MAIA Framework

Theoretical and Analytical Foundation of the MAIA Framework

The MAIA (Model of AI Literacy) Framework is proposed as an experimental and analytical system aimed at the meta-evaluation and architectural validation of generative artificial intelligence models. It emerges in response to the growing reliance on large-scale language models within educational and research contexts, where the apparent diversity of systems may obscure underlying epistemological convergence.

Rather than evaluating isolated outputs in terms of correctness or performance, the MAIA Framework shifts the analytical focus toward the relational dynamics between multiple generative systems. In doing so, it enables the identification of patterns of convergence, redundancy, and divergence across outputs, thereby providing a basis for examining the structural behavior of distinct model architectures through their linguistic productions.

Central to this framework is the concept of monocentric dependency, here theorized as “CHAT-GPTlogy”, which refers to the tendency of different generative systems to reproduce semantically aligned responses, reinforcing a dominant epistemic structure while simulating diversity. This phenomenon raises critical concerns regarding the formation of knowledge, particularly in pedagogical environments where such systems are increasingly integrated.

To operationalize this analysis, the system introduces two core metrics: the Monocentric Dependency Index (IDM), which quantifies the degree of convergence among model outputs, and the Coefficient of Plurality (CPA), which measures the extent of divergence and potential epistemic variability. These indicators are not intended as definitive measures of truth or quality, but as heuristic tools for exploring relational patterns within AI-generated content.

The MAIA Framework is particularly situated within the domain of ELT (English Language Teaching), where the integration of generative AI tools necessitates a critical approach to technological mediation in learning processes. By foregrounding plurality and encouraging comparative engagement with multiple systems, the framework seeks to foster a form of AI literacy grounded in epistemological awareness and methodological reflexivity.

As a browser-based prototype, this system enables localized experimentation, iterative data accumulation, and user-driven interpretation, positioning itself not as a closed evaluative instrument, but as an open, evolving environment for academic inquiry and conceptual development.

Academic Note: This tool is experimental and intended for research and prototyping in AI evaluation and literacy studies.

IDM/CPA Metrics Concepts and Interpretations

To operationalize this analysis, the system introduces two main metrics: the Monocentric Dependency Index (IDM), which quantifies the degree of convergence between the outputs of the models, and the Plurality Coefficient (CPA), which measures the extent of divergence and the potential epistemic variability between generated responses.

• The IDM: can be understood as an indicator of the tendency of systems to produce structurally and semantically aligned responses, signaling possible zones of epistemological redundancy.
• The CPA :acts as an analytical counterpoint, evidencing the presence of variation, conceptual displacement or discursive reconfiguration between the compared outputs.

This system operationalizes two analytical indicators: IDM and CPA.

The analytical classification follows:

• Monocentric: IDM ≥ 70
• Hybrid: Range between 40 and 69
• Plural: IDM < 40

The scatter plot represents epistemological positioning: X-axis → IDM (convergence) Y-axis → CPA (divergence)

It is important to highlight that such metrics are not intended to establish normative criteria of truth, accuracy or quality. Instead, they operate as heuristic tools, whose function is to make visible relational patterns between textual productions of different systems, allowing the researcher to identify dynamics of convergence and divergence within generative ecosystems.

The validation of the data produced by the system does not take place in an automated way, but depends on an interpretative analytical process conducted by the researcher. Such validation occurs through the critical observation of the outputs, considering elements such as:

• recurrence of discursive structures;
• repetition of key concepts;
• argumentative alignment between different models;
• significant semantic variations;
• presence of epistemological displacements or conceptual reformulations.

The use of the framework implies an articulation between quantitative measurement (IDM/CPA) and qualitative interpretation, in which the researcher plays an active role in reading the data, either by analytical perception or by directed exploration of emerging concepts and structures during experimentation.

The MAIA Framework is particularly situated in the field of teaching English as a foreign language (ELT), where the integration of generative AI tools requires a critical approach to technological mediation in learning processes. By privileging plurality and encouraging comparative engagement with multiple systems, the framework seeks to foster a form of AI literacy grounded in epistemological awareness and methodological reflexivity.

As a browser-based prototype, this system allows for localized experimentation, iterative accumulation of data, and user-driven interpretation, positioning itself not as a closed evaluative instrument, but as an open and constantly evolving environment for academic inquiry and conceptual development.

Researcher Procedure Guidelines

System Limitations

• This prototype uses a simplified similarity method based on word overlap, not deep semantic analysis.
• Results may not accurately reflect true conceptual similarity between model outputs.
• All data is stored locally in the browser and may be lost if cache is cleared.
• No server-side validation or persistence is implemented.
• Graphical visualization is basic and intended for exploratory analysis only.
• Manual editing of values may introduce inconsistencies if not carefully validated.
• The system does not prevent duplicated or biased inputs.

References

Sugestions (APP)