LLM-D12-SP), confirming its two-factor structure of instrumental and relationship dependency and establishing its reliability and validity for assessing psychological reliance on LLMs in Spanish-speaking populations.

Original authors: Tran Gia Bao, Mo El-Haj, Sameha Al-Shakhsi, Antonio Garcia-Cabot, Raian Ali, Ala Yankouskaya

Published 2026-07-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tran Gia Bao, Mo El-Haj, Sameha Al-Shakhsi, Antonio Garcia-Cabot, Raian Ali, Ala Yankouskaya

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart digital assistant that can write your essays, plan your weekend, and even chat with you when you're feeling lonely. For a long time, we thought of computers as just tools, like hammers or calculators—things you pick up to do a job and then put down. But lately, these tools have started acting more like friends. They remember your secrets, they cheer you on, and sometimes, we start feeling like we can't make a decision without asking them first. This shift is happening in a field called human-computer interaction, where scientists study how our brains and machines mix together. The big question isn't just "how often do people use these tools?" but "how much do people need them?" Just like a person might rely on a crutch for a broken leg, or a friend for emotional support, we are starting to wonder if we are becoming too dependent on these digital brains. This matters because if we lean on them too hard, we might forget how to think for ourselves or how to connect with real humans.

Enter the researchers, who decided to build a ruler to measure this "digital leaning." They created a special test called the Large Language Model Dependency Scale (LLM-D12). Think of this scale as a two-sided ruler. On one side, it measures Instrumental Dependency: this is when you use the AI like a super-powered calculator or a personal secretary to get tasks done. On the other side, it measures Relationship Dependency: this is when you treat the AI like a buddy, sharing your feelings or feeling lonely if you can't talk to it. The original ruler was made for English speakers, and a similar one was made for Arabic speakers. But what about the hundreds of millions of people who speak Spanish? They needed their own ruler to see if the same two sides of dependency exist in their culture.

The team behind this study took the English ruler, carefully translated it into Spanish, and tested it on 386 Spanish-speaking people. They wanted to see if the ruler worked the same way in Spanish as it did in English. Did the "tool" side and the "friend" side stay separate, or did they get all mixed up? They also checked if people who scored high on this ruler were also the ones who felt addicted to the internet or trusted the AI too much.

Here is what they found. The Spanish ruler worked perfectly. It confirmed that there are indeed two distinct ways people get dependent on AI. One group relies on the AI to help them make decisions and finish tasks (Instrumental), while another group relies on it for comfort and social connection (Relationship). The study showed that these two feelings are related but different, just like being good at math is different from being good at art, even if the same person can be good at both.

The researchers also discovered some interesting patterns about who leans on the AI the most. People who used the AI multiple times a day were more likely to rely on it for tasks. Interestingly, people who used their AI on mobile phones felt more of that "friendship" connection than those who used it on a computer. The study also found that people who trusted the AI a lot were more likely to depend on it, whether for work or for feelings. However, the study ruled out one idea: being dependent on AI doesn't seem to be about how much you naturally like to think hard (a concept called "Need for Cognition"). In other words, smart people who love solving puzzles can still get dependent on AI, just like anyone else.

The authors suggest that this new Spanish version of the scale is a reliable tool for scientists and psychologists. It proves that the way we depend on AI isn't just a one-size-fits-all thing; it has different flavors. Whether you are using AI to write a report or to talk about your day, this scale can help measure that reliance. While the study doesn't say that using AI is bad, it gives us a clear way to understand when we might be leaning on our digital friends a little too much, helping us keep a healthy balance between our human brains and our digital partners.

Technical Summary: Developing and Validating the Spanish Version of the Large Language Models Dependency Scale (LLM-D12-SP)

Problem Statement
As Large Language Models (LLMs) transition from occasional tools to habitual cognitive partners in organizational and personal settings, concerns have emerged regarding the erosion of human autonomy and the emergence of psychological reliance. While existing instruments measure "problematic use" or "addiction" based on substance-use frameworks (e.g., withdrawal, tolerance), they often fail to distinguish between instrumental reliance (using LLMs for task execution and decision support) and relational reliance (viewing LLMs as companions or sources of emotional validation). Furthermore, there is a significant gap in psychometrically validated tools for Spanish-speaking populations, despite the rapid expansion of LLM adoption in Spain and Latin America. The absence of a culturally and linguistically appropriate instrument limits the ability to assess the prevalence, structure, and correlates of LLM dependency in this demographic.

Methodology
The study aimed to validate the Spanish adaptation of the Large Language Model Dependency Scale (LLM-D12-SP) through a rigorous psychometric evaluation.

Participants: A final sample of 386 Spanish-speaking participants (Mean age = 28.0 years; 55% male) was recruited via Prolific. Inclusion criteria required active engagement with a primary LLM (e.g., ChatGPT, Gemini) and either frequent use or perceived significant dependence.
Instrument Translation: The scale was developed using a standard back-translation procedure to ensure conceptual equivalence with the original English version.
Measures:
- LLM-D12-SP: A 12-item scale comprising two subscales: Instrumental Dependency (6 items) and Relationship Dependency (6 items), rated on a 6-point Likert scale.
- External Validity Measures: Internet Addiction (IAT-7), Attitudes Toward AI (Acceptance and Fear), Need for Cognition (NFC), and Perceived Trustworthiness of the primary LLM.
Analytical Procedures:
- Monotonicity Check: Mokken scale analysis to assess item scalability.
- Confirmatory Factor Analysis (CFA): To test the two-factor structure against a one-factor alternative.
- Reliability Analysis: Cronbach's alpha and corrected item-total correlations.
- Discriminant Validity: Assessed via the Fornell-Larcker criterion, Heterotrait–Monotrait ratio (HTMT), and chi-square difference tests.
- Network Analysis: EBICglasso algorithm to examine item centrality and network structure.
- External Validation: Multivariate multiple regression to analyze associations with external constructs.

Key Results

Factor Structure: Confirmatory Factor Analysis supported the original two-factor structure (Instrumental and Relationship Dependency). While the Tucker-Lewis Index (TLI = 0.89) was slightly below the 0.90 threshold, the Comparative Fit Index (CFI = 0.91) and Standardized Root Mean Square Residual (SRMR = 0.07) indicated acceptable model fit. The two-factor model fit significantly better than a one-factor model ( $\Delta\chi^2 = 488, p < .001$ ).
Reliability: The scale demonstrated strong internal consistency. The total scale Cronbach's $\alpha$ was 0.89, with subscale alphas of 0.86 for Instrumental Dependency and 0.85 for Relationship Dependency. All items showed significant loadings ( $p < .001$ ), though the reverse-scored item in the Relationship subscale showed a slightly lower loading ( $\lambda = 0.48$ ).
Discriminant Validity: The two subscales were empirically distinct. The Average Variance Extracted (AVE) for both factors (0.52 and 0.48) exceeded the squared correlation between them (0.32). The HTMT ratio was 0.65, well below the 0.85 cutoff.
Network Analysis: Items clustered into two coherent networks. In the Instrumental network, items regarding decision confidence and habitual use were central. In the Relationship network, items regarding social support and reducing loneliness were central.
External Validation: Both dependency dimensions were positively associated with Internet Addiction and perceived trustworthiness of LLMs. Higher dependency was also linked to positive attitudes (Acceptance) toward AI but showed weak or no association with Need for Cognition (NFC) or AI Fear.
Usage Patterns: Higher frequency of LLM use and usage across both stationary and mobile contexts were associated with higher Instrumental Dependency. Mobile device usage was specifically linked to higher Relationship Dependency compared to PC usage.

Significance and Claims
The paper claims to provide the first psychometrically validated Spanish-language instrument for assessing LLM dependency. Its primary contributions are:

Cross-Linguistic Support: The study establishes that the two-dimensional structure of the LLM-D12 (Instrumental vs. Relationship) is robust across English, Arabic, and now Spanish-speaking populations, suggesting these are stable dimensions of human-AI interaction rather than cultural artifacts.
Conceptual Distinction: By validating the separation of instrumental and relational dependency, the scale allows researchers to distinguish between reliance driven by efficiency and task support versus reliance driven by social and emotional needs. This addresses a limitation in prior scales that often conflate these distinct forms of reliance into a single severity continuum.
Practical Utility: The LLM-D12-SP offers a reliable tool for investigating the psychological aspects of LLM use in Spanish-speaking educational, occupational, and clinical contexts. It enables cross-cultural comparisons and supports the development of evidence-informed guidelines for AI integration.
Theoretical Nuance: The findings suggest that LLM dependency is not inherently dysfunctional but represents a specific form of reliance that correlates with trust and acceptance. The lack of association with Need for Cognition indicates that this dependency is not merely a proxy for general cognitive avoidance but a specific technology-related construct.

The authors maintain a modest stance, noting that the scale is an initial operationalization of a developing construct. They emphasize that the study is cross-sectional and relies on self-report, limiting causal inferences about the stability of dependency over time or its direct link to negative outcomes. The paper positions the LLM-D12-SP as a foundational tool for future research to determine under what conditions LLM reliance may become maladaptive.

More like this