03 · What You Need to Know
How AI Can Influence the Visibility of Local and Marginalized Knowledge
Knowledge Visibility Is Not the Same as Knowledge Quality
Research becomes visible through publication, indexing, citation, translation, dissemination, and institutional networks.
These processes do not provide equal exposure to every research community.
A study published in a highly indexed international journal may be easier to retrieve than relevant work distributed through a local institutional repository.
Greater visibility does not automatically establish greater scientific quality, just as limited visibility does not establish that a publication is reliable.
Researchers should evaluate evidence according to its methodological quality, relevance, provenance, and suitability for the research question.
AI-generated recommendations can help identify sources, but their apparent comprehensiveness should not be confused with the actual distribution of knowledge in a field.
Uneven Training Data Can Affect Which Perspectives AI Recognizes
Generative AI systems learn patterns from large collections of text and other information.
These collections may contain unequal representation of languages, geographic regions, institutions, and scholarly traditions.
When a perspective appears infrequently in the available material, the system may be less likely to produce a detailed or contextually accurate account of it.
For example, an AI assistant may provide extensive explanations of internationally established educational technology frameworks while offering comparatively limited discussion of locally developed concepts.
Such differences may reflect biases in AI training data, although retrieval procedures and prompt framing may also contribute.
Researchers should avoid attributing observed omissions to training data without evidence about the relevant system.
Scholarly Indexing Can Reinforce Existing Visibility Differences
AI research assistants may retrieve publications from scholarly databases, citation networks, publisher platforms, or other indexed sources.
These systems differ in their coverage of disciplines, countries, languages, and publication types.
Relevant studies may be absent from a retrieval index even when they are publicly accessible elsewhere.
For example, a researcher examining Philippine educational policy may need official government documents, local institutional reports, and peer-reviewed regional studies that are not equally represented in international citation databases.
If the AI assistant relies on a limited set of sources, its recommendations may overlook these materials.
This connects to bias in AI literature recommendations.
Researchers should examine the sources available to the tool rather than assuming that missing material does not exist.
English-Language Dominance Can Reduce the Visibility of Other Knowledge Traditions
English is prominent in international scholarly communication, but research knowledge is also produced and communicated through many other languages.
Local concepts may be discussed in national journals, community publications, policy documents, and research materials that are not primarily written in English.
AI systems may perform differently across languages because of differences in training resources, evaluation coverage, and linguistic complexity.
Consequently, research conducted in less represented languages may be harder for some systems to retrieve, summarize, or interpret accurately.
Researchers should examine whether language restrictions are appropriate to their research question.
The possibility of English-language conceptual assumptions influencing non-English research becomes particularly important when local concepts are translated into familiar international terminology.
Local Knowledge May Be Embedded in Practices Rather Than Conventional Publications
Not all knowledge relevant to research is communicated through peer-reviewed journal articles.
Community practices, oral histories, institutional experience, professional knowledge, and locally maintained records may contribute to understanding a phenomenon.
For example, teachers working in resource-constrained environments may develop instructional adaptations that are not formally documented in international publications.
These practices can provide valuable research questions and contextual evidence.
However, researchers should distinguish experiential knowledge from empirically established findings and avoid treating anecdotal accounts as equivalent to systematic evidence.
Appropriate research methods can examine local practices while preserving their context and acknowledging limitations.
Indigenous Knowledge Requires Particular Attention to Context and Governance
Indigenous knowledge systems may involve relationships among communities, environments, practices, histories, and responsibilities that do not align neatly with conventional academic categories.
Some knowledge may be collectively held, culturally restricted, or governed by community-specific protocols.
AI-generated summaries may remove these contextual relationships or present knowledge as freely reusable information.
Researchers should not assume that publicly accessible descriptions authorize unrestricted extraction, processing, or publication of Indigenous knowledge.
Relevant community permissions, research ethics, intellectual property arrangements, and data-governance principles should be considered.
The CARE Principles for Indigenous Data Governance provide a useful framework emphasizing collective benefit, authority to control, responsibility, and ethics.
Researchers should engage appropriate community representatives rather than relying on AI to determine how culturally governed knowledge may be used.
AI Can Replace Local Concepts With Familiar Academic Categories
Generative AI may interpret unfamiliar expressions through theoretical concepts prominent in internationally visible scholarship.
For example, a participant's description of interpersonal obligations may be categorized as extrinsic motivation or social support.
These categories may capture part of the experience, but they may also remove distinctions that matter to the participant's account.
Researchers should examine whether the proposed interpretation fits the original expression and surrounding context.
The problem is not the use of an international theory itself. It is the unsupported replacement of one conceptual meaning with another.
Where appropriate, researchers may retain source-language terms with contextual explanations.
Dominant Theories Can Shape Which Research Questions Appear Important
AI-generated theoretical recommendations may emphasize established frameworks because they are prominent in scholarly material.
In educational technology, a system may repeatedly recommend models of technology acceptance when researchers investigate digital learning.
Such models may be appropriate for questions concerning perceived usefulness or intention to use technology.
However, they may not fully address institutional governance, educational inequality, or community-specific learning practices.
Researchers should consider whether AI favors dominant scientific theories at the expense of relevant alternatives.
Theoretical selection should be based on explanatory suitability rather than recommendation frequency.
AI Can Create the Appearance of Universal Research Findings
A finding established in one population or institutional setting may not generalize automatically to another.
For example, research involving students at well-resourced universities may identify particular patterns of technology adoption.
These patterns may differ in settings where learners face connectivity limitations, shared-device access, or different institutional policies.
When AI summarizes the literature without preserving population and contextual qualifications, findings may appear more universal than the evidence supports.
Researchers should examine the populations, settings, and methods underlying consequential claims.
Contextual relevance is particularly important when international research is used to justify local educational interventions.
Local Research Should Not Be Included Merely for Geographic Representation
Addressing visibility inequalities does not mean that researchers must include every local publication regardless of relevance or quality.
Evidence selection should follow criteria appropriate to the research question.
A poorly designed local study should not automatically receive greater evidentiary weight than a rigorous international study.
Likewise, international findings should not be treated as universally applicable without considering contextual differences.
The objective is to evaluate all relevant evidence fairly rather than replace one visibility hierarchy with another.
Researchers should distinguish inclusive evidence identification from uncritical acceptance of every perspective.
AI Can Overlook Knowledge From Resource-Limited Institutions
Institutions with limited research funding, publication support, or access to international dissemination networks may produce work that receives comparatively little visibility.
Some research outputs may appear in institutional repositories, technical reports, or local journals rather than widely indexed publications.
AI systems that rely heavily on prominent scholarly sources may underrepresent this work.
For example, a study of technology-supported instruction in provincial institutions may require evidence from settings that differ substantially from highly resourced universities.
Researchers should seek relevant sources through appropriate national repositories, institutional archives, and disciplinary networks.
They should also evaluate the methodological limitations of each source.
Knowledge Omission Can Influence Research Gap Identification
Researchers sometimes ask AI to identify topics that have not been sufficiently studied.
However, a model may describe a topic as underexplored because relevant local or non-English studies are missing from its accessible information.
For example, an AI assistant might suggest that little research exists on digital learning among working students in Southeast Asia while overlooking locally published studies.
Such a claim should be treated as a hypothesis for verification.
Researchers should conduct appropriate literature searches before asserting that a research gap exists.
The absence of evidence from an AI-generated response is not evidence that the research itself is absent.
AI Can Influence Which Voices Are Quoted or Summarized
In qualitative research, AI may assist with selecting representative quotations or summarizing participant accounts.
These processes involve decisions about which experiences receive attention.
A system may emphasize common or familiar patterns while overlooking contradictory or less frequent responses.
For example, an AI summary of online learning experiences may emphasize flexibility while minimizing accounts from learners who lack reliable access to study spaces.
Researchers should verify whether the selected quotations adequately support the analytical claims.
Less frequent perspectives may be important because they reveal exceptions, contextual differences, or limitations of a proposed interpretation.
Frequency alone does not determine qualitative significance.
Epistemic Injustice Provides a Useful Critical Perspective
Epistemic injustice concerns ways individuals or groups may be disadvantaged in their capacity as knowers.
One form involves unfairly discounting someone's credibility. Another concerns gaps in shared interpretive resources that make particular experiences difficult to communicate or understand.
AI-assisted research may contribute to such problems when it repeatedly excludes relevant perspectives or translates unfamiliar experiences into categories that do not represent them adequately.
However, not every omission constitutes epistemic injustice.
Researchers should examine whether the omission reflects a relevant pattern of exclusion, whose interests are affected, and how the system's use contributes to the problem.
The concept provides a critical lens rather than a substitute for empirical evaluation.
How Can Researchers Evaluate Knowledge Visibility Bias?
Researchers should first define the knowledge domain and the populations or perspectives relevant to the research question.
They can then compare AI-generated recommendations or summaries with independently identified evidence.
For literature discovery, this may involve examining coverage across publication sources, languages, geographic contexts, and theoretical traditions.
For qualitative analysis, researchers may examine whether AI summaries preserve locally meaningful expressions and contradictory accounts.
For theory selection, researchers can evaluate whether the system overlooks relevant conceptual approaches.
Repeated tests may help identify systematic patterns, but the evaluation criteria should be established independently of the AI's outputs.
Researchers Should Avoid Extractive Use of Local Knowledge
Greater visibility does not always mean unrestricted disclosure is appropriate.
Some community knowledge may be sensitive, collectively governed, or shared for limited research purposes.
Researchers should consider whether AI processing complies with participant consent, community expectations, institutional ethics approval, and applicable legal requirements.
Where appropriate, community collaborators should participate in decisions about interpretation, representation, and dissemination.
Researchers should avoid using AI to generate authoritative explanations of communities without adequate engagement with the people whose knowledge is being represented.
Document How Knowledge Sources Were Identified and Evaluated
Researchers should record the sources, search procedures, language restrictions, and criteria used to identify relevant evidence.
Where AI contributes to literature discovery or synthesis, its role and relevant limitations should be documented.
For culturally situated material, researchers should explain how original expressions and contextual interpretations were preserved.
Documentation should also acknowledge when important sources were inaccessible or when the available evidence did not support confident conclusions.
Transparency allows readers to evaluate whose knowledge informed the study and where significant gaps may remain.
