AI Literacy: : How LLMs Work, Their Biases, and Their Operation
Rodríguez Joaquín, René Alexis
Literacidad en IA: Funcionamiento, Sesgos y Operatividad de los LLMs
Tema large language models
Tema tokenization
Tema algorithmic bias
Tema RLHF
Tema benchmarks
Tema corpus transparency
Tema literacidad en IA
Tema modelos grandes de lenguaje
Tema tokenización
Tema sesgo algorítmico
Tema RLHF
Tema benchmarks
Tema transparencia de datos
Descripción Large Language Models (LLMs) have moved from a specialized object of computational linguistics to a routine infrastructure for professional, scientific and administrative production. This essay proposes an AI literacy understood as the technical-epistemic comprehension of the LLM artifact, organized in three axes: how it works, what biases it inscribes, and what responsible operability can be expected from its use. From a qualitative-interpretive paradigm, a documentary research design with a critical-analytical approach is adopted, following Bowen’s (2009) model, drawing on primary literature indexed in Web of Science, Scopus, Google Scholar, arXiv and the ACL Anthology between 2017 and May 2026, on transformer architectures (Vaswani et al., 2017), tokenization (Petrov et al., 2023; Ahia et al., 2023), reinforcement learning from human feedback (Ouyang et al., 2022; Casper et al., 2023), construct validity (Raji et al., 2021; Bowman and Dahl, 2021), and data coloniality (Quijano, 2000; Couldry and Mejias, 2019). The analysis identifies three central technical operations (tokenization, attention and next-token prediction) whose design decisions penalize Spanish relative to English with a tokenization premium of 1.55 to 1 in GPT-3.5 and GPT-4; three structural layers of bias inscription (corpus, RLHF alignment and benchmarks) that invalidate the neutrality claim of commercial systems; a drop in aggregate transparency of foundation models from 58 to 40.69 points out of 100 between 2024 and 2025; and a gap of up to 30.2 percentage points between the best and worst languages in MMLU-ProX. The article concludes that technical-epistemic literacy is a necessary condition for Latin American academia, public bodies, private sector, and the third sector to articulate situated, transparent, and sovereign AI policies.
Tipo info:eu-repo/semantics/publishedVersion
Tipo Artículo revisado por pares
Formato text/html
Identificador 10.61454/yhetdk61
Fuente Espectro Investigativo Latinoamericano; Vol. 8 No. 2 (2026): Espila; 118-133
Fuente 2710-7515
Fuente 10.61454/zwp9gj54
Relación http://revistas.isaeuniversidad.ac.pa/index.php/espila/article/view/159/236
Derechos https://creativecommons.org/licenses/by-nc-sa/4.0