Manifold Grace
SOCIAL EPISTEMOLOGY · INFORMATION SCIENCE · AI EVALUATION

AI Access Is Not the Limit of Human Knowledge

Why the boundaries of an AI model's training material, searchable records and retrieval behaviour cannot be treated as the boundaries of everything people know.

← Emerging Research Horizon · Interdisciplinary Atlas · Information Science

The central proposition

AI training data and retrieval capabilities do not define the limits of human knowledge. Human knowledge includes embodied expertise, social practice, observation, testimony, unreported findings, protected archives, local and Indigenous knowledge, and work not represented—or not retrievable—within a particular AI system.

A careful argument distinguishes existence from documentation, accessibility, representation and retrievability. Missing search results do not prove nonexistence; nor does a newly encountered claim automatically establish truth or innovation. Claims still require appropriate evidence and testing.

1. Human experience and inquiryObservations, practice, skilled action, social memory, experiments and testimony
2. Recorded or published evidenceOnly some experiences and results are articulated, documented or released
3. AI-accessible informationTraining or retrieval is limited by coverage, permissions, language, indexing and time
4. Model output and recallEven represented information may not be reliably recovered or interpreted

Logical consequence: an AI response reflects a conditional evidence-and-inference process, not a census of all human knowledge. The boundaries are different at every stage.

Concrete examples, not assumptions

Archival example: The National Archives confirms that not all records can be viewed online. Therefore an unsuccessful web search cannot establish that no relevant historical record exists.

Scientific example: The WHO describes incomplete clinical-trial results reporting. A literature search can miss existing experimental findings because those findings were not disclosed.

Social example: UNESCO recognises oral and Indigenous knowledge traditions, while OECD documents digital disparities. Communities may possess and apply knowledge that is poorly represented in searchable digital sources.

Technical example: Google Research (2026) separated encoded facts from recall failures in frontier models; the reported percentages concern its evaluated facts and methodology, not a claim that a model has encoded all possible knowledge.

Knowledge outside digital corpora

Publication and representation gaps

International social statistics

OECD — Digital divides

Access, skills and participation vary across populations; the online record is not equally representative of every group.

Epistemology and social credibility

AI-specific evidence and evaluation

How to investigate an apparently missing advancement

Verification checklist: from missing retrieval to a justified claim
  1. Define the claim and scope: What exactly might be absent? From which model, index, source set or date?
  2. Search relevant repositories: Archives, specialist libraries, languages, field researchers, patents, standards and unpublished or grey literature where lawful and appropriate.
  3. Consult relevant people: Domain practitioners, communities and rights-holders, with consent and due credit.
  4. Check alternative explanations: Search terminology, dataset exclusions, non-digitisation, licensing, indexing failures, obsolete records and retrieval or model recall failures.
  5. Verify evidence: Independently reproduce technical findings or triangulate historical and qualitative sources using domain-appropriate methods.
  6. Record provenance: Author, date, artefact, limitations, corrections, conflicts and uncertainty.
  7. Classify cautiously: Unfound, inaccessible, unpublished, unsupported, previously known, novel and validated are different statuses.
Epistemic distinction: openness is not credulity

Whether knowledge exists is an ontological question; how a person or system knows it is an epistemological question. The fact that systems are incomplete is a reason for inquiry, not a licence to accept extraordinary or unsupported claims. Human testimony, Indigenous knowledge, archival material and expert practice deserve serious assessment, not automatic acceptance or dismissal. Some knowledge must remain private or protected by rights, consent, confidentiality and cultural stewardship.