who-04-slides

TITLE SLIDE AI for Literature Review Co-adaptation Between Human and Machine Intelligence Ukubona LLC • WHO–India Seminar

SLIDE 1 — AI IS AN INFRASTRUCTURE “AI is an infrastructure.” — Jensen Huang

AI is not a tool, not a feature, not a shortcut. It is a layered system of computation, perception, reasoning, synthesis, and action. Understanding these layers prevents misuse, overreliance, and failure.

SLIDE 2 — THE PENTADIC STACK: OVERVIEW A practical model for public health officers:

  1. World AI The foundation. Models trained on global data. What exists. What is missing. What is biased.

  2. Perception AI The layer WHO officers are asking for. Locates, reads, and extracts information from literature.

  3. Agentic AI Plans multi-step tasks. Conducts screening. Manages PRISMA workflows.

  4. Generative AI Summarizes, synthesizes, drafts, and explains evidence.

  5. Embodied AI Auditing and safety in real-world systems (future clinics, hospitals, devices).

Each layer depends on solidity beneath it. Weak foundations produce biased outputs.

SLIDE 3 — WHY THIS MATTERS FOR WHO–INDIA 700+ officers Multiple languages (Hindi, Tamil, regional) Grey literature, field notes, programme reports Fragmented data across ministries and states Policy timelines that outpace traditional review cycles

AI must support these realities. Not replace epidemiological discipline. Not replicate Western bias.

SLIDE 4 — LIMITATIONS: INVARIANT ASPECTS OF SCIENCE & PUBLIC HEALTH Bias must be named. Bias emerges when World AI lacks Indian representation.

Gaps must be acted on. Generative AI can surface missing comparisons or populations.

Evidence must be found. Perception AI is needed to locate documents across languages and formats.

These limits are structural, not temporary.

SLIDE 5 — METHODS: JOURNAL & LANGUAGE AGNOSTIC A robust literature review must include: • English journals • Hindi and Tamil publications • National programme documents • Grey literature • Field notes and facility reports

Perception AI can read these. But Perception AI cannot find them unless humans surface the documents.

Method: Human discovery + AI extraction.

SLIDE 6 — RESULTS: A WORKFLOW TAILORED TO WHO-INDIA

  1. Officers upload documents (PDFs, reports, local studies).
  2. Perception AI extracts key facts across languages.
  3. Agentic AI organizes screening, PRISMA, and inclusion/exclusion.
  4. Generative AI synthesizes findings into draft outputs.
  5. Human experts audit and finalize.

Equity is embedded in every filter, not added later.

SLIDE 7 — THE REALITY OF “PERCEPTION AI” What WHO officers call “AI that can read literature” actually consists of: • OCR and text extraction • Language models • Document understanding • Citation network tools • Database search engines • Grey literature ingestion

These tools are immature unless grounded in World AI built on diverse data.

SLIDE 8 — WHY CURRENT PERCEPTION TOOLS FAIL INDIA Most tools read English PDFs well. Few tools find documents across India’s journals. None solve multilingual discovery at national scale. Western citation networks exclude Indian public health work. Grey literature is invisible to traditional academic tools.

WHO–India must define Perception AI around its own ecosystem.

SLIDE 9 — TOOL LANDSCAPE: WHAT EXISTS TODAY Perception AI (reads documents) Scholarcy, Humata.ai, SciSpace, Paper Digest

Discovery AI (finds documents) Semantic Scholar, Connected Papers, Research Rabbit, Litmaps

Agentic Screening (structured review) Rayyan, DistillerSR, Nested Knowledge

Synthesis AI (answers questions) Elicit, Consensus, SciSpace, OpenEvidence

Each plays one role. None deliver end-to-end, India-ready workflows alone.

SLIDE 10 — THE “PERCEPTION AI” TOOLS THEY WANT Examples WHO-India can safely adopt:

Scholarcy Summarizes PDFs in any language.

SciSpace Claims multilingual model coverage (75+ languages).

Humata Upload a Hindi or Tamil health report; query it in English.

ATLAS.ti (specialized) AI-coding of field notes and interviews in local languages.

These enable officers to work with India’s real documents—not just PubMed.

SLIDE 11 — THE SURFACE EXPERIENCE FOR WHO–INDIA Landing page concept:

AI for Evidence Synthesis
Policy outputs rooted in Indian soil → Co-adaptation between human and machine intelligence

Three things AI must always support: • Evidence must be found (Perception) • Bias must be named (World) • Gaps must be acted on (Generative)

Every officer can click deeper based on appetite: • Literature (India-relevant) • Tools (Perception AI) • Methods (journal/language agnostic) • Vocabulary (stack + governance)

SLIDE 12 — A SAFE, GOVERNED AI WORKFLOW

  1. Officers locate the evidence (human discovery).
  2. Perception AI extracts and translates.
  3. Agentic AI organizes and screens.
  4. Generative AI drafts with citations.
  5. Embodied AI audits future deployments.
  6. Human oversight at every stage.

Ensures: • Transparency • PRISMA compliance • Equity • Scientific rigor

SLIDE 13 — CONCLUSION AI will not replace public health. AI will amplify public health—when grounded in the correct stack.

Understanding the pentadic model lets WHO–India: • adopt AI safely, • adapt workflows intelligently, • and generate evidence that reflects India, not only the global North.

Co-adaptation between human and machine intelligence. Rooted in local data. Governed by public health principles.

SLIDE 14 — APPENDIX: TOOL RANKING TABLE (Insert your full 20-tool table here, already prepared.)

SLIDE 15 — APPENDIX: THE PENTADIC STACK IN DETAIL

  1. Physics (Ontology) → What exists.
  2. Engineering (Orchestration) → How systems coordinate.
  3. Grammar (Flows) → How signals move between layers.
  4. Prosody (Discipline) → How constraints guide safe outputs.
  5. Metaphysics (Epistemology) → How we know what we know.

A practical governance model for AI in public health.

The stack will hold as deep as you want to go.